A Simulation Study on Energy Optimization in Building Control with Reinforcement Learning

We propose and evaluate a deep reinforcement learning control paradigm for building energy systems. In comparison to other advanced control techniques, namely Model Predictive Control, the reinforcement learning paradigm avoids the costs and uncertainties associated with the requirement for a control-oriented model. We apply a mixed agent for the Proximal Policy Optimization algorithm, similar to the algorithm proposed in [7] as well as a non-discounted finite horizon optimization scheme. We investigate the capabilities of the proposed reinforcement learning controller regarding energy efficiency, comparing it against the most widely used rule-based control paradigm as a baseline controller. We verify our proposed paradigm in a simulation study with building models implemented in Dymola.