Simplifying decision-making in the model-based co-design of building energy systems through automatically generated optimal controllers

Simplifying decision-making in model-based co-design of building energy systems through automatically generated optimal controls Karl Walther1,2, Louis Hermans1,2, Lone Meertens1,2, Lieve Helsen1,2 1KU Leuven, Leuven, Belgium 2EnergyVille, Genk, Belgium Abstract Energy-efficient technologies and efficient engineering processes are two major prerequisites for the fast and cost-efficient energy transition of the built environment. This study analyzes the potential of automatically generated optimal controls to not only enable high-performing building operation but also to simplify and streamline the decision-making process in the design phase. As a case study, the re-design of the energy supply system for a historic neighborhood in Bruges, Belgium, fully based on renewable and residual energy sources (R2ES) with air-source and ground-source heat pumps, photovoltaic thermal (PVT), and photovoltaic (PV) systems is used. Two design procedures, either using manually generated Rule-Based Control (RBC) or automatically generated Optimal Control (OC), using the Toolchain for Automated Control and Optimization (TACO), are compared in terms of the engineering workflow and the system performance. The results illustrate that, for this case study, an optimally controlled system uses around 30 % less electricity compared to rulebased control. Moreover, the potential of automatically generated OC to simplify the decision-making process in model-based co-design processes and as system integrator is highlighted. Key innovations •Use of automatically generated OCs for the heating, ventilation, and air conditioning (HVAC) system design. •Integrated co-design of HVAC systems, including air-source and ground-source heat pumps, borefields, passive cooling, and PVT collectors in a historic case study neighborhood. •Comparison of system performance and engineering and decision making workflow for modelbased design using conventional manually developed RBCs and automatically generated OCs. Practical implications Using automatically generated OC already at the design stage has the following implications on practical HVAC and control engineering workflows: (1) the fact that the OC is generated based on a physics-based design model enables taking into account the OC performance for the integrated building and HVAC design. (2) The fact that TACO automatically generates OCs significantly reduces the effort for the control development which is a bottleneck in model-based design using manually developed conventional RBCs. From a technical perspective, the results highlight the potential of integrated hybrid systems including air-source and ground-source heat pumps with PVT for borefield regeneration. Introduction Motivation In the frame of the energy transition, existing fossilfuel based energy systems in buildings and neighborhoods1will be largely re-designed in the next years to integrate R2ES. To find optimal overall solutions, an integrated co-design of HVAC systems, controls, and building physics is required (Kramer et al. 2017). Such integrated co-design is, however, a highly challenging engineering task due to the increasing complexity of heat-pump-based supply concepts that integrate various R2ES and that offer flexibility to heat and electricity networks through demand-side management (DSM). Hence, the methods and tools for the design and decision-making are crucial to efficiently design high-performing project-specific solutions. This paper focuses on the HVAC and control design tasks before the tendering process when HVAC system components, their sizes, their hydraulic connection, and their control are determined. The following options with different process quality levels can be distinguished: Building practice Today’s HVAC control sequences in buildings are predominantly based on RBCs. These RBCs are drafted based on static considerations in single operation modes during the design phase. Design concepts are usually documented using graphical control schemes and textual ‘functional descriptions’ / ‘sequences of operation’ that are manually translated into control code during the execution phase. The programming generally includes (1) the selection of a suitable control approach for the documented control concept (2) the 1The present study uses existing buildings and neighborhoods as examples, however, the methods are equally applicable to new