Optimizing battery health monitoring in electric vehicles using interpretable CART–GX model
Mohnish Karthikeyan B, N. Nikhil Anirudh, Navaneetha Krishnan S, Christopher Columbus C, Aravind Chellachi Kathiresan
Abstract
ABSTRACT The increasing adoption of electric vehicles (EVs) has created a growing demand for accurate and reliable predictions of lithium-ion battery State of Health (SoH) to ensure optimal performance, safety, and longevity. This paper introduces CART-GX, a novel hybrid model integrating advanced deep learning techniques to enhance SoH prediction. The proposed model combines Convolutional Neural Networks (CNNs), Gated Recurrent Units (GRUs), attention mechanisms, residual connections, and transformers to extract spatial and temporal features, prioritize critical information, and model long-range dependencies. The methodology is validated on both the NASA and CALCE battery datasets, demonstrating strong generalization across diverse chemistries and cycling protocols. Moreover, this research employs SHAP values to interpret the results of CART-GX, identifying that the battery capacity in each cycle contributes the most to the estimation of SoH in both datasets. The proposed model achieved exceptional performance, with a root mean square error (RMSE) of 0.0130%, mean absolute error (MAE) of 0.0087%, and an R² value of 99.999% on the NASA dataset, and maintained robust accuracy on the CALCE dataset, achieving RMSEs as low as 1.23% and R² values exceeding 99%. By effectively combining the hybrid model with explainable AI, CART-GX addresses limitations of traditional methods and offers a practical solution for real-world battery health monitoring. Such advancements contribute to the development of a more efficient and sustainable EV ecosystem, supporting the broader adoption of EVs and the transition to cleaner transportation technologies.