Dual‐Stage Predictive Model Incorporating Extreme Gradient Boosting‐Based Data Linearizer for High‐Precision State‐of‐Charge Estimation of a Lithium‐Ion Battery

This is the authors' abstract. We don't add key points for this paper.

Most studies have primarily focused on state‐of‐charge (SOC) estimation models, often neglecting the improvement of battery datasets, resulting in raw data being commonly used to train these models. However, this study emphasizes the quality of the input data for SOC estimation models. A dual‐stage predictive model was proposed in which the first stage, a Ridge data linearizer (RDL), is an algorithm designed to linearize inherently nonlinear lithium‐ion battery (LIB) data, whereas the second stage employs an extreme gradient boosting (XGBoost) model, a machine learning (ML) algorithm used for SOC estimation using the linearized dataset from the RDL as the input. To assess the impact of the linearization process, a conventional XGBoost model was developed and trained on the raw dataset for performance comparison. In addition, a feature engineering technique was introduced during the data preprocessing stage to augment the battery features by manipulating existing data, thereby generating additional features automatically accessible to the estimation model. The SOC estimation comparison of both models highlighted the critical importance of data quality and the potential for data enhancement in improving the accuracy and reliability of SOC estimation models.