buildings. Proceedings of the 19th IBPSA Conference Brisbane, Australia, Aug 24-27, 2025 0001 https://doi.org/10.26868/25222708.2025.1690 selection of suitable control function blocks (proportional–integral (PI), hystersis, etc.), (3) the correct connection of inputs and outputs of function blocks, and (4) the parametrization. Common issues related to this building practice are the lack of integrated design and control verification, oversized HVAC systems, error prone manual processes, or missing digital control documentation (Torabi et al. 2022). Model-based design with manually generated RBCs Model-based design using a physics-based (white-box) approach in Building Performance Simulation (BPS) environments enables the integrated engineering of HVAC systems and controls, and is a strategy to handle the complexity of future supply systems. New toolchains, such as the Modelica-based Control Description Language (CDL) (Wetter et al. 2022), or the IEC 61131-3 programmable logic controller (PLC) code simulation in IDA ICE (Walther 2025), allow replacing analog (textual and graphical) documents by digital formats for the seamless implementation on building controllers. (Walther 2025) A critical practical issue of manually developing custom RBCs in a fully integrated model-based HVAC design process is that the effort in the design phase significantly increases. This is particularly problematic in the context of the shortage of skilled labor. Model-based design with automatically generated OCs It is well documented in the literature that the HVAC system performance can be significantly increased through OC taking into account the future model behavior for a given time horizon (Drgoˇna et al. 2020). OC is usually implemented during operation, often using data-driven approaches (grey-box, black-box), to replace conventional RBCs that were the basis for the initial HVAC design (see previous paragraphs). A drawback of data-based OCs is that the OC performance can not yet be taken into account for the design of new and renovated HVAC systems, because measurement data is not yet available. Potential benefits from the interaction of OC with the HVAC system layout might be overlooked. Physics-based approaches are required to take the OC performance into account during the HVAC design. TACO (Jorissen, Boydens, et al. 2019) enables exactly this approach by using physics-based models in the Modelica language for the OC generation. The particular advantage of TACO regarding the engineering effort is that these optimal control trajectories are automatically generated while custom RBCs need to be largely generated manually. Research gap, research questions, contribution, and structure This study investigates the benefits of using automatically generated OCs already at the design stage. This represents a completely new approach for engineering practice, and stimulates applying modelbased design in the construction industry. The OCs are generated by TACO in a 1 year Optimal Control Problem (OCP).2The application of model-based design with manually generated RBCs is compared against model-based design with automatically generated OCs using a real design process in a case study neighborhood. Existing studies that compare RBC and OC typically focus on the improved HVAC and control performance either in virtual simulation studies (Ramesh et al. 2024) or field implementations (Saloux et al. 2025). The RBC vs. OC comparison in this study focuses on the potential of automatically generated OCs to simplify and facilitate practical design and decision-making processes compared to manually developed RBCs. The central research questions are (1) how do the control development and testing processes in practice differ between manually developed RBCs and automatically generated OCs, and (2) how does each option influence the decision-making process. The remainder of this paper is organized as follows: First, the methodology for the comparison of manually developed RBCs vs. automatically generated OCs, including the case study neighborhood, is presented. Then, the results are discussed with respect to the system performance and the engineering and decision-making workflow. Methodology The differences between manually generated RBCs and automatically generated OCs are evaluated by applying both options to the design and decisionmaking process of the heating supply of the ‘Stijn Streuvelstraat’ case study neighborhood. Thanks to the use of physics-based models in Modelica the method is equally