Enhanced State-of-Health and Remaining Useful Life Predictions for Lithium-Ion Batteries Through Randomized Grid-Search Hyperparameter Optimization

Abstract Accurate estimation of the state-of-health (SOH) and remaining useful life (RUL) of lithium-ion batteries is critical for ensuring their reliability, safety, and efficient utilization in energy storage systems in electric vehicles. In this study, a systematically regularized long short-term memory (LSTM)-based model (Model-A) was first developed, integrating dropout, L2 regularization, adaptive learning rate scheduling, and early stopping to mitigate overfitting and enhance model stability. Building upon this foundation, an optimized LSTM framework (Model-B) was proposed, employing randomized grid-search-based hyperparameter optimization to further improve prediction accuracy and generalization. The performance of the proposed model was evaluated using three benchmark NASA battery datasets under various train–test split ratios. The results revealed that Model-B consistently outperformed Model-A, achieving significant reductions in prediction errors and demonstrating robust learning behavior across all configurations. The model achieved the most balanced performance at intermediate split ratios, reflecting an optimal tradeoff between training sufficiency and testing reliability. The predicted RUL values are closely aligned with experimental observations. Moreover, the model maintains stable SOH prediction accuracy under 5% Gaussian noise, indicating robustness to measurement uncertainty and suitability for real-world battery monitoring. Combined with offline training and lightweight online inference, the approach requires minimal computational resources, making it practical for real-time deployment on embedded battery management system edge devices.