A facile data-driven battery capacity estimation framework for on-road plug-in hybrid electric vehicle

A data-driven framework with strong generalization capabilities is proposed to effectively extract features and easily access battery capacity. This framework can make highly accurate predictions for the battery capacities of plug-in electric vehicles. The feature extraction process is entirely based on statistics, which are always available and can be generalized to various types of battery data. An improved ampere-hour integral method can easily access battery capacity with just short-charging segments lasting 500 ​s. Several machine-learning models are trained to verify the framework's effectiveness, with the best model achieving a test error of 0.84 ​% based on leave-one-out validation. SHAP values are used to provide a reasonable interpretation of the relationships between the constructed features and model outputs. The proposed framework offers advantages such as reduced computational resources, wide generalization, and high prediction accuracy, showing great potential for battery management.