Machine Learning Methods for Feedforward Power Flow Control of Multi-Active-Bridge Converters

Controlling the multiway power flow in a multi-active-bridge (MAB) converter is important for achieving high performance and sophisticated functions. Traditional feedforward methods for MAB converter control rely on precise lumped circuit models. This article presents a machine learning (ML) method for the feedforward power flow control of an MAB converter without a precise circuit model. A feedforward neural network was developed to capture the nonlinear characteristics and predict the phases needed to achieve the targeted power flow. The neural network was trained with a large amount of data, collected with a set of known phase angles. This trained network was used to predict the phases to achieve the targeted power flow. A six-port MAB converter was built and tested to validate the methodology and demonstrate the 'machine-learning-in-the-loop' implementation. Transfer learning was proven to be effective in reducing the size of the training data needed to obtain an accurate ML model. ML-based feedforward power flow control can achieve comparable accuracy as traditional model-based methods and can function without a precise lumped circuit element model of the MAB converter.