Precise Coil Inductance Prediction with Transfer Learning Improved Deep Neural Networks in WPT Systems

Due to the intricate electromagnetic characteristics of wireless power transfer (WPT) systems, conventional inductance modelling methods for coils in WPT systems face the challenge of low accuracy, particularly when the ferrite plates are adopted. This paper introduces a transfer learning improved feedforward neural network (FNN) to precisely predict the inductances of rectangular coils with elliptic corners in WPT systems. First, a parametric structure model is introduced for characterizing the layouts of rectangular coils. Based on this structure model, a FNN model is designed to predict the inductances of different coils under varied misalignments. Moreover, the proposed FNN model can be continuously refined with transfer learning via ongoing input data, thus significantly improving its generalization. The prediction accuracies of the FNN model are validated with three prototype coils under varied misalignments. The experimental results demonstrate that the proposed FNN model can achieve high prediction accuracies with peak mean prediction errors of only 5.52% and 5.42% in self-and mutual inductance prediction.