The relationship between heating energy use, indoor temperature and heating energy demand under reference conditions in residential buildings

Faculty of Engineering University of Porto The relationship between heating energy use, indoor temperature and heating energy demand under reference conditions in residential buildings Sara Moreira Coelho de Magalhães ii Porto January 2016 iii Submitted to the Faculty of Engineering of the University of Porto in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Sustainable Energy Systems Supervisor: Vítor Manuel da Silva Leal Faculty of Engineering of the University of Porto Co-Supervisor: Isabel Maria Noronha de Resende Horta e Costa Faculty of Engineering of the University of Porto Faculty of Engineering of the University of Porto Rua Dr. Roberto Frias 4200-465 Porto, Portugal Sara Moreira Coelho de Magalhães, 2016 Thesis developed in the frame of the MIT-Portugal Program iv i ACKNOWLEDGEMENTS I would like to start by expressing my special appreciation and thanks to my advisor Professor Vítor Leal, who kept a sense of humor when I had lost mine. I would like to thank you for encouraging my research, for your guidance and knowledge over the past four years and for allowing me to grow as a research scientist. I would like to offer my grateful thanks to Professor Isabel Horta, my co-supervisor, by her care, continuous support, persistence and availability, which have been invaluable for the conclusion of this thesis. I also would like to express my gratitude to Professor Eduardo de Oliveira Fernandes for having inspired me to enroll into this area of expertise, for the interesting discussions and for his insights into this research. I can´t forget the warm welcome of Professor Steve Connors at MIT. Thanks for the interesting discussions and, most of all, for your daily positive attitude, which is an inspiration. The joy and enthusiasm you have for your research was contagious and motivational for me. Next, I would like to acknowledge the financial support from the Portuguese Foundation for Science and Technology through the doctoral grant (SFRH/BD/51591/2011) under the MIT Portugal program and the funds from North Regional Coordination and Development Committee for the work developed under the “Energy Efficiency in the Built Environment” project contract NORTE-07-0124-FEDER-000036. The Institute of Science and Innovation in Mechanical and Industrial Engineering (INEGI) at FEUP hosted me during my PhD enrollment, which I am very grateful for. ii I would like to thank the Portuguese Energy Agency, ADENE, for the technical support and for providing access and permission to use ADENE’s SCE database. I am especially thankful to Dr. Paulo Santos and Engª Susana Soares, for their valuable comments on the advanced draft of the published paper. The monitoring campaign study discussed in this thesis would not have been possible without the collaboration of the directors, students and the following professors, from the school year 2013/14: Maria Helena de Almeida, Maria Brazilina Pacheco and Filipe Coimbra, from Escola Secundária Infante D. Henrique in Porto; Carla Carvalhais and Paula Martins, from Escola E.B. Gomes Teixeira in Porto; Alvarino Sampaio from Escola E.B. 2,3 António Feijó in Ponte de Lima; Luísa Fernandes, Célia Costa, Teresa Calvo, Mara Dias and Luís Freitas from Escola Secundária Emídio Garcia in Bragança, and Susana Fraga from Escola E.B. 2,3/S. Miguel Torga in Sabrosa. I would also like to thank Professor Fernanda Sousa from Department of Civil Engineering of FEUP for her technical support during data treatment. I also appreciate the expertise of Professor Szabolcs Varga and Ana Camanho from Department of Mechanical Engineering of FEUP, for their guidance at the beginning of the statistical work addressed in this thesis. My time at the Institute of Science and Innovation in Mechanical and Industrial Engineering (INEGI) was made enjoyable in large part due to the many friends and people who have supported and helped me along the way. A very special acknowledgement to Ana Ferreira for her crucial contribution on technical experience on building construction. The following friends all deserve my recognition by the crucial discussions, fun moments, patience, positive attitude, and by being such a great listeners: Nathan, Mafalda, Isabel, Miguel, Mariana, Vasco, Ana, Zenaida, Hugo and Célia. You are all in my heart forever. A special thanks to my family. Words cannot express how grateful I am to my beloved mother, father, and grandmother, for all of the sacrifices that you’ve made on my behalf, for all the love and for being my rock through all these years. A special thanks to my brother, who was an inspiration for me and said ‘you can do it’ to me countless times over the last few months. I iii would also like to thank all of my closest friends, who supported and incented me to strive towards my goal. Thanks Milena, Milai and beloved Maelena for you endless love and care. Lastly, I would like express appreciation to my supportive, encouraging, and patient love, who have always believed in me and was always my support in the moments when there was no one to answer my queries. Ricardo, your encouragement has bolstered my strength. Thank you. Sara Magalhães 2016 iv v ABSTRACT This thesis started by characterizing the thermal performance of the residential building stock in Portugal mainland and by performing a preliminary assessment of the ‘reference heating gap’ of the stock, using the Portuguese EPBD-derived EPC database. The second research topic concerned the characterization and prediction of indoor temperatures during the winter season, in the residential buildings in Northern Portugal. The work was based on the monitoring campaign carried out at Porto, Ponte de Lima, Bragança and Sabrosa for the winter season period of 2013-14. Models, particularly effective at predicting bedroom and the living room temperatures, were developed using linear regression with panel corrected standard errors. The last major topic was the development of statistical models to characterize the relationship between heating energy use, indoor temperatures and heating energy demand under reference conditions (HDRC) (i.e., values from energy rating/certification scheme´s databases) in the residential buildings. These models are applied to Portugal and to different geographical contexts. The developed models, along with data and assumptions from the previous chapters, assisted the assessment of the value of ‘heating gap’ of the residential building stock in Portugal mainland. It was found that the actual energy use for heating is only about 45% of that occupants would need to maintain a comfortable environment. Motivated by the lack of buildings data, this thesis also focuses on the potential use of energy rating/certification schemes´ databases, as a rich and available source of data for countries´ decision-making and future energy planning. The results of the models proposed will be of outmost interest for the development of energy planning practices regarding the residential building stock. xii xiii CONTENTS ACKNOWLEDGEMENTS ................................................................................................................... i ABSTRACT ...................................................................................................................................... v RESUMO ....................................................................................................................................... vii LIST OF ABBREVIATIONS ................................................................................................................ ix CONTENTS ................................................................................................................................... xiii CHAPTER 1 ..................................................................................................................................... 1 INTRODUCTION ............................................................................................................................. 1 1.1 General context ................................................................................................................. 1 1.2 Motivation and research objectives .................................................................................. 3 1.3 Thesis structure ................................................................................................................. 8 CHAPTER 2 ................................................................................................................................... 11 LITERATURE REVIEW .................................................................................................................... 11 2.1 Assessment of ‘heating gap’ and ‘reference heating gap’ of residential building stock .. 11 2.2 Characterization of indoor temperatures and heating patterns ...................................... 14 2.2.1 Importance of indoor temperatures and heating patterns characterization............... 14 2.2.2 Indoor temperatures and heating patterns ................................................................. 16 2.2.3 Drivers of indoor temperatures ................................................................................... 19 2.3 Modeling heating energy use .......................................................................................... 21 CHAPTER 3 ................................................................................................................................... 29 CHARACTERIZATION OF THE STOCK AND THE ASSESSMENT OF THE ‘REFERENCE HEATING GAP’ ..................................................................................................................................................... 29 3.1 The Energy Performance Building Directive in Portugal .................................................. 30 3.2 Characterization of the Portugal mainland residential building stock in terms of construction period ..................................................................................................................... 31 3.3 Characterization of the thermal performance of the residential building stock .............. 37 3.3.1 Evolution over time and hypothetical effects of regulations ....................................... 37 3.3.2 Disaggregation of the stock by thermal performance levels ....................................... 42 xiv 3.4 Theoretical energy demand under reference conditions: A majorant for the country’s theoretical energy demand under thermal comfort conditions .................................................. 46 3.5 Estimation of the gap of the residential building stock for space and water heating ...... 49 3.5.1 Top-down estimation of actual energy use for heating, cooling and domestic hot water 49 3.5.2 Reference heating gap for space and domestic hot water .......................................... 51 3.6 Conclusions ...................................................................................................................... 52 CHAPTER 4 ................................................................................................................................... 55 CHARACTERIZING AND PREDICTING INDOOR TEMPERATURES IN RESIDENTIAL BUILDINGS ....... 55 4.1 Data collection and sample.............................................................................................. 56 4.1.1 Selection of households ............................................................................................... 56 4.1.2 Indoor and outdoor temperature dataset ................................................................... 57 4.1.3 Socio-economic factors and building characteristics dataset ...................................... 59 4.2 Characterization of indoor temperatures ........................................................................ 66 4.2.1 Overview of daily mean outdoor and indoor temperatures ........................................ 66 4.2.2 Characterization of daily mean indoor and outdoor temperatures profile over time . 71 4.2.3 Thermal comfort levels ................................................................................................ 74 4.2.4 Identification of heating patterns ................................................................................ 76 4.3 Modeling indoor temperatures ....................................................................................... 78 4.3.1 Description of the statistical models ........................................................................... 78 4.3.2 Description of database used in statistical models ...................................................... 80 4.4 Results and discussion ..................................................................................................... 82 4.5 Conclusions ...................................................................................................................... 88 CHAPTER 5 ................................................................................................................................... 91 MODELING HEATING ENERGY USE IN RESIDENTIAL BUILDINGS .................................................. 91 5.1 Modeling concept ............................................................................................................ 93 5.2 Modeling architecture ..................................................................................................... 95 5.2.1 Building archetype ....................................................................................................... 98 5.2.2 Geographical location ................................................................................................ 105 5.2.3 Set point temperature ............................................................................................... 106 5.2.4 Heating patterns and indoor heat gains .................................................................... 106 5.2.4.1 Definition ................................................................................................................... 106 5.2.4.2 Occupancy and occupant´s behaviour characteristics variables to define heating patterns and indoor heat gains .................................................................................................. 108 xv 5.3 Description of variable database used in the statistical models .................................... 113 5.3.1 Description of database: A1 - Varying only building archetype, geographical location and set point temperature ........................................................................................................ 115 5.3.2 Description of database: A2 - Varying building archetypes, geographical locations, set point temperature, heating patterns and indoor heat gains ..................................................... 118 5.4 Description of the statistical models ............................................................................. 121 5.4.1 Statistical models ....................................................................................................... 121 5.4.1.1 Multivariate regression .............................................................................................. 122 5.4.1.2 Artificial neural networks ........................................................................................... 123 5.4.2 Performance evaluation of the models ..................................................................... 126 5.5 Development of the universal models ........................................................................... 128 5.5.1 Example of the applicability of universal models: Portugal ....................................... 134 5.6 Development of Portugal specific models ..................................................................... 136 5.7 Graphical representation of the energy-temperature relationship ............................... 141 5.7.1 Heating energy use as a function of indoor temperature and HDRC ......................... 141 5.7.2 Indoor temperature as function of heating energy use and HDRC ............................ 144 5.8 Estimation of the ‘heating gap’ ...................................................................................... 146 5.9 Conclusions .................................................................................................................... 149 CHAPTER 6 ................................................................................................................................. 153 OVERALL CONCLUSIONS AND FUTURE WORK ........................................................................... 153 6.1 Contributions ................................................................................................................. 154 6.2 Implications for the real practice ................................................................................... 158 6.3 Future work ................................................................................................................... 158 REFERENCES .............................................................................................................................. 161 Appendix A................................................................................................................................. 177 Appendix B ................................................................................................................................. 179 Appendix C ................................................................................................................................. 199 xvi 1 CHAPTER 1 INTRODUCTION 1.1 General context The energy policy context since the beginning of the XXI century requires a strong incentive towards energy demand-side management and energy efficiency. The buildings sector is, along with transports and renewable energy, a key area of intervention. This happens because it accounts for 35% of the world final energy use [1], and also because it is recognized as one of the sectors where carbon abatement can be achieved with lower costs. Among the various energy uses in buildings in developed countries, heating represents about 45% of the total energy use. Properly designed buildings, in view of their local climatic conditions, can lead to drastically moderate or reduce final energy demand for comfort (e.g. heating, cooling, ventilation, and lighting energy services) [2] as a result of careful implementation of what is sometimes called as ‘sufficiency’ strategies [2–4]. The sufficiency strategies give the building, as an ‘energy system’, the ability to catch and manage ambient energy for the purposes of comfort, resorting to external insulation of walls, thermal inertia of the internal walls, orientation and sizing of openings, shading of the glazed surfaces, etc. In this way, the design of a building can pre-empt part or all of the energy demand for comfort with little need for ‘add-on’ energy systems. If still ‘add-on’ systems are required, energy-efficiency concerns [2] are applicable to reduce the energy use to generate the energy service [4]. Because of the innumerous non-technological solutions (sufficiency strategies), at the level of the building operation and envelope, designed to reduce the extent of energy services 2 needed to maintain the required comfort level in a building [4], and because of the technological solutions that are designed to provide energy services at lower levels of energy use (energy efficiency measures) [4], the building sector has been receiving especial attention from policy makers [5–7]. In what regards heating, there is awareness that indoor thermal comfort for a significant number of existing dwellings worldwide is not yet guaranteed. Several studies in the literature point out that indoor temperatures in winter, in many dwellings, are kept below the levels usually deemed as comfortable, and in many cases, even below the recommended levels [8– 16]. It is known that at least in some European countries this reality is associated, in part, with the poor building construction in terms of ‘sufficiency’ and/or the lack of efficient central heating systems [17,18]. For these reasons, this problematic is often named as ‘cold homes’ [14,17,19]. The issue is that, supposing unmet thermal comfort needs is a reality, in the long term, it could have particular influence on the performance outcome of existing building renovation and/or energy efficiency programmes. This happens because these programmes are normally designed assuming reference indoor temperatures/heating patterns that are usually not accurate. It has been widely reported that thermal upgrades, more energy efficient heating systems and better controls, especially in homes which are operated at low indoor temperatures, do not always save as much energy as predicted [20–22] and can lead to unintended consequences [23]. Actually, it might happen that occupant´s thermal comfort expectations get more demanding with the upgrades. For example, Critchley et al. [14], in their study of the impact of Warm-Front efficiency programme, registered occupants reporting ‘‘I have never been used to heating upstairs’’ and ‘‘I noticed the difference (after efficiency measures) though I might have thought differently before I had central heating”. This type of behavioural response to improvements, which leads to shortfall in expected energy savings as a trade-off for warmer temperatures, is a form of the so-called ‘rebound effect’ [24–29]. The register of indoor temperatures lower than those needed for comfort may also represent a potential for future increase in heating energy demand, if the economic conditions improve in the mid-term future; the energy prices go down; and/or there is a demand for higher thermal comfort levels (i.e., increase of thermal comfort expectations). 3 Hereupon, both for energy and health policies, understanding in detail the actual indoor temperatures, the heating patterns, as well as the unmet thermal comfort needs of existing residential buildings, is fundamental as they may give some indication on what to expect regarding the evolution of the energy use for heating, and influence the design of energy and climate plans, as well as the design of new dwellings and refurbishment of existing ones. 1.2 Motivation and research objectives There are two approaches that can provide with an indication of whether or not occupants feel comfortable within their indoor environmental conditions. These are the characterization of indoor temperatures (by evaluating if they are within the commonly recommended values) and the evaluation of the perception of thermal comfort of the occupants to their indoor environment [13,30–51]. But, the comparison between the ‘theoretical heating energy demand under (ensuring) thermal comfort conditions’ and the ‘actual energy use’ for heating, of the residential building stock, can be a useful exercise to assess whether the thermal comfort needs are truly being met. This makes place for the concept of energy use gap, defined as the amount of additional heating that would be needed to ensure that buildings were maintained at comfortable indoor temperatures, in their current physical state. At this stage it is important to recall that this theoretical concept of energy use gap does not apply when the occupants are satisfied with their thermal comfort environment. Also, somehow related to social aspects of thermal comfort expectations, the exercise of the estimation of energy use gap is a challenging one. The estimation of the theoretical value for the entire residential building stock is difficult because the levels of thermal comfort are usually unknown, dependent on each occupant and subjected to changes over time and with age. When aiming at determining the ‘theoretical heating energy demand (THD) under thermal comfort conditions’ of the residential building stock, one faces several possibilities. At the uppermost extreme of these is to assume that all indoor space needs to be maintained at a 4 certain temperature during all the winter/heating season period (e.g. at 20ºC or 18ºC) to ensure comfort indoor environments. This is an assumption often found in energy rating/certification schemes such as those in place after the Energy Performance Buildings Directive (EPBD) and it is usually named as reference conditions. However, more relaxed/reasonable values will probably be achieved if it is taken into account that occupants do not need to heat homes during all time, and not all rooms at the same temperature, and even not the same temperature at all time (e.g. comfort during sleep can be achieved at lower temperatures than during active hours [52]): i.e., assuming more relaxed values for thermal comfort conditions. This leaves two possible but different situations of estimating the ‘THD under thermal comfort conditions’. When the estimation of THD assumes more relaxed levels of thermal comfort conditions, the energy use gap is designated as ‘heating gap’. In turn, when the THD assumes stringent thermal comfort conditions, i.e., when theoretical heating energy demand under reference conditions (HDRC) values, provided, for example, by energy ratings/certifications schemes, are used as direct values of thermal comfort conditions, the computed energy use gap is called as ‘reference heating gap’. Figure 1 illustrates the two variants of the energy use gap that are introduced in this thesis in the context of the residential building stock. Figure 1. Schematic illustration of the different levels of heating energy and the ‘heating gap’ and ‘reference heating gap’ of the residential building stock. 5 ‘Heating gap’ is the difference between (2) and (3), in Figure 1, and ‘reference heating gap’, the difference between (1) and (3), in Figure 1. Currently, there are several energy ratings/certifications schemes in force worldwide and these hold extensive energy performance certificates´ (EPC) databases ready to be exploited. Actually, motivated by the lack of building data, one can think that the estimation of the ‘THD under thermal comfort conditions’ of the residential building stocks can be assisted with those databases, provided that they contain HDRC values. Based on recent literature findings [53–59] the EPBD-derived EPC databases emerge as a good example of application for the European context. The European Commission has put forward the EPBD directive (2002/91/EC) in 2002 [6]. This Directive was enacted for labelling the energy performance improving the energetic quality of new buildings and existing building stocks. A subsequent update under directive 2010/31/EC [7] (the EPBD recast) set more demanding objectives, such as the nearly net-zero energy building. The implementation of the EPBD 2002/91/EC with the attribution of energy performance certificates (EPC) to almost all buildings in Europe has therefore somehow initiated the mapping of thermal performance of the existing European building stock [60]. The EPBD-derived EPC databases compile a great number of energy performance certificates, which have been issued both for new and existent 1 buildings. Each certificate provides a theoretical HDRC value, for each building [61]. The computation methodologies require a number of operating conditions to be defined, such as building’s density of occupation, set point temperature, occupancy profiles and operational schedules of building services. These operating conditions (i.e., occupant behaviour), along with other physical building parameters, are often unknown in the design of new buildings or subject to a lot of uncertainty in existing buildings. This justifies the use of values assuming reference conditions under the EPBD, or under any other energy rating/certification scheme. Even if the purpose of EPBD methodology is mostly to ensure compliance with Building Regulations [62], and even though some thoughts regarding the need for improvements to the 1 Only required when there is a commercial transaction. 12 temperature that has been the main focus [10,47,83–85]. Levels of comfort with respect to those environmental variables are modified according to clothing insulation and activity level [82], but a further potential influencing factor is considered to be adaptation, which is closely related to experience and expectations of each individual [86–88]. For example, it is argued that the effect of adaptation occurs over time according to outdoor conditions, so that higher indoor temperatures are accepted as comfortable when the outdoor temperature is high, and lower indoor temperatures are accepted as comfortable when the outdoor temperature is low [89,90]. Also, in an empirical study to 600 households in Sweden, Linde´n et al. [55] found that those living in detached houses tend to accept lower indoor temperatures than households living in flats. Authors also found that for households living in dwellings where the energy bill is paid collectively the indoor temperature is higher by 2ºC. All these factors may vary according to the members of the household as some occupants may be more susceptible to high or low indoor temperatures than others [83], emphasizing the difference of thermal comfort expectations among occupants. As presented in section 1.2, the definition of energy use gap is associated with occupants living in indoor environment conditions that may not meet their expected level of thermal comfort. It is estimated by evaluating how much the ‘actual energy use’ for space heating is lower than the so called ‘theoretical heating energy demand (THD) under thermal comfort conditions’. Due to the several assumptions and considerations found in literature behind the estimation of THD, it is, at this point, important to recall the fact that, in the estimation of ‘heating gap’, the THD values regards the demand needed to ensure relaxed thermal comfort conditions ((2) in Figure 1, in section 1.2). Also, in the estimation of the ‘reference heating gap’ it regards the demand needed to ensure stringent thermal comfort conditions ((3) in Figure 1, in section 1.2). Several studies on the gap between the theoretical and the actual measured performance of buildings [62,91] can be found in literature. This comparison is termed as ‘performance gap’, by some authors [53,54,62,70,91–97] and it matches the concept of ‘reference heating gap’, introduced in section 1.2, because of the similarities in the methodology of estimation. 13 However, the ‘performance gap’ is not estimated with the purpose of determining whether the gap between the ‘theoretical’ and the ‘actual’ is part the expression of the difference between reference and actual values for heating patterns (i.e., occupant behavior), or other parameters, or if it is also the expression of a ‘deficit of comfort’. Instead, an expression of the first [91], in particular, differences in the occupant behaviour, is pointed out as the major reason for the ‘performance gap’. This is because occupant behaviour plays a central role in controlling heating energy use; and because operating conditions (i.e., occupant behaviour) are normally considered standard/reference rather than actual measured conditions [24,53,54,70,91], which can differ significantly from each other (due to the complex nature of the determinants of occupant behavior [29]). Indications of the ‘performance gap’ started to appear from the mid-1990s [98], until nowadays. From the academic literature reviewed [27,92,99,100] on the comparison between theoretical and actual energy use, most of the authors reporting situations where actual energy use is lower than theoretical values estimated the ‘performance gap’ [53–56,59,70]. From those examples, two actually showed some preoccupations in relation to the future energy savings, which somehow resembles the concept behind the ‘reference heating gap’. For example, Sunikka-Blank and Galvin [57] based their study on existing German datasets that included the calculated thermal performance ratings (i.e., energy performance certificates) and the measured energy use data from around 3400 dwellings. They concluded that occupants use, on average, 30% less heating energy than the calculated rating (estimation is done in kWh/m2.year). The authors called this phenomenon as ‘prebound effect’ and it is referred to the situation before a retrofit, indicating how much less energy is consumed than expected. The authors also considered that the discrepancy suggests less potential for economically feasible savings in Germany’s domestic heating energy than assumed, especially because of the correlation between the ‘prebound’ effect, household income level and energy bills. Also, Tigchelaar et al. [58] in their analysis, from 4700 households in Netherlands, found an identical phenomenon, though they named it the ‘heating factor’ (average of 0.7). The information used came from a database that was previously obtained through a national survey and compared with the calculated energy performance certificate for each respondent´s home (authors claimed that the figure is representative of the Dutch housing stock for the year of 14 2006). They suggested that this ‘heating factor’ severely limits the potential savings through thermal retrofits. One major observation resultant from this analysis is the evidence of similar definitions in literature attributed to those of ‘heating gap’ and ‘reference heating gap’. The analysis of the literature leads to the conclusion that there is no proof of evidence that the concept of ‘heating gap’, with its associated ‘relaxed comfort’, has been studied. Furthermore, studies, where actual energy use is lower than theoretical values assuming reference conditions, were found [27,92,99,100]. Some of these studies aimed to estimate the ‘performance gap’ and others showed preoccupations in relation to the future energy savings, which resembles the concept behind the ‘reference heating gap’ (but at an individual level). Moreover, the study of ‘heating gap’ and ‘reference heating gap’ have never been assessed for an entire residential building stock of a city, region or country. It seems, therefore, there is a lot of ground for exploring this issue. 2.2 Characterization of indoor temperatures and heating patterns 2.2.1 Importance of indoor temperatures and heating patterns characterization Recent medical research has associated low indoor temperatures to various illnesses (e.g. pneumonia, increased blood pressure, asthma, bronchitis, influenza arthritis and heart diseases) and social pathologies (e.g. depression, anxiety, constraints of mobility and isolation) [16,31,32,101–104]. Low indoor temperatures, which are frequently associated with the ´fuel poverty´ phenomena [13,31,101–103], have also a serious impact on mortality [14,16,105]. According to [106], there are approximately 30.000–60.000 excess winter deaths in UK, and 1500–2000 in Ireland related to low indoor temperatures in dwellings. Several international 15 standards define threshold indoor temperatures for health reasons. The proposed indoor temperatures are in the range of 18 to 21ºC, varying as a function of many parameters regulating thermal comfort. For instance, the World Health Organization (WHO) recommends 21ºC in the living rooms and 18ºC in the other occupied rooms to achieve an adequate standard of warmth [107]. Also, the UK Department of Environment proposes as minimum temperatures for health reasons 16ºC in bedrooms and 18ºC in living rooms [108]. Empirical data for residential indoor temperatures and heating patterns have important implications for policymakers in the development of programmes to improve indoor thermal comfort and health conditions. It also has an important role to support energy demand models for the building stock [109] (e.g. more accurate estimations of the actual heating energy use) and energy planning (e.g. studies on the impact of energy efficiency programmes on future energy savings). The provision of accurate information on indoor temperature and heating patterns (i.e., on occupant behaviour [110]) has become increasingly important over the last decade as governments worldwide move to adopt policies aimed at reducing carbon emissions through improvements to the building stock [24]. Despite the availability of many energy models to support energy planning and policy development it is not often clear which assumptions for indoor temperature and heating patterns estimations are being made and their empirical basis. In most cases operating conditions (i.e., occupant behaviour) are considered standard/reference rather than actual measured conditions [24,53,54]. For instance, the British Research Establishment’s Domestic Energy Model (BREDEM) assumes that living room is heated to 21ºC and other premises to 18ºC for 9h on weekdays and for 16h on weekends [24,111]. However, some authors found out that homes displayed on average lower indoor temperatures during assumed heating periods, and significantly shorter durations of heating than models usually assume. For example, Huebner et al. [46] concluded that currently used reference assumptions of heating demand and heating duration do not accurately reflect the living room temperatures in England. Also, Kane et al. [20] studied the heating patterns in 249 dwellings in Leicester in UK and concluded that indoor temperatures were much lower than those often assumed by BREDEM-based energy models. This can poses some limitations in a scenario where actual heating energy use values needs to be estimated. 16 2.2.2 Indoor temperatures and heating patterns Most of the existing studies in the literature analyze indoor temperatures in UK [8,10,12– 14,17,19,31,32,44,50,84,112], southern [33,43,51],southeast European countries [16,24], and non-European countries [37,47–49,52,113–115]. Typically, studies revealed a broad range of indoor temperatures. For example, Yohanis and Mondol [10] measured the indoor temperatures of 25 households in Northern Ireland at four locations (bedrooms, living rooms, halls, and kitchens) and analyzed data on seasonal, monthly and daily bases. The households were selected from 800 Northern Ireland households based on house type, heating system, number of occupants, location and employment status. In 80% of homes, the winter mean daily temperature was between 15ºC and 20ºC, and in summer between 20ºC and 23ºC, maintaining a reasonably comfortable temperature throughout the year. In 14% of homes, the daily mean temperature was above 21ºC throughout the year, suggesting a higher household temperature than required for comfort, which indicates wasteful energy behaviour. In three percent of homes, the heating was not used adequately and the winter mean temperature was below 15ºC. More recently, Kane et al. [20] verified that mean winter temperatures, measured in the individual homes, ranged from 9.7ºC to 25.7ºC in living rooms, and 7.6ºC to 24.2ºC in bedrooms. Some studies report low indoor temperatures [9,11,13,14,16,18]. For example, Hunt and Gidman [9] during February and March 1978 undertook spot measurements of the wet- and dry-bulb temperatures in each room of 1000 homes in UK. The mean of the living-room temperatures recorded was 18.3°C, and the mean temperature of the warmest bedroom was 15.2°C. The average dwelling temperature was 15.8°C. Hutchinson et al. [19] analyzed data from five urban areas of England. Half-hourly livingroom temperatures were recorded for two to four weeks in dwellings over the winter periods (i.e, November to April in 2001–02 and 2002–03). Overall, 21.0% of the dwellings had daytime living-room temperatures lower than 16ºC, and 46.4% had nighttime bedroom-temperatures below the same temperature. Also, Critchley et al. [14] analyzed data from a national survey of 888 dwellings in England occupied by low-income residents over the winters of 2001–02 and 17 2002–03. A total of 222 households were identified as occupying cold homes, with mean bedroom temperature below 16ºC or mean living room temperatures below 18ºC. French et al. [18] undertook indoor temperature monitoring in over 400 dwellings in New Zealand. Temperatures were logged every 10 minutes during one year in bedrooms and living rooms. The mean living room temperature was 17.9ºC. The maximum mean was 23.8ºC, and the minimum mean temperature was 10ºC. The bedrooms on average always seem to be slightly lower than the living rooms (at the most there is a difference of 3.8ºC which occurs during the evening). This is mainly caused by heating occurring in the living room and, typically, very little or no heating in the bedrooms. Santamouris et al. [16] collected indoor temperature and energy data during the winter 2012–13 from 50 low and very low income dwellings in Athens area in Greece. The results show that indoor temperatures were much below the accepted standards and, in many cases, place in risk the health and even the life of the residents. Table 1 presents temperature monitoring studies. Table 1. Temperature monitoring studies with broad range of indoor temperatures. Authors Location of study Mean temperature (ºC) Living room Bedroom Hunt and Gidman (1982) [9] (n= 1000) UK 18.3 15.2a Oreszczyn et al. (2006) [12] (n=1604) UK 19.1 17.1 Huntchison et al. (2006) [19] (n=470) England 18.2 16.4 Summerfield et al. (2007) [8] (n=14) UK 20.1 19.3 Critchley et al. (2007) [14] (n=888) England 18.0 16.0 French et al. (2007) [18] (n=400) New Zealand 17.8b 15b Yohanis and Mondol (2010) [10] (n=25) UK 19.4 18.4 Santamouris et al. (2014) [16] (n=43) Greece 15.9c Kane et al. (2015) [20] (n=249) UK 18.5 17.4 aAverage temperature of the warmest bedroom; bFor the evening period (17:00 to 23:00); cAverage between living room and bedroom temperatures. Heating patterns in the residential sector were also explored by several authors. Kane [116] developed a monitoring campaign in 300 homes in Leicester in UK, and found the following main heating patterns: a) the heating threshold temperature was found to be 8°C to 18°C. This 18 range indicated that some dwellings may be heated throughout the whole year, while others only during the coldest winter months; b) the average duration of daily heating period was 12.6 hours. The longest and shortest heating periods found were 22 hours and 4 hours, respectively; c) the average number of under-heated days was 2.9 in 90 days analyzed; d) two heating patterns dominated the sample; heating was tuned on only once (33%) or twice (51%) per day; e) the most common heating periods were identified (19/20:00 to 23:00) for single heating pattern and (6:00 to 9:00 and 16:00 to 21:00) for double pattern; f) the average temperature during single heating periods was 18.2°C in living rooms and 17.6°C in bedrooms. Dwellings presenting double heating periods achieved an average temperature of 17.5°C in living rooms and 17.0°C in bedrooms in the first heating period, and 19.0°C and 17.8°C in the second heating period, respectively; and g) the average living room temperature was found to be 1.0°C warmer than the average bedroom temperature. Also, Santamouris et al. [16] verified that the absolute energy use for heating purposes in all groups analyzed is quite low. It varied between 4 and 30 kWh/m2, with an average close to 18 kWh/m2. The time of use of the heating systems varied between 0.75 and 3h per day, while heating is necessary for much longer depending on the occupancy of the dwellings. In addition, Audenaert et al. [54] characterized the behaviour regarding energy use of 5 Belgium dwellings through surveys. The authors concluded that all heated their living rooms; 4 of them heated their bathrooms; and 3 heated their bedrooms. It was also roughly estimated that 2 heated more than 75% of the floor area, 1 heated 100%, 1 heated between 50% and 75%, and the other heated between 25 and 50% of the floor area. French et al. [18] reported that only 5% of New Zealand houses have central heating, with most houses only heating one or two rooms. Occupants tended to turn a heater on when they arrive, and off when they leave, or when the room is considered to be warm enough. As a result, temperatures that would be considered comfortable elsewhere in the temperate world were often not achieved. The most commonly heated room was the living room which was heated in the evenings in 90% of houses during weekdays, and in 87% of houses during the weekends. Only in 6% of houses did not heat the living rooms. Conversely in 50% of houses the bedroom was never heated, and 68% of houses did not heat utility areas (laundry, bathroom, corridor, etc.). Also, the average length of the heating season ranged from 8.6 months in the cooler far south and 5.5 in the warmer north. Approximately, 4% of the sample heated the entire year. Conversely, 3% of the houses did not heat at all, but these tend to be in the warmer 19 locations. The mean living room, and bedrooms temperature were the highest during the evening (17:00 to 23:00). The mean living room temperature droped from the evening to the night, again not surprisingly as only 18% of houses heated the living room at night (23:00 to 7:00). Only 15% of houses heated the bedroom during the night, but when coupled with the small heat gains from the occupants and appliances, the bedroom temperatures were closer to the living room temperatures overnight and during the morning. Finally, Burholt and Windle [11] examined a representative sample (N = 421) of older people (aged 70+) living in rural North Wales and concluded that, although only 1% of the respondents did not heat their living room, nearly 18% did not heat a second reception (dining) room. In addition, 16% of respondents did not use any heating in their bedroom. Over one-quarter of respondents did not heat a second bedroom, which may be due to infrequent use of the room. Nearly one-third (31%) of respondents did not heat the kitchen, however, it may be assumed that cooking appliances might increase the heat of the kitchen. Moreover, over one-third (34%) of respondents did not heat the bathroom. From the literature review, authors that characterized homes as ‘cold’ regardless of their geographical location were found, which makes empirical investigation of winter indoor temperatures and heating practices imperative. Even in cases where the mean indoor temperatures are high, there is a significant variation between dwellings. 2.2.3 Drivers of indoor temperatures Several studies scrutinize the driving forces behind indoor temperatures [10,14,16,19,68,115] during winter/heating seasons. The studies analyzed a variety of factors (e.g. climatic conditions, building characteristics and socio-economic factors), that may explain indoor temperatures [70]. For example, Critchely et al. [14], using binary logistic regression to model dwelling and household features, concluded that cold homes predominate in pre-1930 properties where the householder remains dissatisfied with the heating system. 20 Oreszczyn et al. [12] monitored indoor temperatures for a period of two to four weeks in over 1600 low income dwellings, and assessed the determinants of indoor temperatures through tabulation and regression methods. The authors concluded that temperatures were influenced by building characteristics (e.g., the age of construction, and the thermal performance of the building) and household features (e.g., the number of occupants, and the age of household´s representant). Also, Hutchinson et al. [19] investigated the extent to which low indoor temperatures in homes can be due to dwelling and household characteristics using tabulation and logistic regression methods. Data of low-income homes, from five urban areas in England, was analyzed. The authors concluded that property and household characteristics provide only limited justification for low winter indoor temperatures, presumably because of the influence of other factors including personal choice and behaviour. French et al. [18] verified that heating type, climate and house age are the key drivers for the living room temperatures. On average, houses heated by solid fuel are the warmest, and houses heated by portable liquefied petroleum gas and electric heaters are the coldest. Over the winter period, living rooms are below 20ºC for 83% of the time, and living rooms are typically the warmest areas. The relationship between aspects of building quality and indoor temperature has been previously quantified in the study of Haas et al.[117]. Authors registered higher indoor temperatures in more insulated dwellings. Another important factor was whether the heating system was centrally controlled and the surface area of the dwelling [70]. Mateo et al. [118] applied different machine learning techniques along with other classical ones for predicting the temperatures in different rooms. In addition, Kelly et al. [68] predicted indoor temperatures in English homes using panel methods. The model predicted average daily temperatures using both technical and social household variables, explaining about 45% of the variation in indoor temperatures. In particular, the number of occupants, household income and occupant age were found to be the most important drivers of indoor temperature. Finally, Santamouris et al. [16] found strong correlations between the minimum indoor temperatures and the level of thermal losses of the dwellings, and also between the income levels and the environmental and energy parameters, using regression analysis. 21 In summary, there is evidence in the literature that some inferences may be drawn when trying to identify drivers for the indoor temperatures. These, or the relevance of each, tend to vary with the geographical area and none was yet found for Portugal. 2.3 Modeling heating energy use The attention of researchers and experts in building energy performance has traditionally been focused upon a single building rather than on large building stock. This is shown by the increasing number of building thermal behaviour simulation tools on one side, and the increasing interest on certification procedures on the other side [119]. However, when the aim is the evaluation of the global achievable energy savings and Greenhouse Gases reduced emissions, it is also important to widen the focus to the building stock at a regional or national scale [119]. Several methods have been proposed to evaluate the specific energy use of a large building stock as well as of an individual building [120–124]. Commonly they are classified as top-down and bottom-up approaches. Top-down modeling approach starts with aggregate data and then disaggregate these down as far as possible in a bid to provide a comprehensive model [122]. Some examples include the studies developed by Dineen and Gallachóir [67] and Fabbri [59]. This thesis focused on bottom-up models. Typically these models comprise building physics modeling for calculating the energy usage of individual buildings and extrapolation of the results to a region or a country. The bottom-up approaches are mainly divided into statistical and engineering models [120]. The engineering models use physical principles to predict a building dynamic thermal performance [123]. This can be done in two different methods: a) by using simplified heatbalance equations [60,75,119,125–130]; and b) by using simulation tools [128,131–149]. 28 29 CHAPTER 3 CHARACTERIZATION OF THE STOCK AND THE ASSESSMENT OF THE ‘REFERENCE HEATING GAP’ Given the sparse data on the building stock, the existence of databases, such as EPBD- derived EPC databases, with the thermal performance characterization of a large number of buildings, construction periods and building typologies, enables gaining new insights relevant for several dimensions of policy assessment and policy design. This chapter explores the use of these databases to assess two main issues: 1) How does the thermal performance of the existing residential buildings stock vary with the year of construction in Portugal mainland?; 2) What is the difference between the ‘theoretical heating energy demand under stringent comfort conditions’ (THDstcc) ((1) in Figure 1, section 1.2) and the ‘actual energy use’ ((3) in Figure 1, section 1.2) for heating, i.e., the ‘reference heating gap’ for the existing residential building stock in Portugal mainland? From an energy management of a country or region point of view, these objectives provide a new perspective on the characterization of the thermal performance of the existing residential building stock versus the current individual analysis of the buildings. They will also give an indication of the relevant importance of the nature evolution of thermal performance vs the evolution triggered by regulations. The assessment of the existence and quantification of energy use gaps is also relevant for energy planning and policy making as they establish implications in the future energy demand and energy savings (see section 1.2). 30 The remainder of this chapter is structured as follows. Section 3.1 contextualizes the EPBD in Portugal and section 3.2 presents the characterization of the Portugal mainland residential building stock in terms of construction period. Section 3.3 characterizes the thermal performance of the residential building stock, whereas section 3.4 addresses the methodology behind the estimation of the theoretical energy demand for the existing residential building stock. Section 3.5 presents the estimation of the gap between theoretical energy demand and actual energy use for space and water heating. Finally, section 3.6 presents the main conclusions of the study developed under this chapter. 3.1 The Energy Performance Building Directive in Portugal The thermal performance of buildings, both in terms of energy demand and actual energy use, has attracted the attention of several social, industry and policy stakeholders. Most developed countries adopted regulations concerning this issue in the 20th century [201,202]. In Portugal, this was done through the regulation of the characteristics of thermal behaviour of buildings (RCCTE), Decree-law nº 40/90 in 1990 [203]. It was the first legal instrument to impose minimum standards on the thermal quality of the building envelope and it intended to achieve an ‘improvement of the comfort without additional energy use’. This was followed by the first regulation for energy systems and heating, ventilation and air conditioning (HVAC) systems in buildings (Decree nº 118/98) [204]. In 2000, the European Commission fostered the advance of energy efficiency in the building sector by publishing the thermal performance of building Directive (EPBD) in 2002 (Directive 2002/91/EC) [6]. This Directive proposes the adoption of structured methodologies for calculating the energy use in buildings, quality requirements for new and existing buildings, and the periodic inspection of boilers and air conditioning central systems. In addition, it requires the existence of an energy certificate of all buildings when undergoing a commercial transaction. In this regard, the directive changed the focus from new buildings only to the entirety of the building stock. 31 All European Union Member States require an energy performance certificate (EPC) when buildings are constructed, sold and rented. The EPC was considered a pioneering instrument that would overcome a deficit of information, hindering consumer interest in energy efficient dwellings [66]. The copies of all the EPC certificates issued both for new as for existing buildings are compiled in databases. The 2002 Directive was recast on May 2010 as 2010/31/EU Directive [7]. This recast Directive was published and adopted by the European Parliament and the Council of the European Union in order to further boost EU buildings´ energy efficiency following the EPBD. The 2002 European EPBD was transposed to the Portuguese legislation in 2006 through three Decrees: Decree 78/2006 created and defined the Portuguese thermal performance certification of buildings system (EPC) [205]; Decree 79/2006 updated the regulation for building energy systems and HVAC of buildings (RSECE) [206]; Decree 80/2006 updated the regulation on the characteristics of thermal behaviour of buildings (RCCTE) [72]. The overwhelming majority of residential buildings are covered by the first and the third ones, which will be closely referred in the present work. ADENE (the Portuguese Energy Agency) is the regulatory authority for building energy certification and energy efficiency under the supervision of the General Directorate of Energy and Geology (DGEG), and the Portuguese Agency of Environment (PAA) deals with issues related to the indoor air quality in buildings. The key objective for implementation was to save energy while ensuring comfortable indoor conditions and acceptable indoor air quality [207]. In turn, the recast EPBD (2010/31/EU Directive [7]) was transposed into Portuguese legislation in 2013 through the Law Decree 118/2013 [73]. 3.2 Characterization of the Portugal mainland residential building stock in terms of construction period In order to characterize the Portuguese residential building stock, ADENE’s National energy performance certification system (EPC)´s database was accessed in February 2012. The EPC´s database is the Portuguese database derived from the EPBD Directive that holds the energy 32 performance certificates issued, i.e., the Portuguese EPBD-derived EPC database. Since 2009, the ADENE has been compiling statistical data, aiming at characterizing different aspects related to the energy performance of the building stock. These include general aspects, (e.g., the distribution of the ratings of EPCs issued for new and existing buildings) and also detailed technical information (e.g., the average envelope characteristics for new construction in different decades) [208]. At that point of time, the EPC´s database only included certificates issued under the former RCCTE regulation [72]. Each certificate corresponds to an autonomous fraction (i.e., an apartment, detached or semi-detached dwelling). Both certificates for new and existing buildings, with or without heating or cooling systems were considered. Provisional certificates for new buildings still not completed were left out. The analysis covers the whole of the Portugal mainland municipalities, leaving out the buildings from the autonomous regions of Azores and Madeira that have autonomous and different databases. There were 259775 certificates able to be included in the analysis, which represent 5% of the total Portugal mainland residential building stock as of 2011. The National Institute of Statistics (INE), the entity responsible for ensuring the production and dissemination of official statistical information, designates the Portuguese residential building stock as the total of the usual residence, secondary residence and non-occupied autonomous fractions. It is important to introduce the main theoretical evaluation values of RCCTE used within the EPC. For each autonomous fraction (i.e., an apartment or a detached dwelling) under assessment, it is necessary to compute the theoretical energy demand for heating (HDRC) 3 and cooling (CDRC) 4 under reference conditions (in kWh/m2.year of ‘useful’ 5 energy), as well as theoretical energy demand for domestic hot water under reference conditions (DWDRC) 6 (in kWh/m2.year of ‘final’ energy). The calculation method is detailed in [72] and it follows the methodology of EN 13790 [209]. These are then integrated into a theoretical value of fossil 3 Named as ‘Nic’ in RCCTE regulation [72]. 4 Named as ‘Nvc’ in RCCTE regulation [72]. 5 ‘Useful energy’ is not what consumer buys (i.e., ‘final energy) but rather that from which the consumer derives benefits, after losses in the technical systems installed in the building have been taken into account [146]. 6 Names as ‘Nac’ in RCCTE regulation [72]. 33 ‘primary energy’ demand under reference conditions (PDRC) 7 (in kgoe/m2.year), as represented in Eq. 3.1 [210]. All of these evaluation values are computed considering reference operating conditions, which consider that buildings are kept at 20ºC during the whole heating season and at 25ºC during the whole cooling season, and also that each occupant uses 40l/day of domestic hot water at 60ºC. These four values must be lower than the reference limit. 𝑃𝐷𝑅𝐶= 0.1 × 𝐻𝐷𝑅𝐶 𝜂𝑖×𝐹𝑖+0.1×𝐶𝐷𝑅𝐶 𝜂𝑣×𝐹𝑣+𝐷𝑊𝐷𝑅𝐶×𝐹𝑎<𝑁𝑡=0.9×(0.01𝑁𝑖+0.01𝑁𝑣+0.15𝑁𝑎) 1 [kgoe/𝑚2.year] Eq. 3.1 where, ηi and ηv are the conversion efficiency from ‘useful’ to ‘final’ energy for heating energy demand and cooling energy demand, respectively. Fi, Fv and Fa are the conversion efficiency from ‘final’ to ‘primary’ energy for heating energy demand, cooling energy demand, and water heating energy demand, respectively. Ni, Nv, and Na are the reference limit of ‘useful’ heating energy needs, cooling energy needs and domestic water heating energy needs, respectively. Regarding the quality of the assessment performed, it has been shown that the values derived through the RCCTE methodology are correlated with those obtained through dynamic building simulation for the same buildings [211]. The main sources of disparities, during the evaluation of existing buildings, arise when some important characteristics are not known neither can be assessed in situ. In the absence of better information for a certain required parameter the experts must use the reference values, suggested by the energy certification system [72,210,212,213]. During the energy audit of an autonomous fraction, the performance values and other building characteristics are gathered together to build up a certification, collected afterwards into the EPC database. From each certificate included in the database, it is possible to extract information regarding the floor area (m2), year of construction, number of bedrooms, heating energy demand (HDRC), cooling energy demand (CDRC) and domestic water heating energy 7 Named as ‘Ntc’ in RCCTE regulation [72]. 34 demand (DWDRC), all expressed in kWh/m2.year, and the primary energy demand for heating, cooling and domestic hot water (PDRC), expressed in kgoe/m2.year. Other complementary data for Portugal mainland residential building stock were retrieved from the INE [214]. The INE carries out, on a regular basis, the Census, a large statistical survey on the Portuguese population and housing. The most recent one took place in 2011 (Census 2011) [214]. Figure 3 presents the total recorded Portugal mainland residential building stock as of 2011 (from INE´s database [214]) and Figure 4 shows the number of certificates (from EPC database), both per slot of construction period and number of bedrooms. Note that, in the EPC database, data after 2007 regards all buildings finished in a given year, and data before 2006 includes buildings that have undergone a commercial transaction (sale or rental) since 2009. This explains the high quantity of certificates in the last column of Figure 4. Figure 3. Portugal mainland residential building stock per construction period and number of bedrooms as of 2011 (from INE´s database [214]). 35 Figure 4. Number of certificates from EPC database per construction period and number of bedrooms. The Portugal mainland residential building stock was then re-characterized in accordance with the timescale and specific average floor area (per construction period and number of bedrooms) of EPC certificates sample. Figure 5 represents the total built area of the Portugal mainland residential building stock as of 2011 per construction period and number of bedrooms. 36 Figure 5. Portugal mainland residential building stock total built area per construction period and number of bedrooms. The results from Figure 3 make evident that, in terms of number of autonomous fractions, the predominant building slots are those from the decades of 1990 and 1980, followed by the decades of 1970 and 2000. However, analyzing the building stock in terms of total built area (Figure 5), the predominant slots are those from the decades of 1990 and 2000, followed by 1980 and 1970, indicating that apartments and dwellings have become larger. 