A Data-Driven Parameter-Topology Hybrid Optimization of Permanent Magnet Motors Based on Deep Reinforcement Learning

Parameter-topology hybrid optimization is one of the effective approaches to simultaneous optimizations of the shapes of magnets and magnetic cores of interior permanent magnet synchronous motor (IPMSM). However, because the magnet and topology parameters are related to each other, their simultaneous optimizations result in difficult optimization problem. This paper presents a data-driven parameter-topology hybrid optimization for IPMSM based on deep reinforcement learning. The design process of IPMSM is newly formulated as a stochastic process, in which the magnet and topology parameters are assumed to be generated probabilistically in order, depending on design requirement. The present method estimates these probability density functions through a modified actor-critic-based reinforcement learning by utilizing rotor geometry data, which are generated using evolutionary algorithms. Numerical results show that the present reinforcement learning can obtain rotor shapes depending on the requirement through optimizations and learning. It is also shown that the trained actor model can be used as a generative model for the rotor shape, that is, the trained actor model can generate various rotor shapes depending on input requirement, through which the actor model can also be refined.