Building demand response control through constrained reinforcement learning with linear policies

Recent advancements in model-free control strategies, particularly reinforcement learning (RL), have enabled more practical and scalable solutions for controlling building energy systems. These strategies rely solely on data, eliminating the need for complex models of building dynamics during control decision making, the development of which is expensive involving significant engineering efforts. Conventional unconstrained RL controllers typically manage indoor comfort by incorporating a penalty for comfort violations into the reward function. This penalty function approach leads to control performance very sensitive to the penalty factor setting. A low comfort penalty factor can result in significant violations of comfort constraints while a high penalty factor tends to degrade economic performance. To address this issue, the present study presents a constrained RL-based control strategy for building demand response that explicitly learns a constraint value function from operation data. This study considers both linear mapping and deep neural networks for value and policy function approximation to evaluate their training stability and control performance in terms of economic return and constraint satisfaction. Simulation tests of the proposed strategy, as well as baseline model predictive controllers (MPC) and unconstrained RL strategies, demonstrate that the constrained RL approach could achieve utility cost savings of up to 16.1 %, comparable to those achieved with MPC baselines, while minimizing constraint violations. In contrast, the unconstrained RL controllers either lead to high utility costs or significant constraint violations, depending on the penalty factor settings. The constrained RL strategy with linear policy and value functions shows more stable training and offers 4 % additional cost savings with reduced constraint violations compared to constrained RL controllers with neural networks.