37 3.3 Characterization of the thermal performance of the residential building stock 3.3.1 Evolution over time and hypothetical effects of regulations Currently, the analysis regarding the impact of the energy performance certification system becomes possible in the European Union. As the energy performance certification system was implemented several years ago, it is now possible to examine its impact [66]. Several studies have looked into EPBD implementation in different European countries showing a potential to increase energy efficiency (see, for instance, the study performed by Dascalacki et al. [215] in Greece, Amecke [65] in Germany, Tronchin and Fabbri [216] and Salvalai [217] in Italy, Ekins and Lees [218] in UK, Araùjo et al. [219] in Portugal, Murphy [66] in Netherlands and Gangolells et al. [220] in Spain). In addition, Casals [221] analyzed the building energy regulation and certification in Europe in terms of their role, limitations and differences. More recently, D´Agostino [222] provided an overview of the European status towards the implementation of nZEBs. Carpio et al. [223] determined the strengths and weaknesses of EPBD regulamentation in Europe by comparing opinions of 105 professionals. Fabbri [59] described how the energy performance certificates can be used as a value to measure features of their own building estates. Furthermore, Dall’O et al. [224] provided results of a benchmarking study on data from the energy cadaster of the Lombardy Region in Italy. The study identified key indicators on the energy performance of existing buildings, which became an effective tool for energy planning at local and regional scales. In order to try to detect the effects of regulations on the thermal performance of the Portugal mainland residential building stock, the average of the theoretical evaluation values (HDRC, DWDRC, CDRC and PDRC), collected from the EPC certificates sample, were plotted according to the year of construction. Only the most represented categories of the number of bedrooms (i.e., 2 to 5 bedrooms) of the residential building stock were considered. 44 Figure 12. Distribution of total built area by levels of theoretical cooling energy demand under reference conditions (CDRC, kWh/m2.year). Figure 13. Distribution of total built area by levels of theoretical domestic hot water energy demand under reference conditions (DWDRC, kWh/m2.year). 45 Figure 14. Distribution of total built area by levels of theoretical fossil primary energy demand under reference conditions(PDRC, kgoe/m2.year). The results from Figure 10 confirm that the recent buildings present better thermal performance as the biggest share of certificates are within lower levels of HDRC (heating demand in ‘useful’ energy). Most of the buildings constructed in 2011 have HDRC average values in the range of 61-100 kWh/m2.year, whereas the second group is in the range of 31-60 kWh/m2.year. Most buildings completed before 2000 have HDRC values higher than 100 kWh/m2.year. Figure 11 shows that the fraction of total built area with HDRC values lower than 50 kWh/m2.year is negligible, about 80% has heating demand higher than 100 kWh/m2 per year, and about 20% has heating demand higher than 200 kWh/m2 per year. Regarding the levels of CDRC (Figure 12, cooling demand in ‘useful’ energy), it can be observed that, after 1950, most of the certificates have levels of CDRC lower than 15 kWh/m2.year. The share of certificates with levels lower than 15 kWh/m2.year increased until 1991-00 and then decreased until 2008, remaining nearly constant since then. This is even more evident when observing certificates with values lower than 5 kWh/m2.year. 46 In terms of DWDRC values (Figure 13, demand for DHW in ‘final’ energy), it is observed an increase in the share of certificates with low reference values from buildings constructed from the 1980’s onwards, with some acceleration 2000 onwards. Finally, regarding the PDRC (Figure 14, primary fossil energy for heating, cooling and DWH), it is clear that the number of buildings with PDRC demand lower than 5 kgoe/m2.year have been increasing more expressively since the 1990’s. Nearly all buildings built after 2001 meet this condition. Considering the current policy objectives of achieving ‘nearly zero energy buildings’ adopted in the EBPD recast [7], it is worth mentioning that the absolute values of energy demand are still very high and have to significantly decrease in the near future if policy objectives are to be met (e.g., under the Passivhaus concept the building must be designed to have an annual heating demand of not more than 15 kWh/m² per year in heating, and 15 kWh/m² per year in cooling energy, OR to be designed with a peak heat load of 10 W/m² [226]). 3.4 Theoretical energy demand under reference conditions: A majorant for the country’s theoretical energy demand under thermal comfort conditions Providing an estimate of the theoretical energy demand to meet thermal comfort needs would be of great interest to the modeling and planning of countries’ energy systems. In this regard, data gathered in the previous sections of the present chapter is of most interest for this problem, in particular for space heating. It must be recognized that the theoretical evaluation values of RCCTE, computed under reference conditions (e.g., never less than 20ºC indoors during all winter, and never more than 25ºC during all summer) are most likely too stringent regarding the habits in residential buildings in Portugal and other countries. Nevertheless, at this stage, the theoretical evaluation values will be used to establish a ‘majorant’ objective (i.e., will be used as direct values) of the 47 theoretical energy demand under thermal comfort conditions for the residential building stock of the Portugal mainland (see section 1.2). Following this goal, the HDRC, CDRC and DWDRC values found in section 3.3.1 were multiplied by the respective total built area (m2). The total annual values of theoretical energy demand under reference conditions, resulted from the bottom-up assessment, are presented in Figure 15 (the respective equations in Appendix A). Figure 15. Total annual theoretical energy demand for heating (HDRC), DHW (DWDRC) and cooling (CDRC) under reference conditions (GWh/year), per construction period and number of bedrooms. The results show that the Portugal mainland residential building stock in 2011 is characterized by a total annual theoretical energy demand under reference conditions of 119469 GWh for space heating, 9129 GWh for space cooling and 46885 GWh for domestic water heating (Figure 15). To be noted that, as detailed in section 3.2, HDRC and CDRC are evaluated at the level of ‘useful’ energy, DWDRC is evaluated at the level of ‘final’ (or ‘delivered’) energy and PDRC is evaluated at the level of fossil primary energy. 48 The slots that represent a higher energy demand per construction period are those of the decades of 1971-00 (Figure 15). The absolute energy demand values for buildings constructed after 2006 represent only a small portion of the total. Overall, the 4 bedroom´s dwelling is the most representative in the total annual energy demand of the building stock. As the number of Portugal mainland autonomous fractions, as reported by INE, includes usual residence, secondary residence and non-occupied autonomous fractions, it would be worth disaggregating the annual theoretical energy demand for these three categories. The results of this procedure are presented in Table 4. Table 4. Annual theoretical energy demand under reference conditions of the Portugal mainland residential building stock as of 2011 disaggregated per usual residence, secondary and non-occupied residences. Theoretical energy demand under reference conditions Usual residence Secondary residence Non-occupied residence HDRC (GWh/year) 81044 23380 15045 CDRC (GWh/year) 6094 1775 1260 DWDRC (GWh/year) 31805 9105 5975 It is found that about 1/3 of the total annual theoretical energy demand of the residential building stock is due to secondary and non-occupied residences. It seems advisable to remove those from the figures when trying to compare energy values for heating from bottom-up vs. top-down approaches, as it will be the case of the next section. Also, the bottom-up values for the usual residence energy demand were computed now considering the residential building stock until the year of construction 2010 (ca. 3773956 occupied autonomous fractions [227]). The new annual theoretical values for HDRC, CDRC, and DWDRC, for Portugal mainland usual residence building stock in 2010 are 80313, 6006 and 31631 GWh/year, respectively. 49 3.5 Estimation of the gap of the residential building stock for space and water heating In this section, two main energy use gaps will be computed and named as ‘reference’ as the theoretical energy demand under reference conditions was used as the ‘majorant’ for thermal comfort conditions. In this case, the HDRC values were used for the space heating, and the DWDRC for the domestic hot water. 3.5.1 Top-down estimation of actual energy use for heating, cooling and domestic hot water After the bottom-up assessment of the annual theoretical HDRC, CDRC and DWDRC, it is now intended to obtain a top-down value from the national energy balance for comparison purposes. This can be done combining the statistics of the 2010 Portugal mainland national energy balance [228] with estimates of the 2010 breakdown of the energy use in the residential sector [227], as represented in Figure 16. Figure 16. Schematic representation of the methodology for the disaggregation of actual energy use per end uses (final or ‘delivered’ energy). 50 The total actual use of ‘final’ energy for space heating, domestic water heating and space cooling for Portugal mainland usual residence building stock in 2010 is presented in Table 5. The end uses for cooking, equipment and lighting were not computed as they are out of the scope of this work. The total values obtained are similar to those reported in [227], based on inquiries to 7468 Portuguese homes: 6289 vs 6194 GWh/year for space heating, 157 vs 151 GWh/year for space cooling, and 7046 vs 6464 GWh/year for water heating. Still, the first values were considered more appropriate for calculations as they came from a top-down disaggregation, while the latter ones came from a bottom-up collection [227]. Table 5. Actual energy use (‘final’) of the Portugal mainland usual residence building stock for heating, cooling and DHW in 2010, disaggregated per vectors and end uses. Actual energy use (‘final’ or ‘delivered’) Space heating (GWh/year) Domestic hot water (GWh/year) Space cooling (GWh/year) Electricity 897 230 157 Wood 4218 479 0 Natural gas 109 2157 0 Thermal solar 19 203 0 Heating oil 877 572 0 GPL 168 3405 0 Total (GWh/year) 6289 7046 157 The values of Table 5 are presented in terms of ‘final’ energy, but those of HDRC and CDRC of section 3.4 are presented in terms of ‘useful’ energy. It is necessary to convert the values presented in Table 5 to ‘useful’ energy to enable a comparison. Considering that ADENE’s EPC´s database of 2012 lack of data in what regards heating equipment’s efficiencies, average conversion efficiencies for heating were assumed from references: 100% for electricity, 87% for natural gas, 89% for GPL and heating oil and 40% for wood (resulting from a weighted mix between of open fireplaces and closed fireplaces) [214,229–231]. The actual ‘useful’ energy use values for space heating and space cooling for the Portugal mainland usual residence building stock in 2010 were thus estimated to be 3632 and 157 GWh/year, respectively. 51 3.5.2 Reference heating gap for space and domestic hot water Table 6 shows a comparison between the 2010 theoretical energy demand under reference (´stringent´) conditions computed in section 3.4, and the estimated energy actually delivered for the stock of usual residences of Portugal mainland in 2010 computed in section 3.5.1. It should be noted that space heating and cooling demand are compared in terms of ‘useful’ energy, whereas domestic hot water is compared in terms of ‘final’ energy. Table 6. Comparison between the annual theoretical energy demand under reference conditions (‘stringent’) and the estimated energy actually ‘delivered' for the stock of usual residences of Portugal mainland for 2010. End uses Theoretical energy demand under reference (‘stringent’) conditions Actual energy use Space heating demand (GWh/year) 80313 3632 Space cooling demand (GWh/year) 6006 157 Domestic hot water demand (GWh/year) 31631 7046 The ‘reference heating gap’ is defined as the percent difference between the theoretical heating energy demand under stringent thermal comfort conditions (GWH/year) and the energy actually used (GWh/year) for space heating, for the existing residential building stock, as expressed in Eq. 3.2 (see also Figure 1, section 1.2). A similar procedure can be applied to estimate the ‘reference heating gap’ for domestic hot water. The value for cooling is not computed because it has been demonstrated that indoor temperatures well above 25ºC can be compatible with thermal comfort according to the adaptive comfort standard [232]. 𝑅𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒 ℎ𝑒𝑎𝑡𝑖𝑛𝑔 𝑔𝑎𝑝= 𝑇𝐻𝐷𝑠.𝑡.𝑐.𝑐.−𝐴𝐸𝑈 ×100 [%] Eq. 3.2 𝑇𝐻𝐷𝑠.𝑡.𝑐.𝑐. where THDs.t.c.c. is the theoretical heating energy demand under stringent thermal comfort conditions (GWh/year) and AEU is the actual energy use (GWh/year). 52 Using the values represented in Table 6, the resulting value for ‘reference heating gap’ for space in 2010 is 95%, while the value for water heating for the same year was 78%. This means that the actual space heating usage in reference to the theoretical value under reference conditions for that year is only 5%, which is even lower than the 10% assumed in the calculation of the primary total energy demand in the RCCTE regulation [72]. The domestic water heating usage is 22% of the theoretical value estimated under reference conditions, very far from the 100% assumed in the same regulation. The differences found are very large, much beyond of what could be explained only by the lack of precision of the calculation procedure of the thermal performance assessment (either heating patterns (i.e., occupant behaviour), or other parameters assumed as reference values), or by the assumptions in the top-down breakdown of the national energy balance data. 3.6 Conclusions This chapter has performed a preliminary assessment of the ‘reference heating gap’ and characterized the thermal performance of the residential building stock in Portugal mainland, making use of the Portuguese EPBD-derived EPC database. Overall, this chapter showed that national databases of buildings energy certification schemes can be extremely useful in obtaining relevant insights on the thermal performance of the existing residential building stock. They can allow the identification of the most critical slots of building stock (e.g. in terms of type, age and region), and they can also assist analyzing to what extent the historical evolutions are justified by a natural phenomenon or by the effect of regulations. Furthermore, if combined with top-down energy balances, they can enable the estimation of the energy use gaps, which can be relevant both for health and energy policies. For the specific case of Portugal, the results show that the thermal performance of buildings progressively improved caused by a mix of accelerations pushed essentially by natural market evolution, spill-over effects (as that of the 1960’s) and regulations (as that after 2006). Despite the improvements, it is worth mentioning that the absolute values of theoretical heating energy demand computed under reference conditions (HDRC) for new buildings are far from the near- 53 Passivhaus levels implicit in the concept of near zero energy building required by the EPBD recast [7]. Results show that the majority (about 80%) of the building stock has theoretical HDRC values higher than 100kWh/m2.year, even in a ‘mild climate’. The work also compared the estimated theoretical energy demand under stringent thermal comfort conditions with the estimated actual energy use for space and water heating, for the Portugal mainland residential building stock in 2010. The percent difference between the theoretical and the actual energy use values is referred as ‘reference heating gap’. The results revealed gaps in 2010 of 95% and 78% for space and water heating, respectively. This means that occupants in Portugal mainland in 2010 used only 5% of the energy that they would need to maintain the whole dwelling at all time at a minimum of 20ºC during the winter season and, only 22% of the energy needed for one daily bath of 40 liters at 60ºC per person. Following the empirical evidence from other geographical contexts [20,46], and specially giving attention to the generalized Portuguese cultural aspect of not valuing thermal comfort, such differences (gaps) are probably in a large extent due to differences in occupant behavior. It is a fact that occupants do not require their homes to be permanently to a minimum of 20ºC during the whole winter/heating season, nor to a maximum of 25ºC during the summer/cooling season, neither require a bath of 40 liters of water at 60ºC per day [105,232], in order to achieve levels of thermal comfort satisfaction. Nevertheless, it does not mean that the gap is not associated still to a level of ‘deficit of comfort’. Actually, there are very few studies on determining whether the difference between the ‘theoretical’ and the ‘actual’ is part the expression of the difference between reference and actual values for heating patterns, or other parameters, or if it is also the expression of a ‘deficit of comfort’. The suspicion that there could be occupants living in poor indoor environment conditions in residential buildings in Portugal during the winter season, lead one to conclude that the gaps reported in section 3.5.2 may be also regarded as an indicator of the later. Further studies, involving statistically meaningful field monitoring of indoor temperatures, could contribute to provide with insights on this matter (chapter 4). Also, a more relaxed value of ‘theoretical heating energy demand under thermal comfort conditions’ ((2) in Figure 1, 1.2) is computed to reduce the gap that is explained by differences between reference and actual values for heating patterns (i.e, occupant behaviour). An estimation of the ‘heating gap’ is presented in chapter 5. 