Generative Physics Informed Machine Learning Method for DC-Link Capacitance Estimation

Abstract

This article proposes an innovative generative physics-informed machine learning (GPIML) method for the estimation of dc-link capacitance during the precharging process of the vehicular power systems, which contributes to greatly enhancing the reliability of electrified transportation. Different from the other machine learning-based estimation approaches, the proposed method produces highly accurate results using small input experimental dataset. To enable sufficient neural network training, diffusion algorithm is first adopted in the proposed method to augment the training data based on small input dataset. Then, the augmented data is fed to a physics-informed long short-term memory (PILSTM) algorithm to estimate the dc-link capacitance. Superior accuracy and strong robustness to measurement noises are achieved. The effectiveness of the proposed method is validated through experimental studies.