Optimal control for An Active Phase Change Material System Using Reinforcement Learning

Optimal control of buildings equipped with active phase change material (PCM) systems is challenging due to the rigorous models involved. Besides the nonlinear models, binary variables appear in the optimal control formulation in this work. Deep Q-learning network (DQN) reinforcement learning was adopted to address the complexity of the optimal control formulation. DQN includes deep neural networks which can map binary actions to continuous states. The DQN was trained and tested through simulation using a validated model of the whole system. Using the electricity price, hut dynamics and ambient conditions information as states, the proposed new control strategy can achieve an optimal balance between thermal comfort and energy cost saving.