Energy Management of a Residential Heating System Through Deep Reinforcement Learning

In this study, a controller based on deep reinforcement learning was tested for a residential building equipped with a radiant heating system. In detail, a Soft Actor-Critic (SAC) algorithm was implemented to optimize the operation of the heating system while ensuring adequate levels of indoor temperature. A probabilistic window opening behavior model was implemented within the simulation framework in order to emulate the interaction of the occupants with the building. A sensitivity analysis on SAC hyperparameters was carried out to determine the best configuration that was then deployed in four different scenarios in order to analyze the adaptability of the controller to different boundary conditions. The performance of the reinforcement learning agent was evaluated against a baseline strategy which combines rule-based and climatic control. The developed agent was able to achieve a saving of heating energy provided to the building in the range between 2 and 6% while increasing temperature control performance up to 65% in the four scenarios investigated.