Enhanced Strategies for Off-Policy Reinforcement Learning Algorithms in HVAC Control

This paper investigates policy optimization methods for HVAC control in complex dynamic environments, proposing a series of enhancement strategies for off-policy reinforcement learning algorithms to tackle this challenge. Initially, by integrating an improved principal component analysis, redundant and similar features within the state vector are analyzed, and feature extraction is conducted to reduce the dimensionality of the state space. Moreover, accounting for the influence of time-varying factors, the corrected data is upgraded to better reflect real-world conditions. Additionally, to enhance the sampling efficiency of off-policy learning, a fusion of HVAC control characteristics with the experience replay pool is employed, integrating fragmentary trajectories into the learning process. In summary, our methods offer effective policy optimization solutions for HVAC control in complex dynamic environments.