An Online Reinforcement Learning Method for Multi-Zone Ventilation Control With Pre-Training

Abstract

This article proposes an online reinforcement learning with pretraining (ORLPT) method for multizone ventilation control. The proposed ORLPT method contains two stages, which are pretraining and online reinforcement learning (RL). In the pretraining stage, the proposed ORLPT method is trained on a simulation model of the ventilation system, and a dynamic-target RL algorithm is proposed to improve the generality of the learned policy, which can achieve accurate multi-zone ventilation control. In the online RL stage, the neural network settings and parameters obtained from the pretraining stage are employed as the initial parameters for online RL, so that the algorithm starts with a good estimate and can converge speedily. In the lab, a real 5-terminal ventilation duct system is fabricated to validate the proposed ORLPT method. The experimental results demonstrate that the proposed ORLPT method can achieve accurate ventilation control, and the pretraining scheme can significantly reduce the time cost. Besides, compared with the traditional ventilation control method, the proposed ORLPT also shows significantly better time efficiency.