60 Table 8. Variables and its categories. Variable name Items inquired Categories Building characteristics Age of construction Age of dwelling´s construction [Until 1919-1945; 1946-1960; 1961-1970; 1971- 1980; 1981-1990; 1991-2000; 2001-2011] Apart./house Apartment or house [Apartment; house] Type dwelling Type of dwelling [Detached-house;semi-detached house; terracehouse; 1 façade to exterior] Apart. Position Position of the apartment in the building [Not applicable; ground-floor; first floor over garage/store; between floors; last floor] Wall thickness Thickness of external wall [<20 cm; 20-40cm; >40 cm] Air cavity wall Existence of air cavity on external wall [Yes; No] Wall insulation Existence of external wall insulation [No; yes, until 4cm; yes, over 4cm] Window frame Type of window frame [Aluminium; Wood; PVC] Type w. glazings Type of window glazing [Single; Double] Window orient. (bed./liv.) Bedroom/living room´s window orientation [North; East; South; West] Roof insulation Existence of roof insulation [No; yes, until 4cm; yes, over 4cm] Total area Total dwelling area (m2) [<30; 30-49; 50-79; 80-99; 100-119; 120-149; 150- 199; 200-249; 300-349; >349] Area (bed./liv.) Bedroom /living room area (m2) [0-15;16-20;21-25;26-30;>30] No. Bedrooms No. of bedrooms in the house [T0; T1; T2; T3; T4; T5; T6; T7; >T7] Category equip (bed./liv.) Categories of heating equipment existent in the bedroom/living room [None; Portable; Fixed] Type equip (bed./liv.) Type of heating equipment existent that supplies heat to the bedroom/living room [None; electrical radiator; thermoventilator; open fireplace; closed fireplace; air conditioner; gas boiler; diesel boiler; wood boiler; salamander; gas heater] 61 Table 8. Variables and its categories (continuation). Variable name Items inquired Categories Socio-economic factors Rep. school. Maximum schooling of household representants [Primary school; secondary education; degree; master; PhD] Household size No. of people living in the dwelling [0-2; 3-4; 5-6; 7-8] Household Type of household [Children, adults and older people; children and adults; adults (>14-65); adults and older people; children and older people] Min. occp. Day Minimum occupied period during the day [Always at home, except evening; always at home, except in the afternoon; always at home,, except in the morning; just in the evening; just in the afternoon; just in the evening; just in the morning; all day long] Profe. Situation Professional situation of active households [Mostly employed; half employed; no active households; all unemployed; all employed; mostly unemployed] Tenure Type of tenure [Landlord rented housing; cooperatives rented housing; Private housing with bank loans; Private housing with no bank loans] 1 Person bed. Just 1 person sleeping in the bedroom [Yes; No] Monthly net income Monthly net income (€/month) [0-350; 351-750; 741-1250; 1251-2000; 2001-3000; 3001-5000; >5000] Value comfort (bed./liv.) If occupants value comfort in bedrooms/living rooms [Does not value comfort; value comfort] Table B.1 and Table B.2, in Appendix B, report some of the characteristics of the sample analyzed based on the responses in the survey related to socio-economic aspects and building characteristics, respectively. From the responses gathered, it was possible to infer that the majority of the households analyzed were composed by 3 to 4 people (76%) and had all the active households employed (67%). Around 45% of the households were in the income range between 751 and 2000€ per month. The majority of the dwellings were houses (57%) and a high portion was constructed between 2001 and 2011 (32%) and in the decade of 1990 (28%). The presence of heating equipment varies a lot, being the electric radiator the most common 62 equipment (16%). Salamander (11%), open fireplaces (10%) and closed fireplaces (10%) were also commonly presented. Data related to heating patterns, adaptation strategies and thermal comfort preferences was collected from a detailed survey distributed to households at the end of the monitoring campaign (Appendix B). To evaluate the heating patterns the survey included questions related to the following aspects: length of heating period during the winter season, length of heating period during the day and type of heated rooms. To evaluate the thermal adaptation strategies households were asked about which adaptation measures were taken during the monitoring campaign. Households were also asked to scale their thermal comfort preferences in accordance to indoor temperatures. Finally, unmet but desired heating patterns were evaluated by asking households about their preferences on the length of the heating period during the winter season and during the day, as well as the type of rooms that they would like to have heated. In what regards heating patterns, namely the length of the heating period (in number of months) of the bedrooms, the majority of households (54%) mentioned 0 (i.e., ‘no heating’) or 1 month. However, for living rooms, a high percentage of the households (45%) responded 4 to 5 months. Also, concerning the length of the heating period during the day, around 49% of the households mentioned having heated the bedrooms for only less than half of the day and 37% mentioned not having heated the bedrooms at all. Table 9 and Table 10 present the length of the heating period for bedroom and living rooms during the winter season and during the day, respectively. 63 Table 9. Length of heating period during the heating period (in terms of number of months) for bedroom and living room. Length of heating (no. of months) Bedroom Living room 0-1 53.7% 28.6% 2-3 10.5% 8.8% 4-5 24.2% 45.1% 6-7 11.6% 17.6% Total 100.0% 100.0% Total answers 95 91 Table 10. Length of heating period during the day for bedroom and living room. Length of heating Bedroom Living room No heating 34.1% 17.4% Less than half of the day 48.8% 72.0% More than half of the day 8.5% 3.8% All day 8.5% 6.8% Total 100.0% 100.0% Total answers 129 132 Concerning the type of heated rooms, 33% of the households confirmed having heated the living room, bedrooms, kitchen and bathrooms; 12% only the living rooms; 8% the living room and bedrooms and other 8% the living room, the bedrooms and bathrooms (see highlighted column in Table B.3, Appendix B). In terms of unmet but desired heating patterns, 48% of the households answered that they would prefer to have heated the living room, kitchens, bathrooms and bedrooms, 12% living room, kitchen and bedrooms, and 12% just living rooms and bedrooms (see highlighted column in Table B.4, Appendix B). 29% of the households answered that they would prefer to have heated for longer periods during the winter season (see highlighted column in Table B.5, Appendix B) and 54% also admitted that they would prefer to have heated for longer periods during the day (see highlighted column in Table B.6, Appendix B). 88% of the households assumed to have embraced adaptation strategies, such as resorting to warmer blankets, clothing adjustment or hot drinks. Nevertheless, when asked to scale their 64 thermal preferences in accordance to the indoor temperatures registered during the monitoring campaign (see highlighted column, Table B.7, in Appendix B) the majority of households living in Porto (81%), Bragança (61%), Sabrosa (63%) answered that they would prefer warmer temperatures. In turn, the majority of households in Ponte de Lima (54%) were satisfied with the indoor temperatures. From those resorting to thermal adaptation strategies 68% pointed out their preference for warmer temperatures. 65 Table 11 shows a comparison between the sample analyzed and the values of Northern Portugal [239]. This table presents a subset of indications as these are the only data available in the National Institute of Statistics. According to this table, it is possible to verify that the sample of households selected for this study represents reasonably the dwellings of the Northern Portugal (see the similar values of % Northern Portugal and % sample reported on the table). Table 11. Comparison between sample and values of the Northern Portugal. % N.Portugal % Sample % N. Portugal % Sample No. of households (N=139) Tenure (N=136) Ponte Lima 11% 30% Rented to landlord 18% 18% Sabrosa 2% 19% Rented to state/cooperatives 3% 10% Porto 76% 29% Pay provision to the bank 30% 30% Bragança 11% 22% Own housing 49% 42% Building age (N= 107) No. of bedrooms (N=112) Until 1919 to 80 41% 9% 0 bedrooms 0% 0% 1981-90 19% 12% 1 bedroom 2% 4% 1991-00 23% 36% 2 bedrooms 8% 7% 2001-11 16% 42% 3 bedrooms 26% 15% Apartment or house (N=137) 4 bedrooms 34% 24% Apartment 45% 41% 5 bedrooms 14% 16% House 55% 59% 6 bedrooms 7% 15% 7 bedrooms 4% 7% >7 bedrooms 4% 12% Household annual net income (€/year) (N=116) Type of equipment (N=101) 9634 20% 20% Ar conditioning ---- 1% 14800 20% 32% Gas boiler 0% 10% 19061 20% 23% Diesel boiler 0% 5% 25770 20% 16% Wood & salamander boiler 5% 25% 49539 20% 9% Open fireplace 19% 14% Closed fireplace 8% 14% Employment status (N=134) Eletric radiator & thermoventil. 31% 27% None active 27% 3% No heating 12% 5% Half employed 7% 18% Mostly employed 3% 1% All employed 56% 70% Mostly unemployed 1% 1% All unemployed 5% 7% 66 4.2 Characterization of indoor temperatures 4.2.1 Overview of daily mean outdoor and indoor temperatures Porto has a Mediterranean climate, in particular, winter outdoor temperatures typically range between 5ºC and 14ºC, rarely dropping below 0ºC. It is also typical during this season that rainy periods alternate with cooler days and clear skies [240]. Ponte de Lima, located at north of Porto, but still near the Atlantic coast, presents similar outdoor temperature ranges [241]. In turn, Sabrosa is situated in the interior of the country and surrounded by mountains and presents a sharper winter with temperatures ranging from 4ºC and 10ºC [242]. In Bragança, the winter is long, cold and wet with mean min values frequently below 0ºC and mean high values of 10ºC [243]. Table 12 presents the typical daily mean outdoor temperature according to the historical records [235,236] and to the values for the four locations gathered from the monitoring campaign. It can be observed that the monitored winter of 2013-2014 was slightly milder than typical years. The difference is particularly evident in terms of the mean low temperature (e.g., in Bragança, it was over the average by +2.9ºC). Table 12. Daily mean outdoor temperature records and monitored values for the four locations. Locations Winter period Mean high (ºC) Daily mean (ºC) Mean low (ºC) Porto Typical year 14 10 6 Monitored period 14 11 8 Ponte de Lima Typical year 14 10 7 Monitored period 14 10 8 Sabrosa Typical year 11 7 4 Monitored period 9 7 5 Bragança Typical year 10 5 1 Monitored period 13 8 4 67 In what concerns the daily mean indoor temperatures, there is no typical value for the dwellings analyzed. The daily mean bedroom temperatures for the occupied period, were higher than 20ºC in only 4% of the dwellings, between 18ºC to 20ºC in 6% of dwellings, between 14ºC and 16ºC in 31% of the dwellings, and below 14ºC in 35% of the dwellings (with 2% of the dwellings exhibiting temperatures lower than 10ºC). The daily mean indoor temperatures for the 24h period present a similar distribution. The daily mean living room temperatures in the 24h period were maintained in the range of 14ºC to 16ºC for 29% of dwellings, and between 16ºC and 18ºC for 27% of dwellings. In 20% of dwellings, temperatures were above 18ºC (with 6% above 20ºC). There are still 24% of households with living room temperatures below 14ºC, from which 5% were in the range 10ºC and 12ºC, and 1% reaching less than 10ºC. Regarding the occupied period, the large shares lie between 16ºC and 18ºC (26% of dwellings), 14ºC and 16ºC (24% of dwellings), 18ºC and 20ºC (18% of dwellings) and above 20ºC (14% of dwellings). Still, there are dwellings registering daily mean temperatures between 10ºC and 12ºC (4% of dwellings) and lower than 10ºC (1% of dwellings). 68 Table 13 presents an overview of the daily mean indoor temperatures for bedroom and living room, for the occupied and 24h period. Overall, the results from Table 13 show that indoor temperatures are significantly below the levels putatively recommended by the WHO below (21ºC at living rooms, and 18 ºC in the other occupied rooms) [244] . Table 13. Daily mean indoor temperatures for bedroom and living room. Bedroom Living room Sabrosa Bragança Pt. de Lima Porto All locations Sabrosa Bragança Pt. de Lima Porto All locations 24h period Mean daily 14 16 14 16 15 16 17 15 16 16 Standard deviation 3.3 3.0 2.9 2.3 3.1 3.9 3.0 3.0 2.4 3.2 Mean high 15 17 15 17 16 19 19 16 17 18 Mean low 12 15 13 15 14 14 16 14 16 15 30% Percentile 12 14 12 15 13 14 15 13 15 14 70% Percentile 15 18 16 17 16 18 19 16 17 17 Occupied period Mean daily 14 16 14 16 15 18 17 15 17 17 Standard deviation 3.4 3.0 3.0 2.4 3.1 4.4 3.1 3.3 2.4 3.4 Mean high 15 17 15 17 16 19 18 16 17 17 Mean low 13 15 13 16 14 16 17 15 16 15 30% Percentile 12 14 12 15 14 15 16 14 16 14 70% Percentile 15 18 16 17 16 20 19 17 18 18 It should be noted that almost 30% of the time, households in the two poorest locations (Ponte de Lima and Sabrosa) have bedroom temperatures lower than 12ºC even in the occupied period. A second major conclusion withdrawn from Table 13 is that living rooms are usually warmer than bedrooms by a difference of 1.7ºC in the occupied period and by 1ºC in the 24h period. Actually, in contrast to 24% of the households, 42% have affirmed that there is no need to have temperatures higher in the bedroom than in the living room. From Table 13, it is also evident that there is no considerable difference between the daily mean bedroom temperatures in the occupied (15ºC) and 24h period (15ºC). However, for the living room, the difference is almost 1.0 ºC between the two different periods. For example, in Sabrosa, the difference in living room temperatures is quite evidential (18ºC and 16ºC for the occupied and 24h period, respectively). In addition, the mean living room temperature at the 69 occupied period is, on average, 3.7ºC warmer than bedroom temperatures, which can be an indication that those rooms are highly valued during that period. A paired sample t-test was conducted in order to check if there are statistically significant differences between the daily mean indoor temperatures of ‘all locations’ (see seventh and twelfth columns in Table 13) in the living rooms and in the bedrooms, during the occupied and 24h period. Table 14 presents the results obtained in the four paired sample t-tests conducted. The results indicated that the differences are statistically significant (p-values < 0.0001). This suggests that a separate analysis should be done for mean bedroom temperature in the occupied period and 24h period, and for the mean living room temperature in the occupied period and 24h period. Besides the existence of a significant difference between daily mean bedroom temperature in the occupied and 24h period, the focus will be on the occupied period hereafter because it is believed that a mean difference of 0.155 in the Celsius scale might not have an impact on the thermal comfort of occupants. Results from Table 14 also indicate that the difference between mean indoor bedroom and living rooms´ temperatures is significant either in the 24h period or during the occupied period. Table 14. T-tests. T Df p-value Mean Difference Bedroom (24h vs occ. period) -8.978 140 .000 -.155 Living room (24h vs occ. period) -9.831 139 .000 -.750 24h period (bed. vs liv.) -6.383 139 .000 -1.05 Occupied period (bed. vs liv.) -7.624 139 .000 -1.65 Comparing the mean indoor and outdoor temperatures, it is possible to infer that lower outdoor temperatures do not necessarily imply lower indoor temperatures. For example, when comparing the daily mean temperature values from Table 12 and Table 13, it can be observed that although Bragança and Porto present similar indoor temperatures, the outdoor temperatures in Bragança are lower than in Porto. Other factors may have therefore 76 Table 16. Indoor neutral temperatures according to Humphreys and Nicol model [87], and the percentage of time that these are achieved or exceeded in each location for bedroom (Bed.) and living rooms (Liv.). In the points marked with an *, the model was applied beyond the declared applicability range in terms of outdoor temperatures. Sabrosa Bragança Ponte de Lima Porto T (ºC) Bed. Liv. T (ºC) Bed. Liv. T (ºC) Bed. Liv. T (ºC) Bed. Liv. 24h period December 15* 34% 63% 16* 55% 59% 18* 12% 22% 18* 23% 30% January 17* 22% 46% 17* 51% 62% 19 7% 16% 19 14% 20% February 17* 14% 37% 17* 40% 54% 19 4% 11% 19 11% 15% Occupied period December 15* 50% 76% 16* 58% 68% 18 14% 29% 18 24% 35% January 17* 24% 62% 17* 53% 73% 19 8% 22% 19 18% 23% February 17* 16% 54% 17* 42% 66% 19 4% 17% 19 13% 18% Depending on the month, the neutral temperatures ranged between 15 and 17ºC, 16 and 17ºC, and 18 and 19ºC for households living in Sabrosa, Bragança and in Ponte de Lima and Porto, respectively. Results indicated that indoor temperatures are far from the neutral temperatures (the closest are the living rooms at occupied period, especially in Sabrosa and Bragança) based on the Humphreys and Nicol (1998)´s adaptive comfort model [87]. Though, this is a general model that can only be considered for indicative purposes. The cultural aspect may be a key factor when considering the adaptation of Portuguese households to indoor temperatures. 4.2.4 Identification of heating patterns Analyzing the mean temperature in each hour for the period studied, along with the responses to surveys regarding the length of heating period during the day (based on Table 10), it was possible to infer about the heating patterns for each dwelling. Table B.8 and Figure B.2, in Appendix B, present the hourly mean bedroom temperature distribution for a sample of dwellings (i.e., 7 dwellings), the responses to surveys, and the corresponding inferred heating patterns. The distributions provided information on the hourly mean temperature patterns during the day. In the case of unclear patterns, the heating pattern 77 was supported by the responses to surveys. In the case of ambiguous patterns with no responses, dwellings were not classified with a heating pattern (i.e., ‘non-identified’ pattern). Taking this process to the 141 dwellings analyzed, Table 17 illustrates the heating patterns inferred (‘no heating’; heating all day’, ‘heating during evening’, ‘heating all night’ and ‘other’) and the corresponding share of each heating pattern. Table 17. Distribution of heating patterns per location of study for bedroom and living rooms. Heating patterns inferred Bedroom Sabrosa Bragança Pt. Lima Porto Living room Sabrosa Bragança Pt. Lima Porto No heating 33% 44% 10% 44% 33% 18% 11% 0% 21% 32% All day 9% 4% 21% 0% 11% 7% 0% 18% 0% 12% During evening 49% 40% 55% 56% 42% 71% 85% 79% 76% 47% All night 9% 12% 14% 0% 14% 3% 0% 4% 0% 9% Other 0% 0% 0% 0% 0% 2% 4% 0% 2% 0% Total 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% N inferred 129 25 29 39 36 131 27 28 42 34 Non identified 12 9 For the bedrooms, the largest share of households (49%) exhibits the ‘heating during the evening’ pattern. This means that at least 49% of the households have heated their bedroom for some time just before going to bed. The ‘no heating’ pattern is also common for a high number of households (33%). For a ‘free-running’ building, indoor temperatures are a consequence of heat gained/stored from solar heating and occupants, offset by heat lost through the building fabric and by infiltration [10]. For the living rooms, it is possible to conclude that the share of the ‘no heating’ pattern (18%) is lower than the bedroom patterns, but the inverse occurs to the share of ‘heating during the evening’ pattern (71%). This is a clear indication that households heat more often the living rooms than the bedrooms. 53% of the households that have heated the living room only during the evening actually occupied their homes only at that period of time. The problem lies when there is still a great share of households (47%) with that heating behaviour but has at least one person staying home all day when no living rooms are being heated (see Table B.1, in Appendix B). 78 Regarding the expression of each heating pattern, for the bedrooms, per location, (presented in Table B.9, in Appendix B), Bragança is the location with the lowest share of ‘no heating’ pattern (7%), in contrast with Ponte de Lima (40%). Actually, Bragança is the location where more households heat their bedrooms all day long, followed by Porto. Similar conclusions can be applied to the heating patterns that characterize the living rooms. 4.3 Modeling indoor temperatures 4.3.1 Description of the statistical models Three models were developed for the actual indoor temperatures of the residential building stock in Northern Portugal, taking the form of what is usually known in statistic sciences as ‘prediction models’. Most building stock models would benefit from robust estimates of indoor dwelling temperatures [68]. In particular, the models will predict three models: the daily mean bedroom or living room temperatures for the occupied period, and the daily mean living room temperatures for the 24h period. Beyond the prediction purposes, the models will enable users to identify and quantify the importance of socio-economics factors, building characteristics and climate factors to explain indoor temperatures. Following the approach proposed by Kelly et al. [68], the models developed were based on linear regression with panel-corrected standard errors. In particular, this statistical model enables users to account for the panel nature of the data gathered from the monitoring campaign, which covered 141 dwellings during a period of 94 days. It presents important benefits for modeling indoor temperatures over the standard linear regression as it enables users to control for heterocedasticity (each unit has its own variance) and contemporaneous correlation across units (each pair of units has its own covariance), instead of assuming that the errors are independent and identically distributed. 79 The general form of the model developed [253] can be presented as follows: 𝑇𝑖𝑡=𝛼0+𝛽1(𝑉1)𝑖𝑡+ 𝛽2(𝑉2)𝑖𝑡+ …+ 𝛽𝑘(𝑉𝑘)𝑖𝑡+𝜀𝑖𝑡 [ºC] Eq. 4.2 where Tit is the daily mean indoor temperature associated with unit (i.e., dwelling) i (i=1,…,n), in day t (t=1,…,s). It corresponds to the mean of the bedroom or living room temperature over 24h or over the occupied time period for a specific dwelling i, in a specific day t. Vk represents the variables (k=1, …,k), related to socio-economic factors, building characteristics and climatic conditions (i.e., outdoor temperatures), and βk represents the parameters to be estimated (i.e., coefficients of each variable); α is a constant term; 𝜀𝑖𝑡 is a disturbance that may be autocorrelated along t or contemporaneously correlated across i [253]. To run model (Eq. 4.2), STATA11 software was used, invoking the command xtpcse [253] using the correlation(psar1) option, which considers panel-specific autocorrelation. Accuracy metrics evaluate the performance of a model by comparing the observed with predicted values. Three accuracy metrics widely used were selected [200,254]. These are the mean absolute error (MAE), the root mean square error (RMSE) and the coefficient of determination (R2). MAE is most commonly used and easiest to interpret directly, representing the average of the absolute errors [118,158,170,255,256]. RMSE [68,161,168,170,170,257] is also very common and it exaggerate the presence of outliers. Both accuracy metrics depend on the scaling of the variables, which may be inconvenient if the criteria are used for comparing predicting accuracy across different variables [258]. Both MAE and RMSE avoid the negative values to cancel the positive ones and both have the same units as the quantified plotted. 