applicable, and the experiences are equally valid, for other building types, supply systems, or room-side emission systems in varying climatic contexts or with different energy carriers. For the comparison, the same physics-based model in Modelica is used as the basis for the manual development of RBCs and automatically generated OCs (see Figure 1). Typical engineering procedures, as they would be carried out by engineers in practice, are applied for the OC and RBC development to guarantee a fair and realistic comparison. Case study ‘Stijn Streuvelstraat’ ’Stijn Streuvelstraat’ is a historical neighborhood in Bruges (Belgium) with 15 dwelling units for assisted living (living room, sleeping room, bathroom, all on the ground floor) and a total net heated floor area of about 1,000 m2(see Figure 2 and Figure 3). The target of the renovation is, next to the building retrofit, to design and implement a collective heat/cold supply system that is 100 % R2ES-based, future-proof (fac2We use the term ‘OC (Optimal Control)’ for design applications using a 1 year OCP, in contrast to ‘MPC (Model predictive control)’ for operational applications with a receding horizon. Proceedings of the 19th IBPSA Conference Brisbane, Australia, Aug 24-27, 2025 0002 https://doi.org/10.26868/25222708.2025.1690 Building and HVAC modeling (Modelica) Manual development rule-based control Simulation (Dymola) Energy profiles Borefield sizing Life cycle costs Automatic generation of optimal control trajectories (TACO) Energy profiles Borefield sizing Life cycle costs Man. developed RBC Auto. generated OC Figure 1: Methodology Figure 2: Building with 2 (out of 15) dwelling units Figure 3: Site plan (source: Compagnie Costume) ing the effects of climate change), cost-efficient, and that provides indoor thermal comfort. Within this study, the following centralized supply system configurations are compared (see Table 1 and Figure 4): •ref-1: air-source heat pump (ASHP)-only. •ref-2: ground-source heat pump (GSHP)-only. •c-1: hybrid concept including ASHP and GSHP. •c-2: extension of c-1 with 20 m2PVT panels for regeneration of the borefield. •c-3: variation of c-2 with 100 m2PVT. Based on experiences in a previous project, a hybrid system including air-source and ground-source heat pumps was aimed for in order to optimize annual system performance. The cases ref-1/2 are used as single-source references. Both ASHP and GSHP are modulating. Due to acoustic constraints, ASHPs Table 1: Energy supply options (ASHP: air-source heat pump, GSHP: ground-source heat pump, PC: passive cooling, PV(T): photovoltaic (thermal) Case ASHP GSHP PC PVT PV ref-1 x - - - 100 m2 ref-2 - x x - 100 m2 c-1 x x x - 100 m2 c-2 x x x 20 m280 m2 c-3 x x x 100 m2are limited to 28 kW with max. 80 % operation during night hours. A buffer tank of 0.5 m3is used in all cases between the supply and demand side. In cases with a borefield, passive cooling is possible via a heat exchanger enabling additional borefield regeneration in summer (see top in Figure 4). The available roof area (taking heritage constraints into account) that is not covered with PVT is assumed to be used for PV panels with a maximum of, in this study, 100 m2. This study does not include domestic hot water (DHW), which is provided decentrally. Floor heating/cooling was selected as room-side emission system. Temperature setpoints are defined at 21 ◦C (heating) and 26 ◦C (cooling). Internal insulation is used for the external walls (U-value 0.27 W/m2K) because the heritage character does not allow external insulation. The renovated attic floor has a U-value of 0.17 W/m2K. Each dwelling unit has a mechanical ventilation with heat recovery and the following air volume flows: (1) Living room: 75/125 m3/h supply and extract (2) Sleeping room: 50 m3/h supply (overflow) (3) Bathroom: 50 m3/h extract (overflow). Building and HVAC modeling The building and HVAC systems are physically modeled using the IDEAS Modelica library (Jorissen, Reynders, et al. 2018) (see Figure 4) with in total 61 thermal zones. Component models are linearized and/or simplified, if needed, for use in optimization. To ensure comparable results, the RBC and the OC use the same model (Figure 1). Measured hourly weather data from Leuven, Belgium, which includes both hot and cold extremes (−10 ◦C< Toutdoor < +40 ◦C), is used. The maximum available heat pump condenser capacities are limited to 28 kWth (ASHP) and 44 kWth (GSHP). The ASHP and GSHP models are calibrated based on technical manufacturer data