Y. Jeong
University of Rhode Island, Rhode Island, United States
Keywords: Power Electronics, Power converter, deep reinforcement leaning, electric design automatoin
The design process is instrumental in defining power conversion systems (PCSs) performance. For a long time, the industry has depended on human experts and ingrained internal standards for PCS design methods that are often labor-intensive, time-intensive, and might not yield the best outcomes. Yet, there’s a wind of change: the increasing trend towards automated design methods, particularly those based on machine learning (ML), demonstrates immense potential for identifying optimal design parameters for power converter systems (PCSs) with greater precision and efficiency. Currently, the grand challenge of automated PCS design is the long computing time. The slow time-domain circuit simulation is the bottleneck. The sequential nature of time-domain circuit simulation, required for PCSs, precludes the possibility of parallelizing operations across different timesteps on multi-core CPUs or GPUs. Studies have indicated that a staggering 99% of training time goes into circuit simulation when optimizing a circuit. This research explores the application of deep reinforcement learning (RL) to improve the design of power conversion systems. It demonstrates that RL models can find globally optimal designs more efficiently, and using FPGA-based accelerators can further enhance simulation speeds by 60x, making the RL-driven design process more practical for real-world use.