Training neural-network-based controller on distributed machine learning platform for power electronics systems

A new training scheme for neural-network-based controller for power electronics systems is proposed. It utilizes the circuit model of the power conversion stage (PCS) in the training process. The training algorithm is a distributed form of evolutionary computation, being able to run on a computer cluster equipped with multiple graphics processing units (GPUs). As a design example, a boost converter has been built and evaluated to exemplify the performance of the neural-network controller trained by the proposed scheme. The experimental results showed that the neural-network controller excels in speed of response and rejection of disturbance once trained properly.