Safe reinforcement learning for smart building’s energy optimization with vehicle-to-everything (V2X) technology

This research explores Safe Reinforcement Learning (Safe RL) for optimizing energy management in smart buildings integrated with Photovoltaics (PV) systems and electric vehicle (EV) charging infrastructure. Traditional RL methods often lead to unsafe actions that violate the operational constraints, reducing the efficiency and safety of the energy systems. This study evaluates three Safe RL methods, penalty-based RL, action replacement, and action masking, to determine the effectiveness in improving self-consumption, PV utilization, and energy cost reduction.

The results show that the Maskable Proximal Policy Optimization (Maskable PPO) achieves the highest energy cost reduction while maintaining the self-consumption and PV utilization rates. These findings provide valuable insights into Safe RL-based energy management strategies, contributing to the advancement of smart energy systems.