Deep reinforcement learning for the control of a desiccant wheel-based air-conditioning system

Moisture removal is an important concern in many industries, e.g. producing food and pharmacy, and manufacturing lithium-ion batteries. However, humidity control has been paid less attention in research compared to temperature control. In this paper, the control of a desiccant wheel-based air conditioning (AC) system under the influence of changing climate is investigated. A one-dimensional gas-side resistance model is employed to mimic a silica gel desiccant wheel, and one of the state-of-the-art policy gradient algorithms, proximal policy optimization (PPO), is utilized to train a feedback controller for this AC system. Compared with model predictive control, the advantage of this strategy is that the policy network obtained by PPO can be used without involving online optimization. Numerical simulations based on real environmental data show that the controller is effective to keep indoor humidity and temperature to their set points. Compared with deep deterministic policy gradient algorithm, the obtained controller has higher reward and less regulation variance.