Artificial neural networks and physical modeling for determination of baseline consumption of CHP plants

Artificial neural networks and physical modeling for determination of baseline consumption of CHP plants Francesco Rossi a, ⇑ , David Velázquez a , Iñigo Monedero b , Félix Biscarri b a Department of Energy Engineering, Universidad de Sevilla, Spain b Electronic Technology Department, Universidad de Sevilla, Spain Keywords: Baseline energy consumption Industry Cogeneration ANN modeling Thermodynamic modeling abstract An effective modeling technique is proposed for determining baseline energy consumption in the industry. A CHP plant is considered in the study that was subjected to a retrofit, which consisted of the implementation of some energy-saving measures. This study aims to recreate the post-retrofit energy consumption and production of the system in case it would be operating in its past configuration (before retrofit) i.e., the current consumption and production in the event that no energy-saving measures had been implemented. Two different modeling methodologies are applied to the CHP plant: thermodynamic modeling and artificial neural networks (ANN). Satisfactory results are obtained with both modeling techniques. Acceptable accuracy levels of prediction are detected, confirming good capability of the models for predicting plant behavior and their suitability for baseline energy consumption determining purposes. High level of robustness is observed for ANN against uncertainty affecting measured values of variables used as input in the models. The study demonstrates ANN great potential for assessing baseline consumption in energyintensive industry. Application of ANN technique would also help to overcome the limited availability of on-shelf thermodynamic software for modeling all specific typologies of existing industrial processes. 1. Introduction The industry represents about 20% of the European Union (EU)’s primary energy consumption. Although 30% improvement in energy intensity has been achieved in this sector during the last two decades, relevant potential remains for energy efficiency enhancement (Energy Efficiency Plan, 2011). The European Commission is developing systematic plans aimed at rational energy use, like regular and mandatory energy audits in large industrial sites and incentives that EU member countries should establish for companies to implement energy management systems. Intensification of high-efficiency cogeneration plants is also included as an effective potential contributor to the energy efficiency of European industry. In this scenario, energy studies acquire a decisive role in the competitiveness of the industrial sector, the main objective of energy audits being the identification of measures aimed at increasing equipment efficiency and reduction of energy consumption. Once the saving measures proposed in the energy audit have been implemented, an issue of relevant importance is the assessment of real benefits generated by the changes. Although it is to be hoped that the actual effects are not too different from the anticipated savings, a significant margin of error exists due to, among other factors, the reference conditions which the study was based on and the variability of those conditions in the real operation of the production process. The new consumption of the factory after the retrofit can be easily obtained, provided that adequate measuring instruments are installed and functioning properly. The energy saving achieved, on the contrary, cannot be measured, since this would suppose being able to measure the current consumption of industry in its past configuration (before retrofit). For these reasons, the impact of programs and projects for improving energy efficiency in the industry can only be estimated, as the difference between the current energy consumption (directly measurable) and consumption that would occur in the same period in the event that no saving measures had been implanted. This second term is often called by the name of baseline energy consumption. The International Performance Measurement and Verification Protocol (Efficiency Valuation Organization (EVO), IPMVP Volume 1, 2012) can be considered as the international reference text on measurement and verification (M&V) of energy savings from projects aimed at improving energy efficiency and reducing energy costs. One of the aspects highlighted in the protocol is that a key factor in the planning phase of a program of M&V of savings is the selection of the method that aims to solve the problem of ⇑ Corresponding author. Tel.: +34 647004965. E-mail addresses: [email protected],[email protected] (F. Rossi). determining the baseline consumption. The need for a clear formalization of the baseline development process is also stressed in the interesting work of (Reichl & Kollmann, 2011). Among the various approaches proposed for determining the baseline consumption (U.S. EPA State Clean Energy, 2009), the most easily applicable is the engineering method, based on the use of standard formulas and assumptions for calculating the energy use before retrofit. However, the simplicity and practicality guaranteed by its implementation involves an associated level of uncertainty (Kelly Kissock & Eger, 2008) that makes it unsuitable for systematic use in the industrial field. Statistical analyses are also contemplated, but this general term is actually used to refer to models of limited complexity aimed at determining ‘‘before’’ and ‘‘after’’ consumption taking into consideration few factors like changes in weather, facility occupancy and factory operating hours. Other suggested methods contemplate computer simulation for determining system energy performance. Finally, integrative methods can be used, combining some of the aforementioned approaches. As evidenced by the literature survey (discussed below), no relevant applications of ANN