Machine-Learning-Based Optimal Design of a 100 kW, 99% Efficiency, 70 kV Insulated, Leakage-Inductance-Integrated Medium-Frequency Transformer

This article proposes a machine-learning-based optimal design methodology for a leakage-inductance-integrated medium-frequency transformer (LI2MFT), intended for medium-voltage (MV) grid-connected solid-state transformer (SST) systems. The LI2MFT must satisfy multiple design objectives simultaneously, including MV-level galvanic isolation, target leakage inductance for dual-active-bridge (DAB) operation, and high efficiency, power density, and thermal reliability. Owing to the nonlinear dependencies between design parameters, such as core shape, winding configuration, and spatial layout, and performance metrics, such as leakage inductance and losses, closed-form analytical solutions are not available. Therefore, a data-driven and simulation-based design approach is adopted. In this study, machine-learning regression models were trained on 30 000 FEA simulations to accurately predict leakage inductance, losses, and hotspot temperature rise. These models enable efficient and high-fidelity exploration of the design space. A Pareto front was generated through multiobjective optimization using the nondominated sorting genetic algorithm II (NSGA-II), considering efficiency and volume. The final design was selected from this Pareto front based on specific application requirements. A 100 kW, 99%-efficiency, and 70 kV insulated LI2MFT prototype was fabricated and experimentally tested to evaluate the proposed design methodology. The measured results showed good agreement with model predictions, validating the effectiveness of the proposed methodology.