Hybrid XGBoost-LSTM Framework for Accurate SOC, SOH, DOD and Internal Resistance Estimation in Li-ion Cells

Accurate estimation of State of Charge (SOC), State of Health (SOH), Depth of Discharge (DOD), and internal resistance is critical for Battery Management Systems (BMS) in electric vehicles and energy storage. Conventional methods fail to capture the nonlinear and temporal dynamics of lithium-ion cells, while existing machine learning approaches lack systematic benchmarking for embedded deployment. This study evaluates three hybrid models XGBoost-LSTM, XGBoost-SVR, and Linear Regression-Random Forest on high-resolution Samsung 30T single-cell data (five cycles, 6,081 timesteps). Models used 35 mutual information-selected features, identical preprocessing, and Bayesian hyperparameter optimization. XGBoost-LSTM achieved superior accuracy: SOC (R²=0.983), SOH (R²=0.985), DOD (R²=0.977), and internal resistance (R²=0.972), outperforming baselines significantly (Wilcoxon p<0.05). Computational profiling showed 15 ms inference latency and 60 MB memory usage, suitable for real-time BMS at 10 Hz. Results indicate that hybrid temporal learning improves battery diagnostics, while further validation across multiple chemistries, extended temperatures, multi-cell setups, and longer cycles is recommended for practical deployment.