Deep reinforcement learning to optimise indoor temperature control and heating energy consumption in buildings

In this work, Deep Reinforcement Learning (DRL) is implemented to control the supply water temperature setpoint to terminal units of a heating system. The experiment was carried out for an office building in an integrated simulation environment. A sensitivity analysis is carried out on relevant hyperparameters to identify their optimal configuration. Moreover, two sets of input variables were considered for assessing their impact on the adaptability capabilities of the DRL controller. In this context a static and dynamic deployment of the DRL controller is performed. The trained control agent is tested for four different scenarios to determine its adaptability to the variation of forcing variables such as weather conditions, occupant presence patterns and different indoor temperature setpoint requirements. The performance of the agent is evaluated against a reference controller that implements a combination of rule-based and climatic-based logics. As a result, when the set of variables are adequately selected a heating energy saving ranging between 5 and 12% is obtained with an enhanced indoor temperature control with both static and dynamic deployment. Eventually the study proves that if the set of input variables are not carefully selected a dynamic deployment is strictly required for obtaining good performance.