80 The accuracy metrics MAE and RMSE can be calculated using Eq. 4.3 and Eq. 4.4, respectively. 𝑀𝐴𝐸= ∑‖𝑦𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑−𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑‖ 𝑛 𝑖=1 𝑖=1,2,…𝑛 [º𝐶] Eq. 4.3 𝑛 𝑅𝑀𝑆𝐸= √∑[𝑦𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑−𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑]2 𝑛 𝑖=1 𝑛 𝑖=1,2,…𝑛 [º𝐶] Eq. 4.4 where, yobserved is the observed values and ypredicted is the predicted values and n stands for the total number of observations. The coefficient of determination (R2), which is a default outcome of the statistical analysis performed in STATA11 software, was also used and provides information about the goodness of fit of a model. R2 is a figure between 0 and 1, is a measure of effect size and it represents the proportion of the variation in the dependent variable that is attributable to the explanatory variables [116]. The higher the R2 and the smaller the error values are, the better the model is to predict the actual measurements [68]. 4.3.2 Description of database used in statistical models The independent variables considered for the statistical analysis are classified in three primary groups: a) climatic conditions, captured through the outdoor temperatures; b) socioeconomic factors, aiming to depict household characteristics; and c) building characteristics, reflecting the physical characteristics of the dwelling and heating systems. 81 In order to capture the complexities inherent within the residential building stock, a database that contains as much information as possible on the three groups is needed [68]. The socio-economic factors and building characteristics dataset was gathered through the survey (see section 4.1.3). These variables and its categories are listed in Table 8. The outdoor temperatures resulted from the undertaken monitoring campaign (section 4.1.2). The daily mean outdoor temperatures dataset considered a range of values from 1 to 14ºC for the three different predicting models. Before applying the linear regression with panel-corrected standard errors, an initial screening was conducted to assess the potential impact of each variable in the indoor temperature. The screening involved two main steps. The first concerned the analysis of the sample size for each variable collected through the surveys. The second step was the assessment of multicolinearity among the subset of variables selected from the first step. In the first criterion, it was established that the size of each variable (N) should be greater than 70, which guarantees that approximately 50% of the responses for each variable are available. This implied the exclusion of the following variables: total area; bedroom and living room´s area; and roof insulation. In the second step, multicolinearity was analysed using correlation coefficients. The correlations between nominal variables (and between nominal and ordinal variables) were checked using the Cramer´s V coefficient, and the correlations between ordinal variables were verified using the Spearman correlation coefficient. Table B.10, in Appendix B, presents the correlation coefficients between all variables. In this step, it was decided to exclude from the analysis variables with correlation coefficients higher than 0.6 in order to avoid problems with the estimations. Table 18 presents the final set of independent variables that constitute the database used in the statistical models and also the type of variables. 82 Table 18. Set of independent variables included in the prediction models. Variable name Type of variable Bedroom Living room Household Nominal X Household size Ordinal X Monthly net income Ordinal X Professional situation Nominal X Value comfort (bed./liv.) Binary X X Apart/house Nominal X X Age construction Ordinal X X Wall insulation Ordinal X Wall thickness Ordinal X Window frame Nominal X X Window orientation (bed./liv.) Nominal X X Type equipment (bed./liv.) Nominal X X Tout Continuous X X 4.4 Results and discussion The results obtained for the three predicting models developed are reported in Table 19. In particular, Table 19 reports the estimates from the linear regression with panel-corrected standard errors (i.e., the coefficients, standard errors, and p-values), the number of observations included in each model, the coefficient of determination (R2), the significance (Sig.), and the accuracy metrics MAE and RMSE (see section 4.3.1). The prediction models were found to be statistically significant (χ2 test with p-value < 0.0001). The model for the bedrooms explains 89% of the variability of indoor temperatures in the occupied period. The models for the living rooms in the occupied and in the 24h period explain 91% and 90% of the variability of indoor temperatures, respectively. These values are within the range of results found in Kelly et al. [68] who achieved R2 values between 0.45 and 0.88 when applying linear regression with panel-corrected standard errors. In addition, when analyzing the accuracy of the models it can be observed that models present 83 relatively low errors. In particular, MAE is approximately 1ºC for all the models, and RMSE varies between 1.3 to 1.6ºC, depending on the model. Also, in order to explore the face validity of the panel regression results, a standard linear regression was estimated. The results obtained are in line with panel regression results. In particular, the values of R2 were 0.82, 0.78 and 0.80 for bedrooms in the occupied period, and living rooms in the occupied and 24h period, respectively. Nevertheless, the results of the panel regression (see R2 values in Table 19) are more accurate, indicating that it is a better option for modeling indoor temperatures. From Table 19 it is possible to conclude that socio-economic factors and building characteristics, as well as climatic conditions affect significantly the bedroom and living rooms indoor temperatures. It is worth mentioning that approximately 73% of the variability presented in bedroom temperatures in the occupied period is explained by the building characteristics (i.e., age of construction; apartment/house; wall insulation; window frame; window orientation and type of equipment). Socio-economic factors (i.e., type of household; no. of households; comfort value and professional situation of the active households) are able to explain 14% of the bedroom temperatures in the occupied period, whereas the outdoor temperatures only explain 3%. In addition, approximately 85% of the variability of living room temperatures for the occupied period, are explained by the building characteristics (i.e., age of construction; apartment/house; wall thickness; window frame; window orientation and type of equipment), 4% by the socio-economic factors (i.e., type of household; no. of households; comfort value and monthly net income) and 1% are explained by the outdoor temperatures. Similarly, the indoor living room temperatures in the 24h period are mainly explained by building characteristics (81%) and socio-economic factors (7%), followed by outdoor temperatures (2%). These results are in line with several studies that highlight the importance of the building characteristics as one of the main factors influencing indoor temperatures [12,16,19,117,259]. 84 Table 19. Results of the linear regression with panel-corrected standard errors models. Variable name Bedroom occupied period Living room occupied period Living room 24h period Coef. Std. Err. P-value Coef. Std. Err. P-value Coef. Std. Err. P-value Tout 0.0896 0.01666 0.000 0.137 0.01513 0.000 0.125 0.02021 0.000 Age of construction 1946-1960 (base) (base) -4.598 0.7379 0.000 1961-1970 5.326 0.8609 0.000 9.287 0.6920 0.000 4.307 0.5006 0.000 1971-1980 1.136 0.3324 0.001 3.535 0.9774 0.000 -5.071 0.6232 0.000 1981-1990 5.098 0.5636 0.000 3.997 0.4980 0.000 -1.757 0.3665 0.000 1991-2000 5.131 0.6367 0.000 8.752 0.5967 0.000 1.942 0.2781 0.000 2001-2011 8.501 0.8187 0.000 6.663 0.6758 0.000 (base) Apart./House Apartment (base) (base) (base) House -5.875 0.4108 0.000 2.396 0.9691 0.013 5.284 1.0460 0.000 Wall thickness <20 cm (base) -3.534 0.3222 0.000 20-40 cm 1.630 0.3144 0.000 -3.537 0.1992 0.000 >40 cm 5.167 0.3963 0.000 (base) Wall insulation None (base) Until 4 cm 2.291 0.2621 0.000 Over 4 cm 3.564 0.6996 0.000 Window frame Aluminium (base) (base) (base) Wood 3.016 0.5432 0.000 -6.806 0.3343 0.000 -4.785 0.3177 0.000 PVC 2.262 0.4393 0.000 -0.663 0.3408 0.052 -0.794 0.2641 0.003 Window orientation (bed./liv.) North 2.236 1.0823 0.039 (base) -3.203 0.3476 0.000 East 2.815 0.5663 0.000 2.728 0.7274 0.000 -0.928 0.4005 0.021 South -2.022 0.3982 0.000 -0.563 0.2845 0.048 -3.158 0.3474 0.000 West (base) 1.595 0.4274 0.000 (base) Type equip (bed./liv.) No equipment (base) -6.376 0.4613 0.000 (base) Electrical radiador 4.664 0.4100 0.000 -3.815 0.8939 0.000 5.159 0.5693 0.000 Thermoventilator 6.198 0.7039 0.000 -1.907 1.1433 0.095 6.724 0.9439 0.000 Open fireplace --- --- --- -9.159 0.3863 0.000 -2.124 0.3858 0.000 Closed fireplace 7.386 0.6051 0.000 -7.701 0.6677 0.000 -3.2763 0.8808 0.000 Ar condicioner -8.313 0.6024 0.000 --- --- --- --- --- --- Gas boiler -1.227 0.9768 0.209 -3.169 0.9519 0.001 4.769 0.6553 0.000 Diesel boiler 2.073 0.9252 0.025 --- --- --- --- --- --- Wood boiler 5.397 0.4160 0.000 -5.833 0.7720 0.000 -1.742 1.0306 0.091 Salamander 3.080 0.7353 0.000 (base) 3.427 0.4480 0.000 85 Table 19. Results of the linear regression with panel-corrected standard errors models (continuation). Variable name Bedroom occupied period Living room occupied period Living room 24h period Coef. Std. Err. P-value Coef. Std. Err. P-value Coef. Std. Err. P-value Household Children and adults and older people (base) (base) (base) Children and adults 6.515 0.9178 0.000 5.794 0.8992 0.000 8.332 0.9200 0.000 Adults 8.127 0.8193 0.000 4.055 1.0988 0.000 7.551 1.0826 0.000 Adults and older people 11.681 1.4674 0.000 9.117 1.1087 0.000 10.614 1.0596 0.000 Household size 0-2 (base) 3-4 2.190 0.5332 0.000 (base) (base) 5-6 8.846 0.7153 0.000 -2.405 0.2833 0.000 -1.579 0.2727 0.000 7-8 5.923 1.0097 0.000 2.509 0.7386 0.001 0.926 0.8149 0.256 Comfort value (bed./liv.) No value (base) (base) (base) Value -2.652 0.4057 0.000 3.168 0.4952 0.000 0.687 0.5041 0.173 Profess. Situation Half employed 5.790 1.0357 0.000 No active households 11.799 1.5589 0.000 All unemployed 6.033 0.9501 0.000 All employed 5.411 1.0766 0.000 Mostly unemployed (base) 0.000 Monthly net income 351-750 €/month (base) (base) 751-1250 €/month 3.465 0.4445 0.000 1.477 0.3838 0.000 1251-2000 €/month 4.345 0.6100 0.000 5.565 0.5963 0.000 2001-3000 €/month 4.748 0.6315 0.000 5.133 0.6327 0.000 3001-5000 €/month 5.843 0.6879 0.000 5.290 0.6583 0.000 Const -6.285 1.9628 0.001 -0.148 1.488 0.921 4.746 1.0981 0.000 R2 0.892 0.910 0.897 Sig. 0.000 0.000 0.000 No. observations 3630 (41 dwellings) 3321 (37 dwellings) 3322 (37 dwellings) MAE (ºC) 1.12 1.02 0.99 RMSE (ºC) 1.51 1.55 1.30 92 between the reference energy demand, indoor temperature, and the heating energy use was known. For example, given the reference energy demand and a certain indoor temperature intended, it would be possible to estimate the heating required; Or, given the maximum amount of heating than can be afforded, it would be possible to estimate the resultant indoor temperatures. There is a constant search for user-friendly models to predict heating energy use in the residential buildings with a greater degree of freedom in terms of assumptions for the occupant behaviour (i.e., operating conditions) [63]. Furthermore, databases, such as the EPBD-derived EPC databases, are very attractive tools to estimate heating energy use (and indoor temperatures) as they hold an extensive number of certificates already issued with important information on building data (e.g., theoretical heating energy demand under reference conditions (HDRC)). Provided that there is an understanding of the relationship between the heating energy use, occupant behavior (e.g. indoor temperatures) and HDRC, and using the HDRC values from EPC databases, it is possible to estimate heating energy use or indoor temperatures values for different levels of occupant behaviour. Dynamic modeling of individual dwellings can estimate accurately the heating energy use, but it is complex and time-consuming. This chapter explores the use of statistical models coupled with simulations and calculation data to predict heating energy use or indoor temperatures. The models are applicable to residential buildings in any geographical context and in the Portuguese context. In addition, this chapter proposes the use of the models developed to assess the ‘heating gap’ of the residential building stock in Portugal mainland (once the estimated ‘reference heating gap’ value, in chapter 3, was based on stringent thermal comfort values). The developed models in this chapter are useful tools for policy formation [260] and an important contribution to the modeling and planning of countries’ whole energy systems. The structure of this chapter is the following. Section 5.1 gives a brief contextualization of the modeling concept and section 5.2 describes the architecture of the developed models. Section 5.3 describes the databases used in the statistical models. Section 5.4 presents the statistical models adopted. Section 5.5 presents the models developed with values of HDRC that 93 come directly from building simulation with ESP-r, and therefore in principle can be used for any geographical context – hence the label ‘universal’. Section 5.6 uses values of HDRC computed according to the methodology of the Portuguese energy certification system, and therefore are labelled as ‘Portugal-specific’. Section 5.7 presents the graphical representation of the energytemperature relationship, whereas section 5.8 exhibits the assessment of the ‘heating gap’. Section 5.9 presents the main conclusions of the study developed under this chapter. 5.1 Modeling concept Each certificate, representing a specific building, issued in energy rating/certifications schemes´ databases, can provide HDRC values. In chapter 3, the HDRC served as ‘majorant value’ for theoretical heating energy demand (THD) under thermal comfort conditions, but aligned with a stringent perspective of thermal comfort ((3) in Figure 1, section 1.2) to support the preliminary assessment of the ‘reference heating gap’ of the residential building stock. In the Portuguese case, the HDRC values are calculated as ‘useful’ energy values. The HDRC, of a particular building, is dependent on several factors, such as physical characteristics of the building archetype (M); geographical location of the building (L); reference heating pattern (HPatref); reference indoor heat gains (HGref) and reference set point temperature (Tspref). This relationship can be expressed, in general, by Eq. 5.1 as follows: 𝐻𝐷𝑅𝐶=𝑓(𝑀,𝐿,𝐻𝑃𝑎𝑡𝑟𝑒𝑓,𝐻𝐺𝑟𝑒𝑓,𝑇𝑠𝑝𝑟𝑒𝑓) [𝑘𝑊ℎ/𝑚2.𝑦𝑒𝑎𝑟] Eq. 5.1 The heating energy use (HEU) of a particular building, which is also calculated as ‘useful’ energy, depends on the building characteristics, climate conditions, and occupant´s heating behaviour [12,17,29,54,63,117,261–264]. More precisely on: geographical location of the building (L); physical characteristics of the building archetype (M); heating pattern (HPat); indoor heat gains (HG) and set point temperature (Tsp). The latter three are defined by the occupant´s behaviour. The heating energy use can be therefore expressed, in a simple manner, by the relationship described in Eq. 5.2. 94 𝐻𝐸𝑈=𝑓(𝑀,𝐿,𝐻𝑃𝑎𝑡,𝐻𝐺,𝑇𝑠𝑝) [𝑘𝑊ℎ/𝑚2.𝑦𝑒𝑎𝑟] Eq. 5.2 It is possible to create a more operational and simpler version of the model (E.q. 5.2) replacing the building archetype (M), the geographical location (L), the heating patterns (HPat) and indoor heat gains (HG) variables by the theoretical heating energy demand under reference heating conditions (HDRC). Besides all the limitations behind the methodology used to estimate HDRC (derived mainly from assumptions, see section 1.2), the convenience of using the HDRC as a proxy variable lies on the fact that, in principle, this information can be directly gathered from any energy rating/certification´s databases. Other major benefit of using HDRC is that it overcomes the difficulty of getting access to detailed data at building level. The form intended for the model proposed is represented in Eq. 5.3 and describes the HEU as a function of HDRC and Tsp. The model generates results in terms of useful or ‘net’ energy. 𝐻𝐸𝑈=𝑓(𝐻𝐷𝑅𝐶,𝑇𝑠𝑝) [𝑘𝑊ℎ/𝑚2.𝑦𝑒𝑎𝑟] Eq. 5.3 Based on Eq. 5.3, it is also possible to model the Tsp as a function of HEU and HDRC. This relationship is expressed in the form of Eq. 5.4 and provides information regarding the minimal guaranteed indoor temperature in spaces when heated that corresponds to a specific combination of HEU and reference HDRC: 𝑇𝑠𝑝=𝑓(𝐻𝐷𝑅𝐶,𝐻𝐸𝑈) [𝑘𝑊ℎ/𝑚2.𝑦𝑒𝑎𝑟] Eq. 5.4 With the specifications above, the HDRC is considered a key variable for modeling the heating energy use (HEU) or the indoor temperature, named here as set point temperature (Tsp). In this study, the modeling of HEU or Tsp was obtained using HDRC reflecting three different contexts: 1) a standard reference conditions dataset (HDRCst) derived from thermal building simulations. The temperature profiles and heating patterns were based on the RCCTE regulation´s [72] reference heating conditions. RCCTE regulation is the former 95 transposition of EPBD [6] for residential buildings. This analysis aims to create models that can be applicable to different geographical contexts; 2) RCCTE reference conditions dataset (HDRCRCCTE) derived from the RCCTE regulation´s energy calculation model [72]; 3) REH reference conditions dataset (HDRCREH) derived from the REH regulation´s energy calculation model [73]. REH regulation is the current transposition of the recast EPBD [7] for residential buildings. The two later analyses intend to develop models specifically tailored to the Portuguese context (called Portugal specific models: the RCCTE specific models and the REH specific models). Although the former RCCTE regulation is no longer in place, there are currently several certificates issued that were derived from this regulation. In summary, firstly, it was developed HEU or Tsp universal prediction models considering theoretical heating energy demand under standard reference conditions (HDRCst). Secondly, it was developed HEU or Tsp prediction models that are Portugal specific, considering theoretical heating energy demand under RCCTE and REH reference conditions (HDRCRCCTE and HDRCREH). 5.2 Modeling architecture The development of the models implied the construction of a variable database to be used in the statistical models. In order to characterize the relationships explained by Eq. 5.3 and 5.4, the database needed to be composed at least by three different variable datasets that will support the models: HEU, Tsp and HDRC datasets. Figure 26 schematizes the origin of the three datasets that compose the database used in the statistical models. 96 Figure 26. Schematic illustration of the construction of the database used in the statistical models. For both universal and Portugal specific models, the HEU dataset resulted from dynamic building thermal simulations and the Tsp dataset is an input to those HEU simulations. The only difference between the universal and the Portugal specific models is that, in the first, the HDRC dataset (HDRCst) resulted from dynamic building thermal simulations, whereas in the second, the HDRC variable dataset, i.e. the HDRCRCCTE and HDRCREH, resulted from the RCCTE and REH regulation’s building energy calculation models, respectively. All the dynamic thermal building simulations were run with the building energy model ESP-r. ESP-r is a well proven and validated tool that has been used in several research studies in the context of thermal buildings field [34,265,266]. The database used in the statistical models was therefore composed by inputs to and outputs from simulations and calculations methods. 