Efficient Multi-Objective Coil Design with Deep Neural Network-Accelerated PSO in WPT Systems

Traditional multi-objective coil design methods for wireless power transfer (WPT) systems heavily rely on empirical intuitions and finite element method (FEM) simulations, which are laborious and time-consuming. To expedite the multi-objective coil structure design process for different WPT applications, this paper proposes a deep neural network-accelerated particle swarm optimization (DNN-PSO) method. The DNN-PSO integrates a deep feedforward neural network to efficiently predict the inductances of different coils under varied misalignments. Moreover, a unique multi-variate reward function is proposed to numerically evaluate the fitness between the obtained coil structures and the given multidimensional design objectives. Furthermore, a forbidden searching list (FSL) scheme is introduced to dynamically adjust the searching space by excluding the coil structures with low reward values from the subsequent design process, thus greatly improving the converging speed of the DNN-PSO. The effectiveness of the DNN-PSO is validated with a design case for an LCC/S-compensated WPT system. The proposed DNN-PSO can pinpoint the optimal coil design within 11.20s, based on which a prototype coil is manufactured. The transfer efficiency of the WPT system with the prototype coil is measured at a transmitted power of 1kW and 2kW, demonstrating high transfer efficiencies of 94.64% and 95.64%, respectively.