Machine-Learning Based Optimization for Double-Layer Planar Spiral Coil Design in High-Frequency Wireless Power Transfer Systems

Abstract

In Wireless Power Transfer (WPT) systems, the performance with high-frequency in the MHz range depends on several major parts of the system, for example, switching losses of the inverter, the quality factor (Q) of the transmitter and receiver coil for WPT system. Minimized switching losses offer higher efficiency for the inverter. Furthermore, the higher Q of the coil implies higher efficiency. Therefore, it is necessary to improve their performance to achieve a higher efficiency of the overall WPT system, which can be applied in electric vehicles (EVs), unmanned aerial vehicles (UAVs), and drones. This paper presents the optimization of self-resonant coil design in a two-layer planar spiral structure using the machine learning method in the high-frequency WPT system for small and lightweight applications such as drones. Firstly, a deep feed-forward neural network (D-FFNN) model is trained for Q estimation. The training strategies, such as weight initialization, K-fold cross-validation, and hyper-parameter tuning, are adopted to improve the model performance, represented for Q-factor prediction accuracy. Then, optimizing the coil design is approached to explore the optimal design geometric parameters with the high Q. This is done using the APOPT algorithm to solve the mixed integer nonlinear programming problem for high-Q double-layer spiral coil design with the constrained design requirement and resonant condition. Lastly, the coil prototype is implemented for high efficiency, achieving 777.53 of Q with a 20-cm outer diameter, 0.5-mm thickness of the coil, and 3.38-mm thickness of the dielectric layer and operating at 13.56MHz from the proposed method's design suggestion.