A hybrid deep learning model for robust and data-efficient lithium-ion battery remaining useful life prediction

Lithium-ion batteries are extensively utilized in applications like new energy vehicles and aerospace owing to their high energy density and safety, but their service life diminishes due to irreversible capacity degradation from repeated charge-discharge cycles, making accurate remaining useful life (RUL) prediction critical for reliability and operational safety. Current data-driven methods often struggle with long-range dependencies, noise from capacity regeneration, and efficient data utilization. To address these challenges, this study introduces a novel hybrid neural network architecture that integrates Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) for denoising preprocessing with a Transformer-Convolutional Neural Network (CNN)- Bi-directional Gated Recurrent Unit (BiGRU) model. The framework employs health indicators extracted from voltage and current profiles as inputs, where CEEMDAN mitigates interference effects, the Transformer captures global degradation trends via self-attention mechanisms, the CNN extracts localized short-term features, and the bidirectional GRU models temporal dependencies bidirectionally. Experimental validation on NASA and our own test datasets demonstrates that the proposed approach significantly outperforms other models in key metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE), achieving high accuracy even with minimal training data (e.g., only 40% of cycles). Furthermore, cross-dataset validation demonstrates robust generalization of our model, achieving MAPE below 3.5% when transferring between NASA and our battery data without retraining. This hybrid model offers a robust, data-efficient solution for enhancing RUL prediction in practical battery management systems, with strong generalization across diverse battery types.