Contribution to improving the efficiency of a wireless power transfer system using artificial intelligence techniques

Abstract

Wireless Power Transfer (WPT) technology is an innovative method for powering devices without physical wires, which has been used here to provide power to bioimplantable devices. The main design constraints are to achieve maximum transfer efficiency while keeping the implant size small enough to be suitable for the living subject's body. Magnetic Resonant Coupling Wireless Power Transfer (MRCWPT), which uses pairs of inductor coils in the external and implant circuits, is a method actively researched for this type of power transmission. The objective of this thesis is to design and optimize a high-efficiency WPT receiving coil for biomedical applications. Traditionally, optimizing WPT systems based on mathematical equations or numerical models is often time-consuming and may not yield optimal designs. To address these limitations, this thesis introduces a novel approach that integrates a machine-learning model with metaheuristic methods for design and optimization. The primary goal is to maximize the transfer efficiency for an implantable coil with dimensions of 20 mm and a transfer distance of 30 mm, operating at a frequency of 13.56 MHz. To achieve this goal, we firstly identified the critical geometric coil parameters that significantly influence the WPT system's efficiency. A model-based Artificial Neural Network (ANN) was then constructed and trained on a comprehensive dataset generated through Finite Element Method (FEM) simulations. This model predicts efficiency based on geometric coil parameters, eliminating the need for complex calculations. Subsequently, two metaheuristic algorithms: the Genetic Algorithm (GA) and the Coyote Optimization Algorithm (COA), were employed to find the optimal parameters that maximize efficiency. The proposed ANN model demonstrates exceptional accuracy, exceeding 97%. Furthermore, this WPT coil design approach significantly enhances transfer efficiency by up to 76% while drastically reducing computation time compared to conventional methods