Short-term discharge-based state of health estimation of lithium-ion batteries

Accurate and timely prediction of lithium-ion battery State of Health (SoH) is critical for ensuring the safety, lifespan, and reliability of Battery Management Systems (BMS) in electric vehicles (EVs) and other energy storage applications. Conventional SoH estimation techniques often rely on full discharge cycles, which are both time-consuming and unsuitable for real-time deployment in embedded systems. This study proposes a data-efficient framework that leverages partial discharge profiles, as short as 10 min, for rapid and reliable SoH prediction. Four machine learning (ML) models were comprehensively evaluated using NASA’s battery degradation dataset: a standard Long Short-Term Memory (LSTM), a Bidirectional Long Short-Term Memory network with Multi-Head Self-Attention (BiLSTM-MHSA), a Convolutional Neural Network (CNN), and CatBoost. In this study, we first establish a baseline by training and evaluating models on complete discharge cycles from NASA’s battery degradation dataset. Among the models tested, the LSTM achieved the highest accuracy (R² = 0.9036, RMSE = 0.0302, MAE = 0.0139), followed by BiLSTM-MHSA (R² = 0.8714), CatBoost (R² = 0.8463), and CNN (R² = 0.7269). Building on this baseline, we introduce a data-efficient framework that leverages partial discharge profiles—as short as 10 min—for rapid SoH prediction. Results demonstrate that LSTM maintained superior performance across all durations, achieving R² = 0.9402 (RMSE = 0.0237, MAE = 0.0123) with 30 min of data and R² = 0.9435 (RMSE = 0.0231, MAE = 0.0113) with only 10 min. BiLSTM-MHSA excelled at the shortest duration (R² = 0.9460 at 10 min) while CatBoost emerged as a competitive lightweight alternative (R² = 0.9326 at 20 min). In contrast, CNN underperformed across all durations due to its limited capacity to model long-term dependencies. To enhance interpretability, SHAP (SHapley Additive exPlanations) analysis identified early voltage dynamics and temperature behavior as the most influential predictors of battery degradation. Overall, this work presents a scalable and interpretable framework for SoH prediction from full and partial discharge data, enabling accurate, real-time diagnostics while reducing testing overhead, paving the way for efficient and lightweight BMS deployment.