modeling technique were detected for the determination of the energy consumption baseline in CHP systems, intended as the consumption of the system before the implementation of the efficiency enhancement measures. The objective of this paper is to present a valid modeling technique allowing an answer to the problem of determining the baseline energy consumption in the industry. The ANN approach is proposed, as an effective way of recognizing complex non-lineal patterns governing energy consumption in industrial production sites. Achievable levels of accuracy and suitability of ANNs in predicting the behavior of the system are also discussed, by comparing the ANN approach with the thoroughly tested thermodynamic modeling technique. A combined heat and power (CHP) plant is considered for the application of the proposed methodology. The extensive historical operating dataset available from the existing monitoring system is used to generate and validate both ANN and thermodynamic models. Cogeneration plants are often present in industrial sites and represent an efficient solution for electricity and heat supply to energy-intensive processes. According to the objective of this work, a CHP system can also be considered as an independent industrial unit, products of which are electricity and heat, e.g., in the form of steam delivered to the productive process. Below is a brief summary of the literature consulted, stating some of the latest applications of ANN modeling to industrial and cogeneration plants. Simplified methodologies have been proposed for determining baseline consumption and for verification of energy savings in the industry, based on lineal regression and multi-regression analysis (Dalgleish & Grobler, 2003; Kelly Kissock & Eger, 2008). Other more developed methods have been applied, like lineal combinations of influence variables and subsequent regression modeling for generating the baseline efficiency in industrial refining sites (Velázquez et al., 2013). The ANN modeling approach is not a novelty in the industry, as evidenced by the applications for optimization of distillation processes and control of wastewater treatment units (Liau, Yang, & Tsai, 2004; Mingzhi, Yongwen, Jinquan, & Yan, 2009; Motlaghi, Jalali, & Nili Ahmadabadi, 2008). Olanrewaju and Jimoh (2014) propose a hybrid approach based on Index Decomposition Analysis (IDA), ANN modeling and Data Envelopment Analysis (DEA), to predict the expected energy consumption in the industry and subsequently optimize the present energy use with reference to the predictedvalues. Monedero et al. (2012) apply ANN approach to model the best past operation of a petrochemical site, for comparison with present operation and efficiency optimizations purposes. ANN modeling has been frequently proposed also for cogenerations, combined cycles and power plants, with multiple objectives: - Power prediction of the plant (Smrekar, Pandit, Fast, Assadi, & De, 2010). - Prediction and monitoring of thermal efficiency and pollutant emissions of the plant (Fantozzi & Desideri, 1998; Flynn, Ritchie, & Cregan, 2005; Lu & Hogg, 2000; Pan, Flynn, & Cregan, 2007; Tronci, Baratti, & Servida, 2002). - Adjustment of power generation to demand profile of the industry (Moghavvemi, Yang, & Kashem, 1998). - Reproducing the behavior of plant components (boilers, steam turbines, and superheaters) (Bekat, Erdogan, Inal, & Genc, 2012; De, Kaiadi, Fast, & Assadi, 2007; Du et al., 2011; Jia & Xu, 2010; Ma, Wang, & Ma, 2008; Olausson, Häggståhl, Arriagada, Dahlquist, & Assadi, 2003). - Optimization of load distribution and load disconnection for stability purposes in the case of fault of the system (Hsu, Chuang, & Chen, 2011; Romero & Shan, 2005). - Optimization of operating parameters for plant efficiency maximization, in some cases through combined implementation of ANN and other optimization techniques, e.g., genetic algorithm (Arslan, 2011; Hajabdollahi, Hajabdollahi, & Hajabdollahi, 2012; Rashidi, Galanis, Nazari, Basiri Parsa, & Shamekhi, 2011; Suresh, Reddy, & Kolar, 2011). The main power generation unit of the CHP plant considered in this study is constituted by a gas turbine. Many examples exist in the literature of the ANN modeling approach applied to gas turbines, for condition monitoring and performance estimation (Asgari, Chen, Menhaj, & Sainudiin, 2013; Barad, Ramaiahb, Giridhara, & Krishnaiahb, 2012; Fast, Assadi, & De, 2009a; Fast, Palmé, & Genrup, 2009b; Lazzaretto & Toffolo, 2001; Nikpey, Assadi, Breuhaus, & Mørkved, 2014; Palme ´, Breuhaus, Assadi, Klein, & Kim, 2011), fault detection (Kong,Ki, Kang, & Kho, 2004; Kong, Park, & Kim, 2008; Simani & Fantuzzi, 2000; Volponi, DePold, Ganguli, & Daguang, 2003), control enhancement(Sisworahardjo,El-Sharkh,&Alamb,2008),andturbine diagnostics (Bettocchi, Pinelli, Spina, Venturini, & Burgio, 2004). Other examples of ANN methodology applied to industry for online monitoring, performance estimation and diagnostic purposes are represented by modeling of boilers (Arriagada, Costantini, Olausson, Assadi, & Torisson, 2003; Chong, Wilcox, & Ward, 2000; Chu et al., 2003; Hao, Kefa, & Jianbo, 2001; Rusinowski & Stanek, 2007; Teruel, Cortes, Diez, & Arauzo, 2005), furnaces (Calisto, Martins, & Afgan, 2008), circulating fluidized bed boilers (Liukkonen et al., 2011), compressors (Ghorbanian & Gholamrezaei, 2009), refrigerating and heating systems (Kizilkan, 2011; Esen and Inalli; 2009), and fluidized bed dryers (Chen, Tsutsumi, Lin, & Otawara, 2005; Satish & Setty, 2005). Other industrial applications of ANN involve more specific areas such as control of nuclear power plants (Oliveira & Soares de Almeida, 2013), performance prediction of internal combustion engines (Canakci, Ozsezen, Arcaklioglu, & Erdil, 2009; Og ˘uz, Saritas, & Baydan, 2010; Tasdemir, Saritas, Ciniviz, & Allahverdi, 2011; Yap & Karri, 2013), modeling and control of fuel cells (Chávez-Ramírez et al., 2010; Hatti & Tioursi, 2009), performance prediction and optimization of solar systems (Kalogirou, 2004; Sözen & Akçayol, 2004), power output forecasting of wind power plants (Yeh, Yeh, Chang, Ke, & Chung, 2014), control of hybrid wind-diesel power generators (Vargas-Martínez, Minchala Avila, Zhang, Garza-Castañón, & Calle Ortiz, 2013), and prediction performance of photovoltaic panels (Karamirad,Omid,Alimardani,Mousazadeh,& NavidHeidari, 2013). 