Logo image
Back
Explainable artificial intelligence (XAI)-enhanced modeling of controlled rectifiers for electric vehicle charging systems
Journal article   Open access   Peer reviewed

Explainable artificial intelligence (XAI)-enhanced modeling of controlled rectifiers for electric vehicle charging systems

Seda Kul, Seyit Alperen Celtek, Selami Balci and Farhad Shahnia
Machine learning with applications, Vol.25, 100962
2026
DOI: https://doi.org/10.1016/j.mlwa.2026.100962
pdf
XAI6.11 MBDownloadView
Open Access CC BY V4.0

Abstract

12-Pulse controlled rectifier CatBoost Electric vehicle charging systems Ensemble learning SHAP XAI
Electric vehicle proliferation necessitates efficient power conversion systems with minimal harmonic distortion. Among the established solutions for electric vehicles’ charging infrastructure, 12-pulse thyristor-based controlled rectifiers integrated with phase-shifting transformers offer superior harmonic performance; however, their advanced design requires iterative refinements to address nonlinear magnetic behavior, loss mechanisms, and complex trade-offs among harmonic mitigation, transformer rating, volume, cost, and thermal limits. This study presents an explainable artificial intelligence (XAI)-enhanced framework that moves beyond simulation-driven, trial-and-error design toward informed data-driven decision-making supported by interpretable feature-importance analysis. The proposed framework transforms black-box ensemble models into transparent decision-support tools, enabling engineers to better understand the influence of key design parameters on grid-current total harmonic distortion (THD). An ANSYS TwinBuilder–Maxwell 3D co-simulation is employed to generate 4750 operating points linking the thyristor firing angle, output filter capacitance, and load resistance to the grid-current THD for training 15 machine learning algorithms. The evaluated machine learning models demonstrate that tree-based ensemble methods provide the highest predictive performance, with CatBoost achieving R² = 0.964 and RMSE = 0.358. SHAP-based explainability analysis identifies the firing angle as the dominant control parameter (54% feature contribution), followed by load resistance (34%) and filter capacitance (12%). Validation against an independent ANSYS Maxwell–Twin Builder co-simulation at a representative operating point resulted in a prediction error of 1.18%, demonstrating good agreement between the surrogate model and the simulation within the investigated parameter space.

Details

Metrics

1 Record Views
Logo image