sheets. Newly developed models are used for the borefield (Hermans 2025) and PVT collectors (Meertens 2025). Control development and testing For both the RBC and the OC case, a switch between heating mode (ASHP and/or GSHP unblocked, passive cooling blocked) and cooling mode (blocking reversed) above or below a 24 h mean outdoor air temProceedings of the 19th IBPSA Conference Brisbane, Australia, Aug 24-27, 2025 0003 https://doi.org/10.26868/25222708.2025.1690 Figure 4: Modelica model. Top: centralized heat/cold supply system, bottom: neighborhood of 15 houses. perature of 15 ◦C is implemented. Rule-based control sequences are manually developed using function blocks from the Modelica OpenBuildingControl (OBC) library (Wetter et al. 2022). A PI controller is used to modulate the heat pumps according to the tank target temperature determined by a heating curve. In cases with two heat pumps (c-1/2/3), a PI sequence controller is used to prioritize either the ASHP (mild ambient conditions) or the GSHP (cold ambient conditions). The priority switches depending on the outdoor air temperature (using a hysteresis controller with uLow = 8 ◦C and uHigh = 12 ◦C). The PVT system (c-2/3) is activated for borefield regeneration (using a hysteresis controller with uLow = 10 ◦C and uHigh = 16 ◦C borefield fluid temperature) if sufficient irradiation is available (using a hysteresis controller with uLow = 100 W/m2 and uHigh = 200 W/m2global horizontal irradiation). The floor heating valves are regulated by PI controllers for the lower and upper temperature bounds. PI controller and heating curve parameters were modified manually to achieve a balance between energy performance and thermal comfort. The model was simulated using Dymola. Optimal controls are automatically generated by TACO. TACO translates an optimization problem into a continuous Non-Linear Programming (NLP) problem and solves this NLP problem using a specifically tailored gradient-based optimizer. The userdefined objective function (see Equation 1) is formulated to minimize (1) the total electricity consumption Jel(t) including heat pumps PHP (ASHP and/or GSHP) and pumps Ppumps (distribution pump and, if present, pump for passive cooling and PVT circuit) minus the electricity generation from PV(T), (2) thermal discomfort in each dwelling unit Scomf(t), and (3) nighttime modulation of the ASHP (to reduce noise levels) Smod,ASHP(t). min o(t)Zt=1y t=0 Jel(t) + Scomf(t) + Smod,ASHP(t)dt (1a) s.t. dx(t) dt =F(x(t),z(t),o(t), t) (1b) 0 = H(x(t),z(t),o(t), t) (1c) x(t0) = x0(1d) Jel(t) = PHP(t) + Ppumps(t)−PPV(t) (1e) Scomf(t) = wcomf Nunits X n=1 scomf,hea,n +scomf,coo,n (1f) Smod,ASHP(t) = wmod,ASHP ·smod,ASHP (1g) (Tset,min(t)−Tz,n(t))2≤scomf,hea,n (1h) (Tz,n(t)−Tset,max(t))2≤scomf,coo,n (1i) (mASHP(t)−mmax,ASHP(t))2≤smod,ASHP (1j) scomf,hea,n, scomf,coo,n, smod,ASHP ≥0 (1k) Fand Hare the governing equations with the vector of state variables x(t) and the vector of algebraic variables z(t). The optimization variables o(t) are: the modulation degree mof ASHP and/or GSHP, the volume flows for passive cooling and (where available) borefield regeneration through PVT, the floor heating valve positions and the bypass ratio of the air handling unit (AHU) heat exchanger. Thermal discomfort and ASHP modulation are implemented as soft constraints with the slack variables scomf,hea,n, scomf,coo,n and smod,ASHP and the weighting factors3 wcomf = 1 ×103W/◦C2and wmod,ASHP = 1 ×104W. Perfect predictions of weather and user behavior are 3The weighting factors are chosen to have both a wellconditioned optimization problem and constraints fulfillment. Proceedings of the 19th IBPSA Conference Brisbane, Australia, Aug 24-27, 2025 0004 https://doi.org/10.26868/25222708.2025.1690 −10 0 10 20 30 40 Toutdoor [◦C] 20 22 24 26 Tindoor [◦C] RBC OC Figure 5: Operative temperature vs. outdoor temperature for a living room (2,190 samples). 