2. Description of CHP plant, implemented energy-saving measures and preliminary data processing The system at issue is a CHP plant at the service of a refining site. Fig. 1shows a simplified schematic of the plant. The main components of the unit can be appreciated: Gas turbine (GT). Generates electricity from fuel burned in the combustion chamber, both for refinery use and for export to the grid of the production which exceeds demand. Heat recovery steam generator (HRSG). Equipment constituted by heat exchangers the function of which is to recover the heat contained in the flue gases at the outlet of the GT and contemporary steam generation. Post-combustion is also used to increase steam generation and meet the thermal demand of the refinery. Steam turbine (ST). Used for additional generation of electricity from steam generated in the HRSG. Part of the steam is extracted from the turbine before it completes its expansion, to be supplied to refinery at the proper pressure level required by the process. The remaining steam completes the expansion in the turbine to be delivered to the process at a lower pressure level. From April to November 2008 a number of energy-saving projects were implemented in the CHP unit, by which the economic benefit of the plant was increased significantly. Fig. 1 evidences areas of implementation of projects, which were: Replacement of existing burners with Low-NO x burners in the GT combustion chamber. It allowed the cessation of steam injection into the combustion chamber, performed until meeting the emission levels required by legislation. Preheating of deaerator feed water and installation of a new economizer in the HRSG. A new heat exchanger was installed that preheats the water fed to the deaerator, contemporarily cooling HRSG feed water. The first effect of this measure was the reduction of steam consumption in the deaerator. A new heat exchanger was also installed in the HRSG (new economizer), to compensate the lower HRSG feed water temperature. As a second effect of the project, exploitation was thus enhanced of the GT gases heat content and stack gas temperature was reduced from 190 to 130 °C. The main effect of all implemented measures was a significant reduction in steam self-consumption of the plant, which explains the subsequent increase of CHP plant profit. Moreover, saving projects impacted on other consumptions and productions, which also contribute significantly to the operating profit of cogeneration. The parameters directly affected by the projects, for which the construction of the baseline is required, are: Electricity production and fuel (natural gas) consumption in the GT. Medium pressure (MP) steam injected into the GT. Low pressure (LP) steam consumed in the deaerator. As mentioned above, two different modeling methodologies were applied to the CHP plant, i.e., thermodynamic modeling and ANN modeling. Prior to that, a filtering of the collected data was necessary so that a valid database could be obtained for development of both models. A comprehensive data collection was conducted for most of the process variables. Records were obtained from distributed control system (DCS), through specific software that allows downloading historical data in Excel format. Plant monitoring is very exhaustive, and hourly data were collected for 134 variables, comprised between 02-16-2007 18:00 and 05-072009 00:00. These variables include the main operating parameters of all the components of the plant: GT: electricity generation, GT regulation parameters (e.g., inlet guide vane opening angle), pressure drop in air filters, properties of fuel, inlet air, exhaust gases and injected steam. HRSG: conditions of fuel fed to post-combustion, properties of exhaust gases and water/steam in the economizer, evaporator and superheaters, blowdown rate and stack exhaust gas temperature. Deaerator: operating pressure, properties of fed steam and boiler feed water. ST: electricity generation, properties of inlet live steam, extraction steam and steam to condenser. Other components: properties of streams in the existing heat exchangers of the plant for boiler feed water preheating, steam de-superheating stations, etc. Previous data treatment was performed and anomalous registers were eliminated, associated with measurement faults or informatics errors in the acquisition process. A valid database for the Evaporator Deaerator Heat Recovery Steam Generator Stack Economizer Gas Turbine Superheater Process Steam Turbine LP steam MP steam Exhaust gases Boiler feedwate r New lowNOx burners Preheating of deaerator feedwater New economizer Ambient air Fuel (postcombustion) Fuel Treated water Returned condensates Fig. 1. Simplified diagram of the CHP plant and areas subject to energy-saving measures. following modeling phase was subsequently selected. The effectiveness and the outcome of this selection process are strongly bound to the level of knowledge of the operation of the plant. A deep mastery of the philosophy of operation and control of the system is fundamental to the purpose, and it always derives not only from the prior knowledge of the system, but also from the observation of the same recorded data. A good understanding of the system is thus necessary, whatever the method of modeling elected. 