A Nonisolated Single-Inductor Multiport DC-DC Topology Deduction Method Based on Reinforcement Learning

Single-inductor multiport converters have great application potential due to their more silicon less magnetic feature. However, the recent topology design methods, such as the forward design and reverse design, require strong power electronics knowledge, which troubles the application engineers lacking power electronics background. In this article, a topology deduction method based on reinforcement learning (RL) is proposed to gain the benefits of both forward and reverse design. The proposed method uses a neural network (NN) for forward design and takes a set of simple rules to give rewards for reverse design. With RL, the NN sums up experience during trial-and-error without human intervention, and finally finds the topology that gains the highest reward. Benefits of this method include: 1) it only requires design specifications such as components and port voltage as inputs and some simple rules as rewards, avoiding complicated feature predesign; 2) it allows the deduction to be started from any preconnection, satisfying the prior constraints of application engineers; and 3) it can recommend several actions at each step, providing good diversity of deduction results. Using this method, many topologies in the literature works are deduced, and some new topologies are also found. One of the topologies is experimentally tested and the results show the validity.

Automated summary

In this article, a topology deduction method based on reinforcement learning (RL) is proposed to design single-inductor multiport converters. The method utilizes a neural network for forward design and simple rules for reverse design, resulting in simplified design process, meeting prior constraints, and providing diverse deduction results.