Deep learning-based RUL and SOH prediction of lithium-ion batteries using LOCO.

Degradation in lithium-ion batteries negatively affects the reliability, safety, and cost-effectiveness of modern energy storage systems, making accurate Remaining Useful Life (RUL) prediction and State of Health (SOH) assessment essential for battery prognostics. In this work, a comprehensive comparative benchmarking framework for battery prognostics is presented using three representative deep learning architectures, namely, Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and Temporal Fusion Transformer (TFT). To provide a fair and reproducible comparison between models and datasets, identical preprocessing, feature engineering, normalization, and training procedures are applied to two publicly available benchmark datasets, namely, the NASA Ames Battery Aging Dataset and the Oxford Battery Degradation Dataset. Using a sliding-window approach, degradation-related features and temporal statistical descriptors are organized as time-series sequences capable of capturing nonlinear degradation behavior. To make a realistic assessment of cross-cell generalization, a strict Leave-One-Cell-Out (LOCO) validation protocol is used, where the models are validated on completely new battery cells. The prediction task is defined as a regression problem for RUL prediction, and the SOH based on the normalized discharge capacity is used to analyze the degradation behavior. Model performance is evaluated by Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and coefficient of determination (R2) and is expressed as mean ± SD standard deviations across repeated random seeds. We differentiate stochastic training effects from variations caused by different LOCO folds by additional statistical analysis with confidence interval estimation, seed-level variability, and evaluation on cell level. The results show that the degradation complexity and cross-cell heterogeneity affect the model performance, which highlights the importance of carefully designed evaluation protocols, statistical reliability analysis, and rigorous cross-cell validation to ensure reliable model benchmarking in lithium-ion batteries' prognostics.