3. ANN modeling ANN is an adaptive non-linear statistical data modeling technique consisting of interconnected artificial neurons processing data in parallel. Exhaustive introductions to ANN modeling are provided in most of the above-referenced works. The multi-layer perceptron (MLP) network type was employed in this study, and the supervised learning process was used for adjustment of network parameters. Before the training, the dataset was randomized and divided into training and test data subsets. The training set is used for weight adjustment and the test set is used after the training to assess the performance i.e., prediction capability of ANN by comparing calculated and desired values for a set of unknown data. 3.1. ANN setup and training A multi-layer feed-forward structure was chosen for the ANN. A specific ANN was generated for each output variable (GT electricity production, GT fuel consumption, MP and LP steam consumptions) and IBM SPSS Modeler 14.1.0 software package was used to setup and train the models. Input variables in the models were selected by the authors, based on the knowledge of system physics. Input used for GT electricity production and GT fuel consumption were the ambient conditions (temperature, atmospheric pressure and humidity), the density and the lower heating value of the fuel, NO x emissions and a counter representing the aging status of the GT. The same inputs were used for GT steam injection, except for fuel density, which has not a significant influence over it. The steam consumption of deaerator was modeled as a function of its operating pressure, the ambient temperature, the steam production of the HRSG and the hot condensate mass flow returned from the process for preheating of deaerator feed water. For each of the generated ANN, the output layer is represented by one neuron, while the number of hidden layers and neurons is not fixed, being two the maximum number of hidden layers contemplated in the study. The final architecture of the network was determined by a trial-and-error process, in which different trainings were performed with varying numbers of hidden layers and neurons, until the network with best prediction accuracy was selected. A first set of trainings was performed, observing the behavior of one-layered ANN with increasing number of neurons from 1 to 15, fixed as the maximum limit of neurons in the analysis. Further increase of ANN complexity was considered unnecessary; as observed by Calisto et al. (2008), an excessive number of neurons, rather than having beneficial effects, may facilitate the overtraining and reduce the generalization capability of the network. ANN structures presenting local maxima in prediction performance were selected and the effect was observed of introducing a second hidden layer. The number of neurons in the second layer was increased for each selected ANN, from one to the same number of neurons of the first hidden layer. As suggested by Kong and Goo (2011), hyperbolic tangent sigmoid function and identity function were used as transfer functions in the hidden and output layers respectively. Range fields of data were rescaled before training to have values between 0 and 1, to match the range of the activation function of the first hidden layer. The common and well proven back-propagation algorithm proposed by Rumelhart, Hinton, and Williams (1986) was applied to train the ANN and update the weights associated with neurons. Data are processed in the forward direction at each iteration (epoch) and output values are generated in the output layer. The calculated error between the desired output and the model prediction subsequently propagates in the backward direction by means of Gradient Descent algorithm with Momentum, a computationally efficient technique that allows minimizing the error by updating the connection weights. This method, widely explained in Moghavvemi et al. (1998), contemplates the use of two coefficients referred to as learning rate ( g ) and momentum rate (m). Coefficient g is used to control the intensity of the weights modification: fast learning is obtained with high g values, but it may lead to instability problems. Momentum is used to correct the current weight variation with a term that reflects the impact of past weight change. The weight vector is updated using the formula w kþ1 ¼w k  g k g k þm D w k where krefers to the number of epochs and g is the gradient of the error function E(w) with respect to the weights, in turn defined as EðwÞ¼X NR r¼1 E ðrÞ ðwÞ being E ðrÞ ðwÞ¼1 2X N n¼1 t ðrÞ n p ðrÞ n  2 where NR is the number of records, Nis the number of network outputs, t n is the target output value for output n and p n is the corresponding prediction generated by the network. The initial value of g was set to 0.4 and mwas fixed to 0.9. As proposed by Liang-yu, Jian-qiang, and Bing-shu (2002) and Kong and Goo (2011), dynamic variation of learning rate was used to increase stability of training process. Variable adjustments were applied to learning rate factor depending on the error behavior: no correction was used for error decrease and strong reduction is applied for error increase (0.5 factor of correction), to prevent network unsteady problems. The factor m was kept constant, except for the situations in which g k jg k jmj D w k j In these cases mvalue is set at m¼0:9 g k jg k j j D w k j This correction being necessary to make sure that the weight change takes place in the right direction for error decrease. 