Remaining Useful Life Prediction of Oxford Lithium-Ion Battery Cells Using a Long Short-Term Memory Neural Network

Abstract Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries underpins the safe and economical operation of electric vehicles, grid-scale energy storage, and portable electronics. This paper presents a Long Short-Term Memory (LSTM) deep learning framework for RUL prediction, validated on the Oxford battery degradation dataset collected by the University of Oxford, comprising eight commercial lithium-iron-phosphate (LFP) pouch cells cycled under controlled laboratory conditions at 40 °C to an 80% capacity failure threshold. Each discharge half-cycle was characterized through ten statistical features extracted from the raw voltage, current, temperature, and time signals, capturing the shape of the discharge curve, thermal behavior, and discharge duration dynamics that evolve systematically as the cells age. The resulting feature sequences were assembled into sliding windows of ten cycles and supplied to an LSTM network with 50 hidden units, dropout regularization at a rate of 0.3, and the Adam optimizer, trained with early stopping to limit overfitting. Generalization to unseen cells was assessed using leave-one-cell-out cross-validation (LOOCV) across all eight Oxford cells. The proposed model achieved an average root mean square error (RMSE) of 5.59 cycles and a coefficient of determination (R²) of 0.769, with Cells 3 through 6 predicted to an R² above 0.90 while two cells with atypical degradation trajectories, Cells 1 and 7, proved more challenging. The LSTM consistently outperformed linear regression, support vector regression, random forest regression, and gated recurrent unit baselines evaluated under the same cross-validation protocol. A subsequent ablation study showed that the model retains useful predictive accuracy even when the capacity feature is withheld, and a window-length sensitivity analysis confirmed that a ten-cycle window offers a favorable balance between historical context and model stability for this dataset. These findings support the use of LSTM-based prognostics for LFP battery health management. Keywords: Remaining Useful Life; Lithium-Ion Battery; Oxford Battery Degradation Dataset; Long Short-Term Memory; Leave-One-Cell-Out Cross-Validation