Paper Conference

Proceedings of Building Simulation 2021: 17th Conference of IBPSA


Comparison of data-driven model-based and model-free control approaches for unlocking building energy flexibility

Anjukan Kathirgamanathan 1,3, Eleni Mangina 2,3, Donal P. Finn 1,3
1 School of Mechanical and Materials Engineering, University College Dublin, Ireland
2 School of Computer Science, University College Dublin, Ireland
3 Energy Institute, University College Dublin, Ireland

Abstract: Commercial buildings are significant end-use energy consumers and given their inherent thermal mass and use of advanced control infrastructure together with heating, ventilation and air conditioning systems, have the potential to offer significant load shifting opportunities. Data-driven control frameworks show promising results as a robust and scalable technique for the heterogeneous building stock that will enable automated demand response whilst ensuring occupant thermal comfort is maintained. This research compares two different approaches to the problem. First, a model based approach is considered and is denoted Data Predictive Control with Ensemble methods (DPC-En). This approach is based on the use of the random forest predictor with the ‘separation of variables' technique to deliver an optimal control problem. Second, a model free approach is considered, which utilises a soft actor critic deep reinforcement learning (SAC DRL) algorithm. The DPCEn technique was able to minimise energy costs by 14.0% compared to the baseline rule-based control. The SAC DRL similarly achieved 10.7% savings compared to the baseline. Both techniques are able to respect occupant thermal comfort constraints.
Keywords: building energy flexibility, data-driven, reinforcement learning
Pages: 503 - 510