3.2. Prediction capability of the model A total of 6405 records were used for the training of the network. Data were pre-randomized and subsequently grouped in two different sets: 70% of data were used for training and the remaining 30% were used for testing at each epoch. The records of the training set were internally divided into a model building set and an overfit prevention set, which is used to track errors during training in order to avoid overtraining of the network, i.e., to prevent the network from becoming overly ‘‘specialized’’ in recognizing the patterns which it was trained with, thereby losing its generalization capability. The percentages of training data employed for model building and overfit prevention were 85% and 15%, respectively, corresponding to 59.5% and 10.5% of total records. The criteria selected for performance evaluation was the accuracy coefficient defined by the formula: Accuracy ¼1 NR X NR r¼1 1jt r p r j max i ðt i Þmin i ðt i Þ 0 @1 A where NR is the number of records, t r is the target output value of record rand p r is the corresponding prediction generated by the network. The terms max(t i ) and min(t i ) in the denominator refer to the maximum and minimum values among all the targets, and subscript i was used to indicate that they do not vary with the index r. Mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination R 2 (square of the Pearson product moment correlation coefficient R) were also calculated, to complete prediction capability analysis of the generated models: MAE ¼1 NR X NR r¼1 t r p r t r  RMSE ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 NR X NR r¼1 t r p r t r  2 v u u t R¼X NR r¼1 ðt r  tÞðp r  pÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi X NR r¼1 ðt r  tÞ 2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi X NR r¼1 ðp r  pÞ 2 q Table 1 summarizes the accuracy levels achieved for GT electric power in the selection process of the final structure of the ANN model. Limited obtained variations of predictive capability confirm the slight influence of hidden neurons number on ANN performance observed by Bettocchi, Pinelli, Spina, and Venturini (2007) when data are affected by measurement errors. Local maxima can be observed for 3, 6 and 8 neurons in the first hidden layer. Maximum accuracy is calculated for 8 neurons and no further improvements are obtained augmenting neurons number. For that reason 3, 6 and 8-layer structures are selected for checking the effect of introducing a second hidden layer. As it may be observed, second hidden layer has no benefit for the case of 3 neurons, whereas slight improvements can be observed for any two-layer structures containing 6 and 8 neurons in the first hidden layer. The final configuration selected for ANN contains 8 neurons in the first layer and 3 neurons in the second, although the accuracy enhancement achieved with the second layer introduction is not relevant. Fig. 2 shows obtained results for all generated ANN models. As it can be observed, the initial date of the modeling period does not coincide with the aforementioned one. Indeed the elimination of registers corresponding to the first months of operation was required, due to a prolonged faulty measurement of NO x . The final dataset was so reduced to data comprised between 06-12-2007 07:00 and 04-21-2008 17:00 (6405 data). The first column of Fig. 2 contains the simultaneous plots of ANN predictions and measured registers, along with corresponding errors for each data point. The same information is presented in the second column, in the form of cross plots of predicted values against measured values. Good agreement between predicted and actual values can be observed from the first column, and correct random (and not systematic) behavior of deviations from the diagonal in the second column. A strong smoothening tendency is observed in the deaerator steam consumption for high steam flow rates (bottom left graph of Fig. 2). This is mainly due to the low number of available data for training the model in that period, which does not reflect the normal operation of the plant. The plant in its original pre-retrofit configuration includes a heat exchanger for preheating of deaerator feed water with hot condensates returned from the process. During the period at issue the heat exchanger was not receiving condensates, thus augmenting the steam consumption in the deaerator. Apart from these aspects, the relative errors calculated in that zone are still acceptable and comparable with those obtained in the rest of the modeled period. Additional graphs are also presented in Fig. 3, showing the same results of Fig. 2 but for a smaller period of time (two weeks). This may be useful to observe the graphical correspondence between the values of accuracy and R 2 achieved and the capacity of the models to reproduce the registered profiles, in terms of flexibility and ability to follow the tendencies and fluctuations of measured data. With reference to Table 2, satisfactory values of all calculated performance parameters may be observed in GT electricity production, demonstrating high model capability in predicting the outputs. Almost 89% of the points present errors lower than 1.0% and only 0.9% are higher than 2%. Lower accuracy was obtained from GT fuel consumption. The tendency of the model to smoothen fluctuations of registered data may be observed in the generated plots and is also reflected by the low value of R 2 . The strong influence of input data quality over model performance has to be considered in this specific case. Indeed the fluctuations of registered Table 1 Selection of ANN structure for modeling of GT electricity production. Obtained prediction accuracies. No. of neurons in the first hidden layer 123456789 101112131415 Obtained accuracy (%) 94,557 94,715 95,400 95,070 94,912 95,330 94,880 95,837 95,150 95,663 95,523 95,524 95,605 95,722 95,620 No. of neurons in the firstsecond hidden layer 3-1 3-2 3-3 Obtained accuracy (%) 94,768 95,399 95,341 No. of neurons in the firstsecond hidden layer 6-1 6-2 6-3 6-4 6-5 6-6 Obtained accuracy (%) 95,150 95,268 95,493 95,393 95,495 95,485 No. of neurons in the firstsecond hidden layer 8-1 8-2 8-3 8-4 8-5 8-6 8-7 8-8 Obtained accuracy (%) 95,310 95,532 95,850 95,338 95,508 95,254 95,091 95,600 One-layer feed-forward structure Two-layer feed-forward structure data are not representative of the real GT operation. They are mainly due to the different frequencies of data acquisition for the two parameters, product