0 50 [kWh/m²a] ref-1 ref-2 c-1 c-2 c-3 78.8 80.2 80.2 80.2 80.2 72.7 72.1 72.3 72.4 72.4 Heat 0 20 [kWh/m²a] 20.3 17.6 18.6 18.5 18.5 13.9 13.3 12.9 12.9 12.8 Electricity 0 5 [-] 3.9 4.5 4.3 4.3 4.3 5.2 5.4 5.6 5.6 5.6 SCOP Control RBC OC Figure 6: Heat generation, electricity consumption, and SCOP of heat pumps (sum of ASHP and GSHP where applicable, space heating only, no DHW). assumed in the 1 year OCP. A time step of 1 h was selected. Borefield sizing and LCC analysis The thermal energy profiles (obtained from either Dymola or TACO) are used for the borefield sizing using the GHEtool4taking into account short-term effects to avoid oversizing (Meertens et al. 2024). For decision-making, the energy supply options are ranked with respect to Life Cycle Costs (LCCs) over 40 years including investment, operation, maintenance, and reinvestment, and with static assumptions for the evolution of construction costs, operational costs, electricity price, inflation rate, and market interest rate. Results and analysis System performance Indoor thermal comfort Figure 5 shows the operative temperature vs. the outdoor temperature for a selected living room. The OC keeps the temperature within the given band of 21 to 26 ◦C. The RBC, on the other hand, is not capable to optimally regulate the slow floor heating/cooling system which results in longer periods below and above the setpoints. HVAC operating characteristics The comparison of annual heat generation and electricity con4https://ghetool.eu/ 0.0 0.2 0.4 0.6 0.8 1.0 Ratio operation time [-] 0 20 40 60 ˙ Qcond,HP [kW] Static heat losses Max. avail. (sim) Control RBC OC HP Sum GSHP ASHP Figure 7: Load-duration curves of heat pump condenser power for option c-1. sumption in Figure 6 highlights that the OC-operated systems perform better both on the demand side (−8 to −10 % heat generation at the condenser) and on the supply side (−24 to −32 % electricity consumption compressor). The heat savings are achieved through a better control of the floor heating system, while the electricity savings are achieved through a better operational management of the heat pumps by taking into account the Coefficient of Performance (COP) dependent on source temperatures (air and ground). An important aspect for the decision-making is the ranking of the electricity consumption. With RBC, the ASHP-only option ref-1 has the highest (20.3 kWh/m2a) and the GSHP-only option the lowest electricity consumption ref-2 (17.6 kWh/m2a). This difference is caused by the RBC activating the ASHP when needed, which is often at night (low COP). The hybrid options are in between. Contrary, with OC, ref-1 still has the highest consumption (13.9 kWh/m2a), however, the hybrid options c-1/2/3 perform best (12.8 kWh/m2a). This reflects the ability of OC to select the heat pump that leads to the highest performance (corresponding to the highest source temperature). This observation is also reflected in the comparison of the Seasonal Coefficient of Performance (SCOP) where with RBC the GSHPonly option ref-2 reaches the highest SCOP of 4.5 while with OC the hybrid options c-1/2/3 reach the highest SCOP of 5.6. The load-duration curves of the heat pump condenser power in Figure 7 illustrate for the hybrid option c-1, firstly, that the RBC-operated system uses the full available nominal heat pump (HP) capacity while the OC-operated system operates at condenser powers far below the nominal capacities. Secondly, RBC results in significantly shorter HP operation (ASHP: 1,937 h; GSHP: 2,901 h; both: 4,516 h) compared to the OC (ASHP: 4,988 h/+157 %; GSHP: 3,847 h/+32 %; both: 5,608 h/+24 %). Figure 8 shows that the average monthly tank temperatures are around 5 ◦C lower for the OC-operated system (except for August when the system is in coolProceedings of the 19th IBPSA Conference Brisbane, Australia, Aug 24-27, 2025 0005 https://doi.org/10.26868/25222708.2025.1690 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 10 20 30 40 [◦C] Control RBC OC Temperature Tank Borefield Figure 8: Monthly tank and borefield temperatures (line: mean; area: min/max) for option c-2. ref-2 c-1 c-2 c-3 0 50 100 [kEUR] Man. developed RBC ref-2 c-1 c-2 c-3 Auto. generated OC Borefield PV PVT Figure 9: Investment costs (without heat pumps) for supply options with GSHP / borefield. ing mode) which further increases the heat pump COP while a sufficiently high temperature level to ensure thermal comfort is still provided (see Figure 5). The borefield temperatures are on the same level for both control strategies. Heat pump and borefield sizing As shown in Figure 7, OC avoids using peak power to eliminate unfavorable operation modes, for example at low outdoor temperatures.5Moreover, the maximum total power (sum of ASHP and GSHP) remains well below the static heat losses. This tendency suggests that selecting even smaller HP capacities would be possible if OC (aware of the available capacity) is employed. In general, this example illustrates that the control approach chosen influences the selected heat pump sizes and as such the investment costs. This, however, raises the question how robust component