97 Figure 27 explains in more detail the methodology behind the construction of the database to model HEU or Tsp. The Tsp, HEU and HDRC datasets are highlighted in red. Figure 27. An outline of the methodology proposed for the development of models. Recalling Eq. 5.2, HEU varies with building archetype (M), geographical location (L), heating patterns (HPat), indoor heat gains (HG) and set point temperature (Tsp). In particular, for each combination of a building archetype and geographical location (e.g. Manteigas), considering a specific heating pattern (e.g. heating 20m2 of the house for 4 hours per day during the winter season) and indoor heat gains (e.g. 3 W/m2), simulations were run for a number (maximum of 6) of different specific values of Tsp (e.g., from 10ºC to 25ºC). The construction of the HEU dataset was therefore built upon various combinations of M, L, HPat and HG, where each takes at a maximum of 6 different Tsp values, resulting each in a different HEU output values. The Tsp dataset corresponds to input values in HEU simulations. Considering Eq. 5.1, HDRC varies with building archetype (M) and geographical location (L), but with reference heating patterns (Hpatref), indoor heat gains (HGref) and set point temperature (Tspref). The same combinations of building archetypes and geographical locations (e.g., Manteigas), used in the construction of the HEU dataset, were run in simulations or calculated considering the reference set point temperature (20ºC or 18ºC, depending on the 98 HDRC variant), the reference heating pattern (heating the house all day long during the entire winter season) and the reference indoor heat gains (4 W/m2). Thus, the construction of HDRC dataset was built upon several combinations of M, L, HPatref, HGref and Tspref, resulting each in a single HDRC output value. At the end, the HEU, HDRC and the Tsp datasets built up the database used in the regression and artificial neural networks (ANN) models applied for modeling HEU or Tsp. The HEU was modeled as a function of the independent variables HDRC and Tsp, whereas the Tsp was modeled as a function of the independent variables HDRC and HEU. As mentioned previously, the HEU dataset resulted from building thermal simulations and the HDRC dataset was obtained either from building thermal simulations (in case of universal models) or building energy calculation models (in case of Portugal specific models). Various building models considered in ESP-r simulations and energy calculation models were required to create the database to be used in the statistical models. The building models implied the need for several combinations of the categories: building archetypes, geographical locations, heating patterns, set point temperatures and indoor heat gains. The range of values of those categories and the combination between them were selected to be the widest and diverse as possible to capture a vast range of HEU, HDRCst, HDRCRCCTE and the HDRCREH output values. I.e., so that HEU or indoor temperature predicting models would perform estimations for a wide range of each independent variable/representativeness [69]. Next sections present the possibilities considered within these categories. 5.2.1 Building archetype Different building archetypes (M) were constructed in ESP-r for the HEU simulations. The selected values for physical characteristics and surroundings of building archetypes are defined in Table 20. 99 Table 20. Physical characteristics and surroundings of the building archetypes (M). Characteristics Discrete values Type of dwelling Detached house; semi-detached house; terrace house; apartments For houses For apartments Construction period <1960; 1961-90; 2006-14 Air infiltration rate/natural ventilation (IR) (ac/h) 0.6; 0.8; 1.0; 1.2; 1.5; 1.7 0.6; 1.0; 1.2; 1.5; 1.7 Floor area (m2) 150I; 225II; 251I; 300III; 350I 100IV; 141IV; 181IV; 200I % of glazing area per facade [190] 10%; 43%; 75% Orientation of the facades All orientations West; North; South No. of floors 2; 3; 4 1;2 Overhang None; Overhang of 1.5m just 1st floor; Overhang of 1.5 m on both floors None; Overhang of 1.5m; Overhang of 0.5 m Type of urbanization No buildings in the surroundings; Houses in the surroundings with same height; Other type of houses in the surroundings No buildings in the surroundings; Apartments in the surroundings I Resultant from 2 floors; II Resultant from 3 floors; III Resultant from 4 floors; IV Resultant from 1 floor. The type of dwelling was composed by six categories: three representing houses (detached house, semi-detached house and terrace house), and the other three, apartments located between other apartments with 1, 2 or 3 external facades (1F, 2F, 3F, respectively). Three values of percentage of glazing area in each facade were considered for both type of buildings: 10%; 43% and 75%. The construction period was categorized in three ‘slices’: <1960; 1961-1990 and 2006-2014. Each slice was characterized by different combinations of building´s construction and materials and air infiltration rates. Details of building´s construction and materials for the three construction periods are presented in Table C.1 and Table C.2, in Appendix C, for houses and apartments, respectively. 100 In terms of number of floors, houses were designed for 2, 3 and 4 floors, whereas apartments considered 1 and 2 floors. The floor area values corresponded to 150, 250, 350m2, for houses with 2 floors and 225m2 and 300m2 for houses with 3 and 4 floors, respectively. For the apartments, often smaller in size, the floor area values corresponded to 100, 140, 180m2 for apartments with 1 floor and 200m2 with 2 floors. These values were defined based on data provided by the National Institute of Statistics (INE) database [214]. In terms of the orientation of facades, it was considered that facades had four possible orientations (i.e., South, North, West and East) for the houses, whilst facades of the apartments had three possible orientations (South, West, or North). Regarding the presence of an overhang that could provide shading to the windows, houses were simulated using three options: a) no overhang (i.e., no shading); b) overhang of 1.5 m length in the first floor; c) overhang of 1.5 m length in the two floors. In turn the apartments were simulated with: a) no overhang; b) overhang of 1.5 m length, and c) 0.5 m length. The type of urbanization recreates different scenarios of shading and infiltration rates. For the houses, the variable type of urbanization was created by running simulations with: a) no buildings in the surroundings; b) house is surrounded mainly by other houses with the same height; and c) house is surrounded by houses and apartment buildings. In turn, the type of urbanization for the apartments was captured by running simulations with: a) no buildings in the surroundings; and b) with apartments in the surroundings. The air infiltration rate/natural ventilation 8 (IR) is influenced by several factors, namely the construction period, the type of urbanization, the type of window´s insulation (e.g. in accordance with the construction period; well insulated; and very well insulated), and the window opening behaviour (e.g. normal patterns; excess in opening windows). 8 ESP-r does not allow the direct input of the natural ventilation values, therefore, they were considered along with air infiltration rates. 101 In reality air infiltration rates/natural ventilation, depending on the situation, can vary with time and space. For simplicity, HEU simulations assumed a constant IR values (between 0.6 to 1.7ach/h) for everyday of the week 9 and every rooms. In terms of window’s control, it was assumed that buildings receive sunlight through windows every day during daytime. Consequently, the thermal conductivity coefficient values of windows were estimated as a weighted value, considering that for half of the time the venetians are open. Also, complex fenestration construction (CFC) files, for each type of window, were implemented in ESP-r using the Glazing Shading Layer Editor (GLSedit) tool. This tool contains an extensive glazing product selection for many manufacturers and was designed for quick synthesis of a glazing product with or without shading components. The output information on optical proprieties of the CFC can then be read by ESP-r [267]. Normal solar, visible and longwave optical proprieties for glazing and venetian blinds layers assumed from the GSLedit´s database, for the three construction periods and for both houses and apartments, are presented in Table C.3, in Appendix C. In addition, thermal conductivity coefficient values were accounted for thermal bridges for each building archetypes. These were considered by including an equivalent thermal conductivity coefficient Ueq for each external facade. The Ueq was estimated accordingly to the physical characteristics of each building archetype, using the Eq. 5.5: 𝑈𝑒𝑞(𝑖)= 𝐴𝑟𝑒𝑎𝑙×𝑈𝑟𝑒𝑎𝑙+𝐴𝑓𝑡𝑏×𝑈𝑓𝑡𝑏+∑(𝐿𝑙𝑡𝑏×𝛹𝑙𝑡𝑏) [𝑊/𝑚2.º𝐶] Eq. 5.5 𝐴𝑡𝑜𝑡𝑎𝑙(𝑖) where, Ueq(i) is the equivalent thermal conductivity coefficient (W/m2.ºC) of facade i; Areal is the area of the construction building material in facade i (m2); Ureal is the thermal conductivity coefficient (W/m2.ºC) of the construction building material in facade i (W/m2.ºC); Aftb is the area 9 Simulations in EPS-r works primarily with weekly air infiltration rate/natural ventilation inputs that are reproduced for all the year. 108 By definition, the HG depends on the following occupancy and occupant’s behaviour (OOB) characteristics, in particular: the occupation patterns, for example, dwellings occupied for longer period would result in higher heat gains; the household size, for example higher number of people would result in higher heat gains; and the level of energy, which defines the lighting and equipment power, for example, higher number of used equipment/bulbs reflect on higher heat gains. 5.2.4.2 Occupancy and occupant´s behaviour characteristics variables to define heating patterns and indoor heat gains The HDRCst simulations and the HDRCRCCTE and HDRCREH calculations took reference values for the heating patterns (Hpatref), as well as for indoor heat gains (HGref). In particular, it was assumed indoor heat gains (HGref) of 4 W/m2, percentage of heated area (HA%ref) of 100%, and entire winter season as the heating period (HPref). In contrast to reference heating conditions, the heating patterns and internal heat gains assumed in the HEU simulations were captured from combinations of occupancy and occupant behaviour characteristics (OOB) in order to increase the range of their values. 109 Table 22 presents the OOB characteristics. Table 22 also indicates which OOB characteristics have an influence on heating patterns and internal heat gains. Table 22. Occupancy and occupant behaviour characteristics of the heating patterns and indoor heat gains. OOB characteristics Discrete values Heating patterns (HPat) and indoor heat gains (HG) Heating schedule When and where occupied (W2); Everywhere, anytime (EA); Everywhere in the occupied period 1 (EO 1) and period 2 (EO 2); Only in common area in the period 1 (CA 1); period 2 (CA 2); and period 3 (CA 3); Specific period (SP) % of heated area Occupation patterns Work time out (WTO); Always at home (ATH); Morning time out (MTO) % of heated area; Indoor heat gains No. bedrooms occupied For houses: 2 and 6 bedrooms occupied; For apartments: 2 and 4 bedrooms occupied % of heated area Level of energy use Low; Normal; High Indoor heat gains Household size 2 people; 4 people; 8 people % of heated area; Indoor heat gains Heating period Winter season period; December; January; February; December to January; December to February; November to January; November to February Heating period The number of bedrooms occupied was different for the houses and apartments. 2 and 6 bedrooms occupied (out of 3 and 6, respectively) were assumed for the houses, whereas for the apartments, 2 and 4 bedrooms were considered occupied (out of 2 and 4, respectively). The level of energy use assumed three levels: a) very low energy use, which may result from the use of more efficient equipment/lighting or the reduced number of equipment/bulbs; low energy use; b) high energy use resultant from the use of less efficient equipment/lighting or the abusive use, or high number of equipment/bulbs; c) an intermediate energy use values. The heating schedule aims to capture the main heating schedules that occur in dwellings. It assumed six options as follows: a) the rooms occupied by the households will be heated only 110 during the time of their occupation (W2); b) all the rooms of the dwelling will be heated during the 24h period, regardless the dwelling occupation (EA); c) all the rooms will be heated regardless their occupation, but only during the dwelling occupied period from to 19:00 to 07:00 (EO 1); d) all the rooms will be heated regardless their occupation, but only during the dwelling occupied period from to 13:00 to 10:00 (EO 2); e) only the common area will be heated during the 24h period (CA 1); f) only the common area will be heated during the occupied period from 19:00 to 23:00 (CA 2); g) only the common area will be heated during the occupied period from 20:00 to 23:00 (CA 3); and during a specific period (from 09:00 to 24:00 in all the rooms and 24h in the bedrooms). The occupation patterns refers to the occupied period and is defined as follows: a) the occupation schedule where occupants leave home in the early morning and arrive in the afternoon (WTO); b) the occupation schedule where occupants stay at home all day (ATH); c) the occupation schedule where occupants leave home in the early morning and arrive after lunch (MTO). In addition, the occupation patterns define the distribution of the households at home (e.g., occupants in the living room during the afternoon and in the bedrooms in the evening). The heating period refers to the length of the heating during the winter season period. Eight periods were assumed: a) November to February; b) December to February; c) November to January; d) December to January; e) January; f) December; g) February; and h) winter season period. The length of the winter season period, for each geographical location, was dependent on the climate conditions. Based on ref. [72] (see subheading bb) in ANEXO II, Definições) heating period was defined from the first ten-days after 1st October, in which, for each geographical location, the daily mean temperature is below 15ºC, ending in the last ten-days before 31st May, in which the referred temperature is still below 15ºC. As explicit in Table 22, the heating period (HP) is determined by the heating period. In Esp-r simulations, in the analysis of the results, one can choose the period of analysis. In RCCTE and REH´s building energy calculation models, the entire winter season is defined by default. 111 From Table 22, it is also possible to verify that the combination of the occupation patterns coupled with heating schedules, household size and number of bedrooms occupied result on different weekly heating patterns (i.e, which rooms are heated and when during a week). For each weekly heating patterns, a correspondent percentage of heated area (HA%) value was estimated, for each value of floor area. In Esp-r simulations, the percentage of heated area is introduced in the form of weekly heating patterns by designating set point temperatures values to each heated rooms during different heating periods, for each day type (weekdays, Saturday and Sunday) of the week. The RCCTE and REH´s building energy calculation models consider by default that all the rooms are heated all day long. Also, when combinations of the level of energy use, occupation patterns and household size are coupled, different weekly 12 indoor heat gains patterns are defined. For each weekly indoor heat gains pattern, a correspondent average indoor heat gains (HG) value was estimated, for each value of floor area. In Esp-r simulations, the indoor heat gains are introduced in the form of weekly indoor heat gains patterns by designating hourly sensible and latent heat gains from equipment, people and lighting, for each day type of the week and rooms. The RCCTE and REH´s building energy calculation models define by default a constant value for indoor heat gains. Due to the great amount of weekly heating patterns and indoor heat gains patterns defined for this work, just one example is illustrated for the case of an apartment 1F with a floor area of 100m2. The apartment has the following characteristics: W2 heating schedule; WTO occupation pattern; 2 bedrooms occupied; normal level of energy use; and 4 people living in the dwelling, resulting in values of 29% and 4.5 W/m2 for the HA% and HG, respectively. Table C.9, in Appendix C, presents the hourly sensible and latent heat gains per room from occupant, lighting and equipment, along with additional information, used to estimate the indoor heat gains, and Table C.10 presents the equipment used in each room, depending on the period of the day. Information on Table C.10 aided the estimation of the hourly sensible heat gains from equipment, shown in Table C.9. Taking the case of the apartment 1F as an example, Figure C.1 and Figure C. 2, in Appendix C, illustrate the heating patterns and indoor heat gains patterns for weekdays, for all the rooms in the dwelling, respectively. 12 Simulations in EPS-r works primarily with weekly inputs that are reproduced for all the year. 112 The usefulness of capturing the influence of the OOB characteristics using heating patterns (HPat) (i.e., HP and HA%) and indoor heat gains (HG) lies on the fact that it enables to reduce the number of variables needed in the model, providing a more intuitive formulation. 110 combinations of the OOB characteristics were performed and resulted in a broad range of HPat and HG values that were assumed in the HEU simulations, especially because HA% and HG are floor area dependent. For example, for a particular combination of OOB characteristics, such as W2 heating schedule; WTO occupation pattern; 2 bedrooms occupied; normal level of energy use; 4 people occupying the dwelling; and the entire winter season as the heating period, the HA% and HG take values of 25% and 3.5 W/m2, respectively, for a dwelling with 150m2, and 29% and 4.5 W/m2 for a dwelling with 100m2 of floor area. The range of HA% values obtained was between 4% and 100%; the HP varied between the possibilities that includes the entire winter season period and 1 to 4 months of heating period; the range of HG values varied between 1.4 and 12.0 W/m2. 113 5.3 Description of variable database used in the statistical models Figure 30 illustrates the main steps taken in the development of the statistical models. The first step, which is the selection of the different possibilities within each category, was presented in the last sections (5.2.1 to 5.2.4). Figure 30. Main steps to the development of the statistical models. As mentioned in section 5.2, the combination of different categories (i.e., building archetypes, geographical locations, set point temperature, heating patterns and indoor heat gains) characterized the building models constructed or considered in the HEU and HDRCst simulations and HDRCRCCTE and HDRCREH calculations. The inputs and the resultant outputs of those simulations/calculations built up the variable database used in the statistical models. This section will describe the combinations of the different 5 categories used to characterize the building models for the simulations/calculations (Step 2) and the database resultant from these simulations/calculations (Step 4). Two different approaches (A1 and A2) were used to combine the 5 categories. Each approach resulted on one different database. The universal and Portugal specific models for heating energy use (HEU) or set point temperature (Tsp) were developed using these two different databases. In this respect, Figure 31 illustrates the total number of models developed 114 under this chapter (four universal models and eight Portugal specific models, i.e., four RCCTE specific models and four REH specific models). In the first approach (A1), the HEU dataset resulted from 745 dynamic hourly simulations combining only the categories: building archetypes, geographical locations and set point temperatures. In the second approach (A2), the HEU dataset resulted from 2611 dynamic hourly simulations combining the categories: building archetypes, geographical locations, set point temperatures, heating patterns and indoor heat gains. In both approaches, the three HDRC datasets (e.g. HDRCst, HDRCRCCTE, HDRCREH) resulted from simulations or calculations (each) combining only the categories building archetypes and geographical locations, as set point temperature, heating patterns and indoor heat gains are reference values. The HDRC are replicated in the databases accordingly (as in Table shown in Figure 27) to perform the total number of observations included in HEU dataset. The database in A1 includes 745 observations, while, in A2, the database totals 2611 observations. Figure 31. Schematization of the different models developed under chapter 5. 