A multi-vehicle study of early-data requirements for state-of-health estimation of electric vehicle batteries using real-world operating data

Le résumé fourni par la source

Estimating battery state of health (SOH) from real-world electric vehicle data is difficult because SOH cannot be measured directly and only limited data are available during the early stage of vehicle operation. Using approximately three years of operating data from 100 Changan electric vehicles, this study investigates SOH estimation and early-data requirements. Reference SOH was derived from charging records using ampere-hour integration, repeated random subinterval estimation, median absolute deviation (MAD)-based outlier detection, and Kalman filtering. Features were extracted from incremental capacity analysis (ICA), cell-voltage and temperature inconsistency, and vehicle operating conditions, and Mamba was used to model degradation across successive cycles. One benchmark vehicle was used to determine the feature set, fixed network architecture, and common modeling settings. The resulting modeling scheme was then applied to the other 99 vehicles, with key hyperparameters selected separately using validation data. For each vehicle, early-life samples ranging from 5% to 40% were used for training, the following 10% for validation, and the remainder for testing. On the benchmark vehicle, Mamba achieved a root mean square error (RMSE) of 0.0777, a mean absolute error (MAE) of 0.0618, and a mean absolute percentage error (MAPE) of 8.10%, outperforming the comparison models. For the other 99 vehicles, training with the first 30% of samples yielded mean RMSE, MAE, and MAPE values of 0.0564, 0.0442, and 5.28%, respectively. Errors decreased as more early data were used, with a clear improvement from 10% to 15% and limited gains beyond approximately 30%. These results provide a reference for selecting the proportion of early data used in SOH modeling for electric vehicles of the same model.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.