Deep reinforcement learning-based position optimization model for intermediate discharge valves in a scroll compressor

Abstract

Heating and cooling cycles can be controlled by changing the suction and discharge pressures of a scroll compressor. Over-pressurization caused by a low compression ratio induces work loss and may damage the structure of the scroll. This over-pressurization can be prevented by discharging compressed refrigerant through intermediate discharge valves (IDVs) before the refrigerant reaches the central discharge port. Determining the optimal positions of IDVs to provide appropriate compression is time-consuming due to the nonlinearity of the design objective. This research focuses on optimizing the positions of the IDVs using deep reinforcement learning (DRL), allowing the compression work to be minimized for any given scroll design, number of IDVs, and compression ratio. A low-order numerical model computes thermodynamic states inside chambers and resulting compression work, and serves as the environment for the DRL agent. Using a proximal policy optimization algorithm, the agent learns a policy that incrementally places the IDVs without geometric overlap. Under the optimal IDV positions, the temporal profile of the chamber pressure is similar to that of the ideal pressure without over-pressurization. Furthermore, by varying the number of IDVs and implementing a cost function for the installation of IDVs, the minimum number of IDVs required to suppress over-pressurization efficiently is determined. The DRL-based optimization approach suggested in this study significantly reduces the time cost involved in the design of the scroll compressor and IDVs.