115 Table 23 summarizes all models developed and the characteristics of the databases for each approach, which will be discussed in more detail in the next sections 5.3.1 and 5.3.2. Table 23. Summary of universal and Portugal specific models. Model designation Dependent variable HEU dataset HDRC dataset Approach A1 Modeluni. HEU.A1 HEU predicting model 745 observations (28 building archetypes x 5 locations x maximum 6 of Tsp) 140 observations (28 building archetypes x 5 locations) ModelRCCTE HEU.A1 ModelREH HEU.A1 Modeluni. Tsp.A1 Tsp predicting model ModelRCCTE Tsp.A1 ModelREH Tsp.A1 Approach A2 Modeluni. HEU.A2 HEU predicting model 2611 observations (267 building archetypes/Hpat/HG x maximum of 5 locations x maximum of 6 Tsp) 220 observations (44 building archetypes x 5 locations) ModelRCCTE HEU.A2 ModelREH HEU.A2 Modeluni. Tsp.A2 Tsp predicting model ModelRCCTE Tsp.A2 ModelREH Tspsp.A2 5.3.1 Description of database: A1 - Varying only building archetype, geographical location and set point temperature The first 28 building archetypes ([1-28], see Table 21, section 5.2.1) were combined with the 5 geographical locations (Manteigas, Bragança, Porto, Lisbon and Faro) to originate 140 combinations. All the HEU simulations considered only one combination of the occupancy and occupant behaviour (OOB) characteristics from the 110 combinations selected (see section 5.2.4.2). Table 24 illustrates the OOB characteristics which reflected a range of different heating patterns (HPat) and indoor heat gains (HG) values, depending on the floor area of each one of the 28 116 building archetypes. The range of values were computed: the indoor heat gains (HG) values within the range of 1.4 to 4.5 W/m2 and the percentage of heated area (HA%) in the range of 13% to 29%. The heating period (HP) was correspondent only to the entire winter season. Each one of the 140 combinations considered a particular value within the range of the HA% and HG values. Table 24. Combination of OOB characteristics, in the first approach. Heating Schedule Occupation patterns Level of energy use Household size No. bedrooms occupied Heating period W2 WTO Normal 4 2 Winter season (see Table 22 in section 5.2.4.2 for definitions.) In particular, the 745 HEU simulations were performed by running each one of the 140 building models for a maximum of 6 set point temperature (Tsp) values (between 10ºC and 28ºC), which retrieved 745 observations to be analyzed. The 140 HDRCst simulations were obtained by running the 140 building models at 20ºC. For the same combinations, the HDRCRCCTE and the HDRCREH calculations (140 each) assumed 20ºC and 18ºC, respectively. The assumed reference heating patterns (HPatref) and internal heat gains (HGref) values were as follows: the indoor heat gains (HGref) equal to 4 W/m2, the percentage of heated area (HA%ref) equal to 100%, and the heating period (HPref) corresponded to the winter season period. It returned with 140 observations to be analysed. The building models were characterized by combinations between different categories. Table 25 presents the possibilities within the different categories used in approach A1 to characterize the building models constructed/considered in the HEU and HDRCst simulations and in the HDRCRCCTE and HDRCREH calculations. Table 25 makes clear that the range of values of the 5 categories varies dependent whether they are building models for the simulations/calculations of HEU, HDRCst, HDRCRCCTE or HDRCREH. 117 Table 25. Comparison between the different possibilities within the categories used to characterize the building models for the HEU and HDRCst simulations, and HDRCRCCTE and HDRCREH calculations (Approach A1). Categories that characterize the building models HEU simulations HDRCst simulations HDRCRCCTE calculations HDRCREH calculations Building archetypes (M) 28 building archetypes 28 building archetypes, excepting constant IRref value (0.95ac/h) Based on the 28 building archetypes. Differences in the calculation methods and in some of the values used Geographical locations (L) Manteigas, Bragança, Porto, Lisbon, Faro using climate files based on RCCTE energy calculation model´s climate files Manteigas, Bragança, Porto, Lisbon, Faro using energy calculation model´s climate files by default Set point temperature (Tsp) Tsp: 10 to 28ºC Tspref: 20ºC Tspref: 20ºC Tspref: 18ºC Heating patterns (HPat) HPat: HPatref: HPatref: HPatref: % of heated area (HA%) HA% - 13% to 29% HA%ref - 100% HA%ref - 100% HA%ref - 100% Heating period (HP) HP - Winter season period HPref - Winter season period HPref - Winter season period, but might be defined differently from those assumed in simulations Indoor heat gains (HG) HG: 1.4 to 4.5 W/m2 HGref: 4 W/m2 HGref: 4 W/m2 HGref: 4 W/m2 The database was resultant from inputs/outputs of the HEU and HDRCst simulations, and HDRCRCCTE and HDRCREH calculations. Table 26 presents only the set of variables that constituted the database (in the first approach) used in the universal and Portugal specific statistical models developed to predict HEU (Eq. 5.3) or Tsp (Eq. 5.4). 124 In this chapter, the ANN analyses were developed using the Neural Toolbox in MATLAB R2014b software [277]. For modeling the problem using an ANN, a feedforward multilayer neural network with a back-propagation technique was used. A feedforward multilayer neural network consists of an input layer, one or more hidden layers, and an output layer. In this work, a single hidden layer was considered to map the function provided suitable hidden neurons. The hidden layer assists to solve non-linear separable problems [199]. In particular, a three-layer feedforward network (one input layer, one hidden layer and one output layer), with a nonlinear activation function in the hidden layer and a linear function in the output layer (see Figure 33), was used. The ANN was trained applying the Levenberg- Marquardt algorithm [278]. The selection of the algorithm depends on many factors, including the complexity of the problem, the number of observations in the training set, the number of weights and biases in the network, the error goal, and the type of use (classification or regression). In this study, Levenberg-Marquardt algorithm was used as it is one of the fastest training functions, and it is suitable to be used in not very large networks (i.e, with thousands of weights) [279]. It is also the default training function for feedforward net in the neural toolbox. Other algorithms were tested (e.g., BFGS Quasi-Newton and Resilient Backpropagation) and they perform worse than the Levenberg-Marquardt algorithm. 125 Figure 33. Schematic of a three-layer feedforward network [Based on Ref. [186]]. In Figure 33, X (X= 1, .., Xp) represents the different independent variables applied to the training model; n (n= 1, …, ni) represents the number of neurons applied to the training model and Y represents the target (i.e., dependent variable) of the training model. For the nonlinear activation function (i.e, first function), each X is connected to each neuron n through weight values (W). Each neuron has a bias b, which is summed with the weighted values as described by the following equation: 𝑛𝑖=𝑊𝑋1,𝑛𝑖+𝑊𝑋2,𝑛𝑖+⋯+𝑊𝑋𝑃,𝑛𝑖+𝑏𝑖 Eq. 5.7 In the linear function (i.e., the second function), each neuron is connected to the target Y by weight values (L). The target Y has also a bias bY associated, which is summed with the weighted values as described by the following equation: 𝑌=𝐿𝑛1,𝑌+𝐿𝑛2,𝑌+⋯+𝐿𝑛𝑖,𝑌+𝑏𝑌 Eq. 5.8 There are no clear rules to choose the ‘best’ number of hidden nodes. Network design is a trial-and-error process and may affect the accuracy of the model. The models were tested for different number of hidden neurons using the constructive method. First, it was tested a small 126 ANN, and then neurons were added until reaching better accuracy values. Note that all variables were normalized to fall between -1 and 1 in order to achieve faster convergence and better accuracy. The process of training involved tuning the values of the weights and biases of the network to optimize network performance, using the mean square error (MSE) as the default performance function [279]. In this work, database was randomly divided into training (50%), validation (25%) and testing (25%) datasets to provide generalization to the model. Training dataset is used to learn the behaviour of input data and to adjust the model coefficients. It was selected the best evaluation criterion (the mean square error (MSE)) of 100 training runs. Validation dataset is used to control the overfitting (overfitting occurs when a statistical model describes random error or noise instead of the underlying relationship). Testing dataset is used to evaluate the models by applying the accuracy metrics (section 5.4.2) [199]. The final models used are explained in detail in sections 5.5 and 5.6. 5.4.2 Performance evaluation of the models Four accuracy metrics were used to evaluate the goodness of fit of the regression and the ANN models. The first is the coefficient of determination (R2) (predictive R2), which indicates how closely predicted values match the actual values. R2 can be calculated as in Eq. 5.9 [199]. 𝑅2= ∑[𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑−𝑦𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑]2 𝑛 𝑖=1 𝑖=1,2,…𝑛 Eq. 5.9 ∑[𝑦𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑−𝑦𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑]2 𝑛 𝑖=1 . where, yobserved is the observed values, 𝑦observed is the average observed values; ypredicted is the predicted values and n stands for the total number of observations. 127 The following metrics evaluate the magnitude of the errors between the observed and predicted values. The mean square error (MSE) can be described by Eq. 5.10 [149,199,254]: 𝑀𝑆𝐸= ∑[𝑦𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑−𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑]2 𝑛 𝑖=1 𝑖=1,2,…𝑛 Eq. 5.10 𝑛 where, yobserved is the observed values and ypredicted is the predicted values and n stands for the total number of observations. This accuracy metrics is introduced in this analysis because the neural network run by MATLAB uses the MSE [199,272,280] as an evaluation criterion of the trained network. The other two accuracy metrics correspond to the mean absolute error (MAE), which was introduced in section 4.3.1, and the mean absolute percent error (MAPE). Percentage errors have the advantage of being scale-independent, and so are frequently used to compare performance between different models. Many organizations focus primarily on MAPE [158,170,254,256] when assessing forecast accuracy. Also, most people are comfortable thinking in percentage terms, making the MAPE easy to interpret. MAPE suits the modeling analysis developed in this chapter as it involves a great amount of data that is guaranteed to be strictly positive [255]. Because percentage errors assume a meaningful zero, a percentage error makes no sense when measuring the accuracy of temperature predictions on the Fahrenheit or Celsius scales[281]. The MAPE accuracy metric can be calculated using Eq. 5.11. 𝑀𝐴𝑃𝐸= ∑‖𝑦𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑−𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑦𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑 ‖ 𝑛 𝑖=1 ×100 𝑖=1,2,…𝑛 [%] Eq. 5.11 𝑛 where, yobserved is the observed values and ypredicted is the predicted values and n stands for the total number of observations. 128 The ‘best’ model is the one that gathers the minimum values of error metrics and the highest values of R2. 5.5 Development of the universal models The universal HEU or Tsp predicting models were developed using two approaches (A1 and A2), considering each two different databases (see Table 26, section 5.3.1 and Table 28, section 5.3.2, respectively). In this research, four universal models (see Table 23, in section 5.3) were developed: Modeluni. HEU.A1; Modeluni. Tsp.A1; Modeluni. HEU.A2 and Modeluni. Tsp.A2. The models developed in the approach A1 are limited to a more narrow range of heating patterns (HPat) and indoor heat gains (HG) scenarios (see, Table 25, second column, in section 5.3.1). In turn, the models developed in the approach A2 can be applied to a broad range of situations (see Table 27, second column, section 5.3.2). However, these models require a higher level of expertise and knowledge in statistical modeling. In the approach A1, the universal modeling of HEU (Modeluni. HEU.A1) and of Tsp (Modeluni. Tsp.A1) were developed applying multivariate regression. In the approach A2, more sophisticated techniques were employed in order to improve the performance of the models due to the nature and size of the database. Firstly, the universal modeling of HEU (Modeluni. HEU.A2) was developed applying the multivariate regression analysis and also ANN models. Secondly, these models were extended using an additional set of independent variables (HA%, HP, HG, IR), as illustrated by Equations 5.12 and 5.13. This set of variables represents inputs to HEU simulations (see Table 28). The modeling of Tsp (Modeluni. Tsp.A2) was developed applying only ANN analysis, using Eq. 5.14. 𝐻𝐸𝑈=𝑓(𝐻𝐷𝑅𝐶,𝑇𝑠𝑝,𝐻𝐴%,𝐻𝐺,𝐻𝑃) [𝑘𝑊ℎ/𝑚2.𝑦𝑒𝑎𝑟] Eq. 5.12 129 𝐻𝐸𝑈=𝑓(𝐻𝐷𝑅𝐶,𝑇𝑠𝑝,𝐻𝐴%,𝐻𝐺,𝐻𝑃,𝐼𝑅) [𝑘𝑊ℎ/𝑚2.𝑦𝑒𝑎𝑟] Eq. 5.13 𝑇𝑠𝑝=𝑓(𝐻𝐷𝑅𝐶,𝐻𝐸𝑈,𝐻𝐴%,𝐻𝐺,𝐻𝑃,𝐼𝑅) [º𝐶] Eq. 5.14 where, HDRC is the theoretical heating demand under reference conditions; and HA% is the percentage of heated area, HG is the indoor heat gains (W/m2), HP is the heating period and the IR is the air infiltration rate/natural ventilation (ac/h) input values assumed in the HEU simulations. HEU is the heating energy use in terms of ‘useful’ energy. Table 29 summarizes the best performing models among all the models performed. In particular, it reports, for each predicting model analyzed (first column), the statistical analyses conducted (Multivariate linear regression – MLR; Multivariate non-linear regression – MNLR; and Artificial neural networks - ANN), the accuracy metrics (R2, MAE, MAPE, MSE), the number of neurons used in the ANN, and the dependent and independent variables. MAE is used in terms of kWh/m2.year for the HEU predicting values, and in terms of ºC for the Tsp predicting values. 130 Table 29. Results of universal models (A1 and A2). Predicting Models Statistical models Testing accuracy metrics No. of neurons Dependent variable Independent variables R2 MAE MAPE MSE ApproachA1 Modeluni. HEU.A1 MLR 0.774 13.4 159% 2.7.E+02 ----- HEU Tsp*; HDRCst* MNLR 0.930 6.3 42% 7.7E+01 ----- HEU Tsp*; HDRCst; Tsp2; HDRCst2; Tsp.HDRCst; Tsp.HDRCst2; Tsp2.HDRCst*; Tsp2.HDRCst2* Modeluni. Tsp.A1 MNLR 0.910 1.8 ------ 2.4E+00 ----- Tsp HEU*; HDCRst*; HEU2*; HEU.HDRCst*; HEU2.HDRCst*; HDRC2*;HEU2.HDRCst2*; HEU.HDRCst2 ApproachA2 Modeluni. HEU. A2 MNLR 0.511 29.0 238% 2.5E+03 ----- HEU Tsp; HDRCst*; Tsp2; HDRCst2; Tsp.HDRCst*; Tsp.HDRCst2; Tsp2.HDRC; Tsp2.HDRCst2 ANN 0.489 32.6 244% 2.8E+03 5 HEU HDRCst; Tsp ANN 0.353 33.9 294% 3.5E+03 22 HEU HDRCst; Tsp MNLR 0.889 14.9 132% 5.5E+02 ----- HEU Tsp; HDRCst*; HA%*; HG*; HP*; Tsp2; HDRCst2; HA%2; Tsp.HDRCst; Tsp.HA%; Tsp.HG; Tsp.HP; Tsp.HDRCst2; Tsp.HA%2; Tsp2.HDRCst*; Tsp2.HA%*; Tsp2.HP; Tsp2.HG*; Tsp2.HDRCst2; Tsp2.HA%2; HDRCst.HA%*; HDRCst.HG*; HDRCst.HP; HDRCst.HA%2*; HDRCst.HG; HDRCst.HP; HDRCst2.HA%*; HDRCst2.HG; HDRCst2.HP; HDRCst2.HA%2; HA%.HG*; HA%.HP; HA%2.HG*; HA%2.HP; HG.HP ANN 0.970 7.1 51% 2.E+02 9 HEU HDRCst, Tsp, HA%, HG, HP ANN 0.988 4.9 45% 6.7E+01 13 HEU HDRCst, Tsp, HA%, HG, HP, IR Modeluni. Tsp. A2 ANN 0.966 0.7 ---- 9.7E-01 15 Tsp HDRCst, Tsp, HA%, HG, HP, IR Only for regression statistical models: significant at 1%: *; significant at 5%: **. ´ 131 Analysing Table 29 it is possible to conclude that three of the four models proposed with universal applicability (shaded in grey) revealed to be good predicting models. These are: a) the model to predict HEU using the first approach: only varying physical characteristics of the building archetypes and geographical locations (Modeluni. HEU.A1, MNLR statistical model); b) the model to predict HEU using the second approach: varying physical characteristics of the building archetypes and geographical locations and occupancy and occupant’s behaviour (OOB) characteristics (Modeluni. HEU.A2, ANN statistical model); c) the model to predict Tsp using the second approach: varying both physical characteristics of the building archetypes and geographical locations and OOB characteristics (Modeluni. Tsp.A2, ANN statistical model). The coefficients of determination (R2) of the three best statistical models are close or even higher that most values found in the literature. For example, Kialashaki and Reisel [272] developed three models for predicting energy demand of the residential sector of USA using ANN statistical models. The respective R2 values were 0.9823, 0.9849 and 0.9896. Paudel et al. [199] obtained a R2 of 0.85 for the building heating energy demand ANN model. Buratti et al. [191] in developing an ANN model obtained a R2 of 0.9957. Also, Aydinalp et al. [166] developed a ANN model for modeling space heating energy demand in the residential sector obtained a R2 of 0.908. The empirical studies that applied standard linear regression presented lower R2 values. This is the case of Kelly (2011) [259], who presented a model with an adjusted R2 of 0.314. Another example is the finding achieved by Santin et al. (2009) [63], which modeled energy use as a function of the building characteristics and achieved a R2 equal to 0.42. The manifested errors are probably mainly derived from the differences between the values taken for the different categories (i.e., building archetypes, geographical locations, set point temperature, heating patterns and indoor heat gains) that characterized the building models used in the HEU and HDRCst simulations. Because databases from approach A1 and A2 were resultant from different combinations of categories, errors might have different origins depending on the approach used to build up the database. 132 In approach A1 (see Table 25, section 5.3.1) errors resulted mainly from HEU simulations assuming heating patterns, indoor heat gains and air infiltration rate/natural ventilation values particularly distinct from the reference values assumed in the HDRCst simulations. The approach A2 (see Table 27, section 5.3.2) attempted to cope with the issue intrinsic to models developed under A1 by including additional independent variables (IR, HA%, HP, HG) to explain better the HEU. Still, part of the errors might also be explained by the limitation inherent to the use of variables in the statistical models that do not account neither with the effect of orientation of the heated spaces nor with the period of heating during the day on the heating energy use. This might be relevant as solar gains and thermal losses vary during the day and depend on the orientation of the dwelling. This is the case of the variables IR and Tsp. For example, for a building archetype with a specific orientation, assuming a Tsp value of 20ºC or an IR value of 1.3ac/h in the statistical model, the model would predict a certain value of HEU, regardless, if the heated rooms are oriented towards South or North and heated during the morning or evening. The same issue is extended to the variables HG and HA%. An in depth analysis regarding the development of the three predicting models is presented next: Concerning the model for predicting HEU using the approach A1 - Modeluni. HEU.A1, the analysis of the errors, measured through the accuracy metrics (MAE, MAPE and MSE) and the R2, indicated that the MNLR statistical model perform better than the corresponding MLR statistical model. The MNLR model predicts relatively well HEU using the Tsp and HDRCst, and other independent variables as presented in Table 29, with a R2 of 0.93. All the detailed results regarding the MNLR statistical model are presented in Appendix C (parameter estimates in Table C.12; comparison between observed and predicted values, using the testing dataset in Figure C.3). 133 In the attempt to develop models to predict HEU using the approach A2 - Modeluni. HEU.A2, six models varying in type of statistical model and independent variables were developed. The evaluation of the accuracy metrics brings the conclusion that none of three first models are particularly effective at predicting the HEU using only the independent variables HDRCst. and Tsp. This may be an indication that the increase of observations with a large variety in terms of heating patterns (HPat) and internal heat gains (HG) input values in the HEU simulations led to models with worse performance. A possible explanation for this performance is that the independent variables HDRCst and Tsp together no longer explain adequately the wide variation of HEU values, especially because HDRCst assumes single reference values for the HPatref and HGref. This can also be explained by the different air infiltration rates/natural ventilation values assumed in both HEU and HDRCst simulations (the latter assumed a reference value (IRref)). Therefore, there was the need to consider an extended Modeluni. HEU.A2 with further independent variables to better explain HEU values. The variables introduced in the model were: the two components of heating patterns (HA% and HP), the internal heat gains (HG) and the air infiltration rates/natural ventilation (IR). The relationships illustrated in Eq. 5.12 and 5.13 were thus analyzed using MNLR and ANN statistical models. From these three last models, it is possible to conclude that the two ANN models are the most promising models to predict the HEU values as they have better accuracy results than regression analyses (see Table 29). The ability of ANN in performing non-linear analysis is therefore an advantage [256] towards the multivariate regression analysis. Because of the learning properties of the ANN model and its sensitivity to fluctuations of the independent variables, the performance of the ANN model is clearly better than the regression models and the results generated by the ANN model are closer to the actual observed data [189]. Furthermore, it can be concluded that variable IR improves the prediction ability of the ANN model, exhibiting higher R2 and lower errors than the ANN model in the absence the IR variable. Therefore, the best model to predict HEU is an ANN statistical model using the variables HDRCst, Tsp, HA%, HP, HG and IR (R2 equal to 0.988). In the end, it can be concluded that the relationship characterized by Eq. 5.13 is therefore the most appropriate to predict HEU over the Eq. 5.3 initially considered. All the detailed results regarding the ANN statistical model are presented in Appendix C (comparison between observed and predicted values, using the testing dataset, in Figure C.4).