Physics-Informed Deep Learning-Based State of Health and Efficiency Prediction Framework for KY Converter-Driven Electric Vehicle Battery Systems

Abstract

Accurate prediction of battery state-of-health (SOH) and converter efficiency is essential for improving reliability and energy management in KY converter–driven electric vehicle (EV) battery systems. Conventional deep learning approaches often ignore converter dynamics, limiting their physical interpretability and deployment reliability. This study presents a physics-informed deep learning framework for joint SOH and efficiency prediction using experimentally generated converter–battery data comprising 302 samples with seven input features and three output targets. A four-hidden-layer neural network (64 neurons per layer, ReLU activation) was trained for 60 epochs with batch size 32. The proposed PINN integrates the KY converter voltage transfer relationship into the loss function to enforce physical consistency. Experimental results show that the baseline ANN achieved RMSE of 0.7073 and MAE of 0.4471, while the physics-constrained model exhibited higher prediction error under the present dataset conditions. The developed model contains 12,995 parameters, requires 50.76 kB memory, and achieves an inference time of 0.00169 s, demonstrating suitability for embedded MCU deployment. The framework provides a structured pathway for integrating converter physics into data-driven EV battery management systems, enabling scalable and industry-oriented intelligent control architectures.