Benchmarking Building Control Systems Thanks to Calibrated Models

Buildings are among the biggest energy consumers and greenhouse gas (GHG) emitters worldwide. Smart control strategies promise an efficient use of e.g., heating and cooling, to reduce both energy consumption and GHG emission.

For controlling Heating Ventilation and Air Conditioning (HVAC) systems in buildings, two strategies are the current standard: rule-based control, where decisions on how to run the heating and cooling equipment are done with a set of fixed rules, and PI control, a classical control strategy where parameters must be tuned by experts for each building separately. These methods generally provide reliable performance with res
pect to temperature requirements but perform sub-optimally regarding energy consumption.

Various studies have shown the potential of more advanced control strategies like Model Predictive Control (MPC), learning-based control, and Reinforcement Learning (RL), but a direct comparison of approaches from those categories remain mostly open.

Energym’s Contribution

This is where Energym comes into play: including 11 simulation models (three Modelica-based and eight EnergyPlus-based models), this Python library offers a framework for comparing controllers in different evaluation scenarios. Table 1 provides an overview of the models’ installed technical equipment and their controllability, as well as the key performance indicator they track.

Table 1. Equipment of the different models in Energym. Th: Thermostat, HP: Heat Pump, Bat: Battery, AHU: Air Handling Unit, EV: Electric Vehicle, PV: Photovoltaic, ✓: present and controllable, #: present but not controllable, ✕: absent. Source: CSEM 

6 Reasons for Using Energym

Usability: The usage of Energym resembles the one of the popular RL benchmarking library Gym and is compatible with different types of controllers like rule-based controllers, PI controllers, MPC, and RL-based controllers. Generating a simulation environment and interacting with it only takes a few lines of code.

Control Specific Features: The environments can be augmented with wrappers, to make them more suitable for the control task at hand. This includes scaling the input and output variables, changing the interaction frequency, or restructuring the outputs to be compliant with the standard set by Gym. Additionally, different weather files can be used to train controllers, and stochastic forecasts are available.

Standardized Evaluation: All stochasticity is removed for the evaluation scenarios, such that comparability between runs can be ensured. This includes a fixed evaluation period and fixed weather. The final performance metric is given by prespecified Key Performance Indicators (KPIs).

Diverse Scenarios: The included models differ in size, number of rooms, usage profile, technical equipment, controllability and climate zone. Furthermore, each model has a predefined control goal, from minimizing the power demand, over minimizing the grid exchange, to minimizing the CO₂ emission, as shown in the right column of Table 1.

Model Calibration: The majority of the models has been calibrated with data from real test-sites. The fitting was done in a two-step process, firstly modelling the building envelope, and secondly adding the technical equipment. By relying on recommendations from the American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) and the International Performance Measurement and Verification Protocol (IPMVP) for model acceptance, the models provide a realistic image of the real systems.

Open Source: Energym is released as open source, which makes the developed code fully transparent. Furthermore, an extension by other researchers is encouraged, increasing the diversity of included models.

Conclusion

Energym serves as a tool to test and benchmark controllers on realistic building models, and provides a broad range of evaluation scenarios. More information on the library can be found in the documentation and the paper.

Acknowledgments

Produced from materials originally authored by P. Scharnhorst (CSEM and EPFL, Switzerland), B. Schubnel (CSEM), C.F. Bandera (University of Navarra), J. Salom (IREC), P. Taddeo (IREC), M. Boegli (CSEM), T. Gorecki (CSEM), Y. Stauffer (CSEM), Antonis Peppas (NTUA), and Chrysa Politi (NTUA).

This information has been sourced, reviewed and adapted from materials provided by CSEM.

For more information on this source, please visit CSEM.