sizing should be handled in OC-based designs in the future (static vs. integrated optimal control and sizing). Figure 9 depicts the investment costs for the borefield (sized with the GHEtool, Meertens et al. (2024)), PV, and PVT. The difference between ref-2 (only GSHP) 5A common approach when simulations are involved in design is to limit the available capacity to the result of a static heat loss calculation, here 44 kW (dash-dotted line in Figure 7). In the case study model, more power is available (here 72 kW) and fully used by the RBC-operated system. The detailed behavior largely depends on how aggressively the PI parameters are set. ref-1 ref-2 c-1 c-2 c-3 0 200 400 [kEUR] 4. 1. 3. 2. 5. Man. developed RBC ref-1 ref-2 c-1 c-2 c-3 4. 1. 3. 2. 5. Auto. generated OC Invest Operation Maintenance Reinvest Figure 10: Life cycle costs over 40 years and ranking. and c-1 (hybrid) illustrates the borefield reduction due to the additional ASHP.6The comparison of c-1 (PV/PVT 100/0 m2) and c-2 (PV/PVT 80/20 m2) illustrates that the borefield cost savings through a small PVT system are almost equivalent to the additional costs for PVT. Using larger PVT systems (c-3) on the other hand, results in the highest overall investment costs. These observations and corresponding rankings are independent on whether RBC or OC is used, however, the better performing OC (see Figure 6) results in smaller borefield sizes. The cost of PVT may come down in the future as the technology matures, making larger PVT systems more attractive, but this example illustrates the need for careful sizing when combined with borefields that provide passive cooling as well. Life cycle costs Figure 10 compares the life cycle costs (present value of future expenses), including investment, operation, maintenance, and reinvestment for all 5 options. Due to the better energy performance, all options with OC have significantly lower life cycle costs. However, the ranking (1. to 5.) of the different option is the same for RBC and OC. This means that, formally, using an RBC or an OC does not influence the LCC-based decision making, however, differences in the individual cost components have to be taken into account. The GSHP-only case ref-2 performs best for both control approaches, however, the difference is mainly rooted in lower reinvestment and maintenance costs, which are significantly higher for cases with ASHPs and/or PVT. If only investment and energy costs were considered, the ASHP-only or GSHP-only options ref-1/2 would perform best and on the same level with RBC, while the ASHP-only option ref-1 would perform best with OC. ref-1 has the lowest investment costs, however, an ASHP-only option was not possible due to noise restrictions. In general, the comparison illustrates that through the significant reduction of energy costs with OC the share and the importance of maintenance and 6An extra investment for an ASHP is of course needed (not included in Figure 9). Proceedings of the 19th IBPSA Conference Brisbane, Australia, Aug 24-27, 2025 0006 https://doi.org/10.26868/25222708.2025.1690 reinvestment costs increase. Design choice A hybrid concept based on c-2 was finally chosen for the renovation of ‘Stijn Streuvelstraat’. The integration of an ASHP, a borefieldcoupled GSHP, and a ‘small’ PVT system (remaining roof area used for PV) is 100 % renewable-based, cost-efficient, future-proof against effects of climate change through passive cooling and borefield regeneration (to balance varying heat demand), and it may provide flexibility options through DSM. OC enables optimal operation of each individual technology in the integrated system which emphasizes the potential of OC as system integrator. Engineering and decision-making workflow The experiences from the design process reveal the following benefits of automatically generated OC: (1) The automatic generation of OCs drastically reduces the effort at the design stage. In the case of RBCs, the selection of function blocks, the connection of inputs and outputs, and particular the manual definition of parameters and the variation in simulation runs are work-intensive tasks. While in practice many RBCs use default parameters (PI controllers, heating curve etc.), they were adapted to the considered system in this study. Thanks to the automatic OC generation with TACO the effort is much lower and limited to the definition of the objective function and constraints (see Equation 1). (2) The automatic generation of OCs enables the time-efficient