of which gives GT fuel consumption in MW: fuel mass flow rate and fuel heating value. The first is measured on an hourly basis and the latter proceeds from the daily laboratory analysis. Nevertheless model prediction capability is still acceptable, since almost 86% of the calculated errors are lower than 2.0% and only 0.4% are higher than 5%. Models generated for GT steam injection and steam consumption of deaerator reproduce correctly tendencies of registered data, as confirmed by the calculated values of R 2 and accuracy. As it was expected, higher MAE and RMSE values were calculated for these two models, due to the relevant error level commonly associated with on-line measurement of steam flow rates. On average, 85% of errors are lower than 5.0% for these two modeled parameters. Summarizing, obtained accuracy levels and errors allow asserting that the generated ANN models captured correctly the behavior of the CHP plant. The generalization capability is regarded as satisfactory for baseline energy consumption determining purposes, since the order of magnitude of the obtained errors is comparable with the uncertainty associated with measuring instruments. 4. Thermodynamic modeling The same energy consumption and generation parameters were modeled using a thermodynamic simulator. The model was performed with Thermoflex (release 23.0), (Thermoflow Inc., 2013) a specific software for thermal systems modeling specially conceived for power generation and CHP plants. 4.1. Model description The generated model reproduces the CHP plant configuration in the period prior to the implementation of saving projects. Fig. 4 shows the diagram of the model used for the simulation. As it can be observed, ST was not included in the model, due to the fact that the operation and electricity production of ST was not affected by energy saving measures. Indeed, the electricity generated by the ST is determined by the conditions (mass flow, temperature and pressure) of inlet live steam and the amount of steam extracted. The production and conditions of live steam entering the turbine are continuously controlled by means of the post-combustion fuel regulation system. It implies that the changes caused by saving -5,0 -2,5 0,0 2,5 5,0 7,5 10,0 12,5 15,0 32 33 34 35 36 37 38 39 40 12/06/2007 10/07/2007 07/08/2007 04/09/2007 02/10/2007 30/10/2007 27/11/2007 25/12/2007 22/01/2008 19/02/2008 18/03/2008 15/04/2008 Error (%) GT electricity production (MW) Time (days) Measured Predicted 33 34 35 36 37 38 39 40 41 34 35 36 37 38 39 40 GT electricity production (MW) precited GT electricity production (MW) measured -9,0 -6,0 -3,0 0,0 3,0 6,0 9,0 12,0 15,0 18,0 21,0 24,0 90 95 100 105 110 115 120 125 130 135 12/06/2007 10/07/2007 07/08/2007 04/09/2007 02/10/2007 30/10/2007 27/11/2007 25/12/2007 22/01/2008 19/02/2008 18/03/2008 15/04/2008 Error (%) GT fuel consumption (MW) Time (days) Measured Predicted 100 105 110 115 120 125 130 135 140 110 114 118 122 126 130 134 GT fuel consumption (MW) precited GT fuel consumption (MW) measured -20,0 -10,0 0,0 10,0 20,0 30,0 40,0 50,0 0 1 2 3 4 5 6 7 12/06/2007 10/07/2007 07/08/2007 04/09/2007 02/10/2007 30/10/2007 27/11/2007 25/12/2007 22/01/2008 19/02/2008 18/03/2008 15/04/2008 Error (%) Steam injection into GT (t/h) Time (days) Measured Predicted 1 2 3 4 5 6 7 8 2345678 Steam injection into GT (t/h) precited Steam injection into GT (t/h) measured -20,0 -10,0 0,0 10,0 20,0 30,0 40,0 50,0 60,0 70,0 0 5 10 15 20 25 12/06/2007 10/07/2007 07/08/2007 04/09/2007 02/10/2007 30/10/2007 27/11/2007 25/12/2007 22/01/2008 19/02/2008 18/03/2008 15/04/2008 Error (%) Steam consumption in deaerator (t/h) Time (days) Measured Predicted 2 6 10 14 18 22 26 30 9 121518212427 Steam cons. in deaerator (t/h) precited Steam cons. in deaerator (t/h) measured Fig. 2. ANN models. Predictions, measured registers and corresponding errors. measures in the conditions of GT exhaust gases do not affect the conditions of the steam entering the ST. On the other hand, the percentage of steam extracted from the turbine only depends on the process demand of MP and LP steam. As previously commented, steam auto-consumptions of the plant decreased as a consequence of the implementation of the energy conservation measures. This caused an increase of MP and LP steam exported to the process. In case the process not having the capacity to absorb these changes, this would have led to modify the form of controlling steam extraction from the ST. Nevertheless, no excess of steam export was registered, since the exportation increase was compensated by load reduction of other auxiliary equipment supplying steam to the process. For these reasons, the operation of the ST did not suffer changes, thus confirming the independency between ST electricity generation and the energy saving measures. The main components of the plant can be observed: GT, HRSG, heat exchangers train for preheating of treated water, and deaerator. The model reproduces the main plant control philosophy: a control loop was implemented for HRSG load, which allows assigning live steam flow produced as an input variable to the model. For model adjustment, the plant was simulated in off-design conditions and model outputs were compared with measured data, until satisfactory matching between data and measurements was obtained. The model tuning was performed manually, although Thermoflow special features would be available for carrying it out automatically. This way, a model was obtained capable of simulating GT partial load operation and reproducing the actuation of the real control system: GT load can be specified in the simulator both in the form of a specific GT electricity production and as a fixed percentage of GT nominal power. The software allowed endowing the model with the appropriate level of detail to reproduce the actual complexity of plant operation. For instance, the boiler feed water turbopump was included in the scheme, driven by MP steam proceeding from desuperheating section, as well as the flash tank for LP steam recovery from HRSG blowdown. 