Evaluating Deep Learning Models for Lithium-Ion Battery RUL Prediction via Unified Benchmarks: A Comparative Study

Lithium-ion battery remaining useful life (RUL) prediction is critical for the safe and efficient operation of energy storage systems. Traditional physics-based and empirical data-driven methods rely heavily on expert knowledge and predefined degradation models, limiting their generalizability across diverse operating conditions. In recent years, deep learning has emerged as a powerful alternative, automatically learning features from raw sensor data without requiring explicit electrochemical assumptions. However, comparative studies of deep learning models for RUL prediction remain fragmented due to inconsistent benchmarks and preprocessing protocols. To address this, we present a unified evaluation of 24 models from six architectural families using the NASA PCoE battery dataset, covering multiple temperatures and discharge currents. All models share standardized inputs and preprocessing, including moving average smoothing with a window of 45 cycles to attenuate measurement noise. Evaluation covers prediction accuracy (RMSE, MAE, R2, MAPE, SMAPE, MedAE) and computational efficiency, with Pareto frontier analysis. Results show that WaveNet achieves the best overall performance (RMSE = 0.0194, R2 = 0.9834), followed by TCN (RMSE = 0.0247) and CNN (RMSE = 0.0307), while performance varies across families depending on the specific architecture and data regime. This study provides, for the first time, a systematic benchmark and practical guidance for deep learning model selection in battery health management.