comparison of a large number of supply options. In the case of RBCs, each case with a different supply system generally requires a different custom controller.7However, reusing and extending parts of control sequences is a strategy to limit the effort. In the case study, the block controlling the heat pumps is reused. However, as only the best performing option will be implemented, the effort for the development of the remaining options (4 out of 5 in the case study example) is wasted. Since the OCs are generated automatically and the effort for the definition of the objective function is very limited, TACO enables the time-efficient comparison of a larger number of supply options. (3) Automatically generated OCs enable a more fair system comparison. The performance of different manually developed and configured RBCs is biased by the knowledge of engineers and the effort invested in the RBC development. In the case study for example, the HP performance is heavily depending on the chosen heating curve parameters: higher supply temperatures would lead to significantly increased HP electricity consumption and would penalize particularly ASHP operation, while lower supply temperatures might result in thermal discomfort. Another 7ASHRAE Guideline 36-2021 addresses this issue by providing pre-configured control macros, however, these macros particularly address HVAC systems common in the U.S.. example is the HP performance in the ASHP-only option ref-1 that could be enhanced by forcing the ASHP to operate more during the day, for example through different day/night setpoints. Finding suitable RBC parameters, however, requires system knowledge (taking the time constants related to floor heating and building inertia into account) and iterative testing, which is, if not automatized, highly workintensive. Contrary, the OC trajectories are all generated by the same algorithm in TACO which enables a more neutral and fair assessment of each supply option. (4) The application of OC at the design stage requires a change in mindset and expertise. Compared to the current building practice (no model-based design), applying model-based design either with RBC or OC requires deep collaboration and integrated codesign of building physics, HVAC systems, and controls. However, for the application of OC, the required expertise shifts from classical control engineering (RBC) to more optimization related topics, including the definition of the objective function, constraints, weighting factors, the tolerance, time steps, or the prediction horizon which has to be done with care to avoid unexpected behavior. Moreover, analyzing externally generated OC (here through TACO) requires understanding of how the optimizer translates the own inputs into an optimal control regime. (5) Simulation time. A crucial aspect for an efficient workflow is the simulation time. In the case study example, a 1 year simulation with RBC takes between 24 and 31 h which is too long for practitioners.8The simulation time for the TACO runs with a time step of 1 h is much shorter, typically between 4 and 7 h. Conclusion and outlook The key findings of this study are: (1) OC increases the energy and comfort performance compared to RBC and the automatic generation of OCs increases the design process efficiency compared to manually developed RBC. (2) The overall ranking of different heating supply options does, in this case, not differ if manually developed RBCs or automatically generated OCs are used. Differences do, however, occur when only investment and energy costs are compared without reinvestment and maintenance costs. (3) OC enables optimized sizing: the maximum HP condenser power with OC remains far below the maximum available capacity and even below the results of a static heat loss calculation. Lower energy use with OC enables smaller borefield sizes. Using the ‘small’ PVT system for borefield regeneration enables further borefield size reduction while the extra costs for PVT balance out the savings by the smaller borefield. The fast and efficient creation of reliable building and 8’Dassl’ solver, tolerance of 1e−6. See Sahlin et al. (2019) for a discussion of the simulation performance of whole building simulations in Modelica. Proceedings of the 19th IBPSA Conference Brisbane, Australia, Aug 24-27, 2025 0007 https://doi.org/10.26868/25222708.2025.1690 HVAC models is a core prerequisite for the successful application of model-based design (either using RBC or OC). While libraries for simulations with RBC are