4.2. Preliminary assumptions According to the plant operation protocol prior to the installation of the new Low-NO x burners, the steam flow rate injected into the GT was manually controlled to maintain fixed levels of NO x emissions in GT exhaust. Such control was hardly reproducible in the model, since the control of GT steam injection for reduction of NO x emissions is not directly implementable in the software. A fixed value of GT steam-to-fuel ratio was assumed for the post-retrofit period, equal to the average recorded value in the pre-retrofit period, thus reducing somewhat the precision of the model. In return, the first months of operation of the CHP plant were not discarded for the simulation, as it was necessary for the case of ANN modeling. Indeed NO x emissions were not used as an input in the model and there was no reason to discard the period of faulty NO x measurement. The adjustment of some parameters depending on operating hours was also necessary to complete the setup of the model, e.g., GT fouling parameters. Since time is not contemplated among the inputs of the thermodynamic model, the effects of aging had to be determined separately for all the input variables directly depending on number of operating hours. A simplification was adopted for those variables, and the same trends observed in the pre-retrofit period were assumed for the GT degradation factors and air pressure losses in GT air admission line. 4.3. Validation and prediction capability of the model For GT full load operation, brief periods of operation were selected to validate the model’s ability to simulate consecutive days of operation. The selection was carried out so that different scenarios of plant operation were covered, according to the main operating parameters of the plant: GT electricity production, fuel burned in post-combustion, steam injection in GT, etc. A total of 500 h of operation were simulated to validate the model at full load. For partial load operation all recorded data were used, due to their smaller amount: 280 measured data. For the validation, a tool was developed which allows performing massive simulations. An Excel spreadsheet was used to link to the simulation file that allowed automatically simulating a number of times set by the user. Thermoflex software provides Visual Basic (VBA) code for that purpose. Only few modifications had to be introduced to adapt it to the specific simulation model. Fig. 5 shows validation results for GT electricity production at full load. Data are presented sorted by date, so that performed selection can be observed, along with the strong influence of ambient temperature over the evolution of GT electrical output. For clarity, the simulated sections were unified and displayed adjacent in the second diagram. 36 37 37 38 38 39 39 40 40 01/12/2007 03/12/2007 05/12/2007 07/12/2007 09/12/2007 11/12/2007 13/12/2007 15/12/2007 GT electricity production (MW) Time (days) Measured Predicted 115 117 119 121 123 125 127 129 131 133 135 01/12/2007 03/12/2007 05/12/2007 07/12/2007 09/12/2007 11/12/2007 13/12/2007 15/12/2007 GT fuel consumption (MW) Time (days) Measured Predicted 3,5 4,0 4,5 5,0 5,5 6,0 6,5 7,0 01/12/2007 03/12/2007 05/12/2007 07/12/2007 09/12/2007 11/12/2007 13/12/2007 15/12/2007 Steam injection into GT (t/h) Time (days) Measured Predicted 12 14 16 18 20 01/12/2007 03/12/2007 05/12/2007 07/12/2007 09/12/2007 11/12/2007 13/12/2007 15/12/2007 Steam consumption in deaerator (t/h) Time (days) Measured Predicted Fig. 3. ANN models. Predictions, measured registers and corresponding errors. Results for one month. Obtained error levels were relatively low, although sensibly higher than the corresponding ANN model. Only 57% of the points presented errors lower than 1.0%, compared to 89% of ANN predictions fulfilling the same condition. Nevertheless, the prediction capability of the model may be still considered satisfactory, as 83% of calculated errors were lower than 2% and only 0.1% higher than 5%. No systematic error behavior was detected and the same random tendency was observed as for ANN models. For the sake of concision, results are not shown for the other modeled parameters. Such results confirmed the more than proven ability of the thermodynamic simulator to predict the behavior of this type of plants, although with somewhat less precision of ANN when it comes to reproducing a set of measured operating data. This can be mainly attributed to the high sensitivity of thermodynamic models to the quality of input data. The well-known data reconciliation (DR) technique would certainly help to correct the measurement errors, thus reducing its adverse effects on the model. For this purpose Thermoflow conceived a special DR software package, the application of which would contribute to enhance the model capacity of prediction. 5. Results Fig. 6 shows the energy baselines obtained for the CHP plant, proceeding with the application of thermodynamic modeling technique. Results are indicated for the main consumption and production parameters of the plant, which were affected by the implementation of the previously described energy conservation measures: Replacement of existing burners with Low-NO x burners in the GT combustion chamber. Preheating of deaerator feed water and installation of a new economizer in the HRSG. The baselines behave as expected. The baselines of electricity production and GT fuel consumption are higher than measurements. It is mainly due to the cessation of steam injection into the GT and the installation of a new stage of filtration in the GT air inlet line; a second filter had to be integrated in the system to prevent the acceleration of GT fouling, since the new Low-NO x burners imposed the cessation of ‘‘carbo-blasting’’ a cleaning