An Enhanced Adaptivity of Reinforcement Learning-Based Temperature Control in Buildings Using Generalized Training

In this paper, we present an adaptive Reinforcement Learning (RL) agent training approach which aims to provide a temperature control adaptable to various types of buildings. The main purpose of the proposed method is to avoid repeating the training of RL agents on every new building and therefore skip the modeling part and ease the spread of RL-based controllers. This study includes analysis of the proposed method working along with ACKTR, a state of the art RL algorithm, regarding parameter tuning, algorithm convergence, consumption reduction and state observation accuracy. The RL based controller is applied first to single rooms and then extended to entire buildings, to demonstrate its generalization efficiency. To measure the performance of the RL controller, comparisons with MPC and ON/OFF based controllers are performed. On each test, the adaptive RL temperature control was similar to MPC and better than ON/OFF control, while being able to reach up to 5% of energy savings compared to MPC and 10% compared to ON/OFF control.