widely available (see Modelica IBPSA library9), corresponding libraries for OC are needed as well. Future work includes performance comparisons of 1 year OCPs (design method, this study) to Model predictive control (MPC) with receding prediction horizon (real operation), analyses of the potential of a split borefield as an extra degree of freedom, methods for the integrated optimal control and sizing, and approaches for the continued use of the design models for fault detection in operation. Acknowledgement The authors are grateful to Filip Jorissen (Builtwins) for all TACO developments, and Sweco Belgium and and Mintus for their contribution in the co-design. The authors acknowledge the funding by the European Union through the SEEDS project under the Horizon Europe Programme (grant Agreement: 101138211). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. References Drgoˇna, J´an et al. (2020). “All You Need to Know about Model Predictive Control for Buildings”. In: Annual Reviews in Control 50, pp. 190–232. issn: 13675788. doi:10.1016/j.arcontrol.2020.09. 001. Hermans, Louis (2025). “Development of an Integrated Optimal Control and Sizing Methodology for a Small Hybrid District Heating System in Belgium”. In: Building Simulation 2025. Brisbane, Australia. Jorissen, Filip, Wim Boydens, and Lieve Helsen (Mar. 4, 2019). “TACO, an Automated Toolchain for Model Predictive Control of Building Systems: Implementation and Verification”. In: Journal of Building Performance Simulation 12.2, pp. 180– 192. issn: 1940-1493, 1940-1507. doi:10 . 1080 / 19401493.2018.1498537. Jorissen, Filip, Glenn Reynders, et al. (Nov. 2, 2018). “Implementation and Verification of the IDEAS Building Energy Simulation Library”. In: Journal of Building Performance Simulation 11.6, pp. 669– 688. issn: 1940-1493. doi:10 . 1080 / 19401493 . 2018.1428361. 9https://github.com/ibpsa/modelica-ibpsa Kramer, R. P., A.W.M. van Schijndel, and H. L. Schellen (2017). “The Importance of Integrally Simulating the Building, HVAC and Control Systems, and Occupants’ Impact for Energy Predictions of Buildings Including Temperature and Humidity Control: Validated Case Study Museum Hermitage Amsterdam”. In: Journal of Building Performance Simulation 10.3, pp. 272–293. issn: 1940-1493. doi:10 . 1080 / 19401493 . 2016 . 1221996. Meertens, L., W. Peere, and L. Helsen (May 28, 2024). “Influence of Short-Term Dynamic Effects on Geothermal Borefield Size”. In: 2024 Research Conference Proceedings of the IGSHPA. International Ground Source Heat Pump Association. International Ground Source Heat Pump Association. doi:10.22488/okstate.24.000004. HDL: 20.500.14446/344451. Meertens, Lone (2025). “Modeling and Validation of a Photovoltaic Thermal System Modelica Model Based on Experimental Measurement Data”. In: Building Simulation 2025. Brisbane, Australia. Ramesh, Rakesh et al. (2024). “Optimising Energy Flexibility in Finnish Residential Buildings: A Comparative Study of Pi, Rule-Based and Model Predictive Control Strategies”. In: doi:10.2139/ ssrn.5002205. Sahlin, Per et al. (Sept. 2–4, 2019). “On the Scalability of Equation-Based Building and District Simulation Models”. In: Proceedings of Building Simulation 2019: 16th Conference of IBPSA. Ed. by Vincenzo Corrado et al. Building Simulation Conference Proceedings. IBPSA, pp. 2584–2590. doi: 10.26868/25222708.2019.210130. Saloux, Etienne et al. (Sept. 2025). “From Theory to Practice: A Critical Review of Model Predictive Control Field Implementations in the Built Environment”. In: Applied Energy 393, p. 126091. issn: 03062619. doi:10.1016/j.apenergy.2025. 126091. Torabi, Narges et al. (Jan. 1, 2022). “Common Human Errors in Design, Installation, and Operation of VAV AHU Control Systems – A Review and a Practitioner Interview”. In: Building and Environment 221, p. 109333. issn: 03601323. doi:10. 1016/j.buildenv.2022.109333. Walther, Karl (Aug. 2025). “Model-Based Design, Digital Delivery, and Implementation of HVAC Controls — Lessons Learned from a Building-Scale Application for AHUs and PLCs”. In: Building and Environment 282, p. 113126. issn: 03601323. doi: 10.1016/j.buildenv.2025.113126. Wetter, Michael et al. (2022). “OpenBuildingControl: Digitizing the Control Delivery from Building Energy Modeling to Specification, Implementation and Formal Verification”. In: Energy 238, p. 121501. issn: 03605442. doi:10 . 1016 / j . energy.2021.121501. Proceedings of the 19th IBPSA Conference Brisbane, Australia, Aug 24-27, 2025 0008 https://doi.org/10.26868/25222708.2025.1690