Entropy-driven deep reinforcement learning for HVAC system optimization
Chen Zhang, Zhi Tan
Abstract
The study explores the use of deep reinforcement learning (DRL) for optimizing heating, ventilation, and air conditioning systems to reduce energy consumption while maintaining indoor comfort. Traditional control methods, such as Model Predictive Control and Proportional-Integral-Derivative control, often struggle with variable environmental conditions, leading to inefficient energy use and discomfort. This research introduces a DRL-based approach, improving the algorithm optimization performance by entropy screening high-value information and reducing value errors. Using the Energym building simulation platform, the performance of various reinforcement learning algorithms, including Deep Deterministic Policy Gradient (DDPG) and Soft Actor-Critic (SAC), was compared in different weather conditions. We demonstrate a novel DRL algorithm, EntropySAC, that demonstrated significant energy savings and improved comfort control. Notably, the optimized algorithm reduced energy consumption by up to 38.95% under high-temperature conditions and consistently outperformed traditional approaches.