method based on injection of solid natural media. Therefore, periodic load reductions can be observed for GT cleaning, which would take place if the new burners had not been installed, as well as a corresponding decrease in steam injection. In the case of ANN modeling it was thought inappropriate to model the cleaning operation mode, due to the small number of records available. This approach has no influence on the results, given the low number of operating hours in cleaning mode and considering the ultimate goal of the study, which is the determination of the economic benefit of implementing energy-saving measures. Among the obtained trends, exceptional crossings occur between measured data and energy baselines. For GT fuel baseline consumption, crossings may be observed during continuous intervals of operation. These isolated cases are likely to be attributed to measurement errors. Actually these values were removed for ANN modeling for being out of range. In Fig. 7 both energy baselines predicted by ANN Fig. 4. Diagram of CHP simulation model (Thermoflex software). -5,0 -2,5 0,0 2,5 5,0 7,5 10,0 12,5 15,0 17,5 20,0 31 32 33 34 35 36 37 38 39 16/02/2007 16/03/2007 13/04/2007 11/05/2007 08/06/2007 06/07/2007 03/08/2007 31/08/2007 28/09/2007 26/10/2007 23/11/2007 21/12/2007 18/01/2008 15/02/2008 14/03/2008 11/04/2008 Error (%) GT electricity production (MW) Time (days) Measured Predicted -5,0 -2,5 0,0 2,5 5,0 7,5 10,0 12,5 15,0 17,5 20,0 31 32 33 34 35 36 37 38 39 0 100 200 300 400 500 600 700 800 Error (%) GT electricity production (MW) Points Measured Predicted Fig. 5. Thermodynamic model. Simulations, measured registers and corresponding errors. GT electricity production at full load. Selection of operation periods (left) and unification of selected periods (right). and thermodynamic models are represented, together with the percentage differences between them. The results for predicted GT electricity production and fuel consumption indicate relatively low differences between generated baselines, in the order of 3–5%. Predictions obtained from ANN modeling for these two parameters show a tendency to exceed thermodynamic baseline in the second part of the observed period. A source of inaccuracy which the differences could be addressed to is the approximate approach that was adopted to reproduce GT aging in thermodynamic modeling. On the one hand, aging parameters were used as an adjustable input to better reproduce measured data when validating the thermodynamic model for pre retrofit period. On the other hand, those parameters generate an environment of uncertainty in the post retrofit period: the same trends were reproduced as for the pre retrofit period and there is no warranty that it corresponds to the real CHP operation. Another possible cause of the discrepancies could be associated with the fuel feed temperature, not contemplated in the ANN model. After the implementation of energy saving measures, fuel preheating was introduced as a practice in plant operation. This resulted in a slight reduction of GT fuel consumption and electricity production. Fuel temperature could not be included as an input to ANN models, since fuel was supplied to GT at ambient temperature in pre retrofit period and the slight fluctuations registered were not sufficient to appreciate its effects. More relevant differences between the models were detected for steam injected into the GT, the differences ranging up to 30–35%. In this case better results were achieved with the ANN modeling approach. The discrepancies are mainly associated with the limitations of thermodynamic model when reproducing the real plant control for NO x emissions reduction, as discussed above. Concerning the effect of the second saving project (pre-heating of deaerator feed water), the differences are 30 32 34 36 38 40 42 17/11/2008 07/12/2008 27/12/2008 16/01/2009 05/02/2009 25/02/2009 17/03/2009 06/04/2009 26/04/2009 GT electricity production (MW) Time (days) Actual Baseline 0,0 0,5 1,0 1,5 2,0 2,5 3,0 3,5 4,0 4,5 5,0 17/11/2008 07/12/2008 27/12/2008 16/01/2009 05/02/2009 25/02/2009 17/03/2009 06/04/2009 26/04/2009 Steam injection into GT (t/h) Time (days) Actual Baseline 100,0 106,0 112,0 118,0 124,0 130,0 136,0 17/11/2008 07/12/2008 27/12/2008 16/01/2009 05/02/2009 25/02/2009 17/03/2009 06/04/2009 26/04/2009 GT fuel consumption (MW) Time (days) Actual Baseline 0 3 6 9 12 15 18 21 17/11/2008 07/12/2008 27/12/2008 16/01/2009 05/02/2009 25/02/2009 17/03/2009 06/04/2009 26/04/2009 Steam consumption in deaerator (t/h) Time (days) Actual Baseline Fig. 6. Predicted and measured energy production and consumptions of CHP in the post-retrofit period. -3,0% 0,0% 3,0% 6,0% 9,0% 12,0% 15,0% 18,0% 25 27 29 31 33 35 37 39 41 17/11/2008 07/12/2008 27/12/2008 16/01/2009 05/02/2009 25/02/2009 17/03/2009 06/04/2009 26/04/2009 GT electricity production (MW) Time (days) Baseline_ANN model Baseline_Termodynamical model Difference (%) -6,00% -3,00% 0,00% 3,00% 6,00% 9,00% 12,00% 15,00% 18,00% 80,00 90,00 100,00 110,00 120,00 130,00 17/11/2008 07/12/2008 27/12/2008 16/01/2009 05/02/2009 25/02/2009 17/03/2009 06/04/2009 26/04/2009 GT fuel consumption (MW) Time (days) Baseline_ANN model Baseline_Termodynamical model Difference (%) -40,0% -20,0% 0,0% 20,0% 40,0% 60,0% 80,0% 100,0% 120,0% 140,0% 160,0% 0,0 1,0 2,0 3,0 4,0 5,0 6,0 7,0 17/11/2008 07/12/2008 27/12/2008 16/01/2009 05/02/2009 25/02/2009 17/03/2009 06/04/2009 26/04/2009 Steam injection into GT (t/h) Time (days) Baseline_ANN model Baseline_Termodynamical model Difference (%) -15,0% 0,0% 15,0% 30,0% 45,0% 60,0% 75,0% 90,0% 105,0% 0 2 4 6 8 10 12 14 16 18 20 22 17/11/2008 07/12/2008 27/12/2008 16/01/2009 05/02/2009 25/02/2009 17/03/2009 06/04/2009 26/04/2009 Steam consumption in deaerator (t/h) Time (days) Baseline_ANN model Baseline_Termodynamical model Difference (%) Fig. 7. Energy baselines obtained with ANN and thermodynamic modeling.