An Improved Hybrid Machine Learning Algorithm for State‐of‐Charge Estimation and Anomaly Detection in Smart Grid Energy Storage Systems
Résumé
ABSTRACT This work proposes a hybrid Piecewise AutoRegressive with eXogenous inputs‐Support Vector Machine (PWARX–SVM) model for accurate State of Charge (SoC) estimation and anomaly detection in Lithium‐ion Batterie Energy Storage System (Li‐BESS). The model identifies discrete operational states rest, charge, and discharge while providing interpretable insights into battery behavior. Validation against experimental data shows high agreement, with R 2 = 0.935 for charging and R 2 = 0.959 for discharging, confirming the model's robustness and its ability to detect early anomalies such as cell degradation or connection faults. A Long Short‐Term Memory (LSTM) network was also implemented for comparison, achieving higher accuracy ( R 2 = 0.998 and R 2 = 0.992) but offering limited interpretability. While LSTM excels in prediction, the PWARX‐SVM model provides a clearer understanding of physical dynamics and operational modes, essential for predictive maintenance. Overall, the proposed hybrid framework effectively bridges data‐driven forecasting and physical interpretability, offering a reliable and explainable solution for real‐time SoC monitoring and fault prevention in smart grid energy storage systems.
