XGBoost-Driven State-of-Power Estimation for Lithium-Ion Batteries in Battery Management System

Abstract Predicting the State of Power (SoP) accurately in lithium-ion batteries is a key to safety and optimal performance in battery-powered systems such as electric vehicles (EVs) and renewable energy storage. Although deep learning models show promise in SoP prediction, they often need large datasets, extensive tuning, and high computational resources, which restrict their practical application in real-time Battery Management Systems (BMS). In this paper, we introduce a new machine learning architecture that combines physically-inspired synthetic data generation with an XGBoost regressor to estimate SoP efficiently and accurately with low computational cost. The XGBoost model is trained on a synthetically generated dataset and assess its performance using standard regression metrics: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R² score. In the proposed method, results are compared with CNN-LSTM-PSO with SHAP based XAI. It yields a Mean Squared Error (MSE) of 6.062 × 10⁻⁷, Root Mean Squared Error of (RMSE) 8.14 × 10⁻⁴, Mean Absolute Error (MAE) of 5.53 × 10⁻⁴, and an R² value of 99.99%. Feature importance is further examined using TreeSHAP to understand how input variables influence SoP. This research offers a scalable, data-efficient method for early-stage development and real-time deployment of SoP estimation models.