A Voltage-Controlled Model Machine Learning Framework for State-of-Charge Estimation in Battery Management System: A Performance Evaluation

Abstract

Accurate State-of-Charge (SoC) estimation is essential for ensuring the reliability, safety, and efficient operation of Battery Energy Storage Systems (BESS), particularly in photovoltaic (PV) applications characterized by intermittent generation. Conventional SoC estimation methods may have limitations in capturing nonlinear and time-dependent voltage behavior under practical operating conditions. This study proposes a data-driven SoC estimation framework that integrates a Voltage-Controlled Model (VCM) with sliding-window temporal feature extraction and multiple machine learning (ML) algorithms. The sliding-window approach transforms sequential voltage measurements into structured lag-based features to represent battery discharge dynamics more effectively. Real-world voltage data were collected from four 45-Ah lead-acid batteries operating in a PV-BESS system from February to December 2024. More than 140,000 charge-discharge records were acquired, from which 6,226 discharge records were extracted for model development and evaluation. Five ML models, namely Support Vector Regression (SVR), Artificial Neural Network (ANN), Random Forest (RF), Long Short-Term Memory (LSTM), and eXtreme Gradient Boosting (XGBoost), were compared using mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE), and computation time. The results show that VCM-LSTM achieves the lowest prediction errors, whereas VCM-XGBoost provides the most favorable balance between predictive accuracy and computational efficiency. VCM-XGBoost achieves MSE = 0.0849, MAE = 0.1401, RMSE = 0.2913, and a computation time of 0.94 seconds. These findings indicate that combining sliding-window temporal representation with VCM and ML can support accurate and computationally efficient SoC estimation, highlighting the potential of VCM-XGBoost for resource-constrained BMS applications, particularly in real-world residential PV-BESS environments with limited computational resources.