Where and when: Thursday, Aug 12 at 2-3pm in 303S-561
Speaker: Tech Logg Ding, Ph.D. student, Department of Mechanical Engineering
Abstract: Refrigeration systems are essential for various cooling applications in modern society. The majority of the refrigeration systems in the world uses the Vapour Compression Refrigeration Cycle (VCRC) technology due to its’ efficiency. In the literature, VCRC control optimisation is proven to be an effective method to increase their reliability and energy efficiency. However, it is often very complicated to implement advanced optimal and predictive control strategies in the industry because of the high nonlinearities, parametric coupling, and time-varying properties of the VCRC. Therefore, my research objective is to explore the application of reinforcement learning to learn an optimal VCRC control strategy. If successful, this opens up a pathway for intelligent VCRC controllers to effectively work with the properties mentioned earlier. Besides that, it could also open up pathways for automated VCRC control design, updates, and more intuitive control strategies. The research currently uses a simulated VCRC model with two controllable actuators, the compressor and the valve, to train the RL agent. The controller’s objective was to track a desired cooling temperature and a superheat temperature using the actuators. Meaningful VCRC states, operating conditions, and previous actions were selected as the inputs to the agent. Next, a continuous reward function supplemented with low error bonuses was also developed based on the control objectives. Finally, using these components, a Twin-Delayed Deep Deterministic (TD3) agent and a Soft Actor-Critic (SAC) agent were trained to learn the optimal VCRC control strategy. The controller performance analysis showed that both agents produced great control performance with low tracking error, high response speed, and stable controller actions. The SAC agent was slightly better than the TD3 in terms of controller performance, policy convergence rate, and ease of implementation. Further discussion on the controllers’ design process and performance will be discussed in the presentation.