An Hybrid Model-Free Reinforcement Learning Approach for HVAC Control

Traditional Heating Ventilation and Air Conditioning (HVAC) systems are extremely energy draining appliances, and their use is ever increasing with urbanisation. For this reason, strong research effort has been put in the development of novel control strategies for the optimal management of HVAC systems, aiming at reducing energy consumption without affecting thermal comfort. In this paper, we propose an hybrid model-free Reinforcement Learning approach for HVAC control able to optimise both energy consumption or users comfort. Our methodology is compared with two baseline solutions in literature based on an EnergyPlus controller and a Model Predictive Control. Results show that our methodology can outperform both baselines in terms of energy consumption reduction or thermal comfort optimisation, given that either of the two objectives is appropriately chosen during the training and the hyperparameters selection phase.