Conference Paper
Proceedings of BSA Conference 2024: Sixth Conference of IBPSA-Italy
Recommendations to Make Reinforcement Learning Practical in Building Control Applications
Recommendations to Make Reinforcement Learning Practical in Building Control Applications
Sourav Dey, Gregor HenzeDOI: https://doi.org/10.13124/9788860462022_44Abstract: The paper provides an analysis of the application of reinforcement learning (RL) in the domain of building controls, summarizing four years of research. The primary focus is exploring RL's potential to adaptively learn from building data, bypassing the need for individualized extensive building modeling efforts and enabling the transfer and adaption of trained agents to similar building environments. Despite its promising prospects, RL faces challenges such as extended training durations, instability during early exploration phases, and issues in interpreting the actions of trained agents. The research was focused on two core areas. The first area investigates strategies to enhance RL agents' learning efficiency and stability in building control contexts with approaches such as imitation learning, inverse RL, and online learning with guided exploration with surrogate models utilizing rule-based controls, showing significant improvements in the training process. The second area addresses the critical aspects of scalability and interpretability of RL agents. It examines the feasibility of transferring trained agents to various buildings, potentially with new objectives, highlighting RL's adaptability and practical applicability in real-world building control scenarios. In summary, this paper consolidates critical findings from the research and offers actionable insights and recommendations for practical deployment and training RL in building energy management systems without extensive building modeling efforts. It emphasizes the transformative potential of RL in this field and suggests avenues for future exploration and development. Pages: 353 - 360 Paper: