Rapid Battery RUL Prediction under Different Fast-Charging Protocols via Domain-Adaptive Deep Learning and Model Compression

Accurate and rapid remaining useful life (RUL) prediction of batteries under various extreme conditions is crucial for battery management systems. However, existing methods often face challenges such as limited datasets under extreme conditions, high model complexity, and weak interpretability. Therefore, this paper proposes a hybrid framework based on pruning domain-adaptive convolutional neural networks (CNN) and long short-term memory (LSTM) to study RUL prediction under different fast-charging conditions using the MIT dataset. First, four voltage-related feature matrices are extracted. Using maximum mean discrepancy (MMD) constraints, the CNN-LSTM is trained with source domain and limited target domain data to align distributions. Neuron pruning is then applied to the fully connected layer to compress the model. Results demonstrate that under sparse target domain data, the domain adaptation approach achieves significantly lower prediction errors than fine-tuning. The pruned model maintains low prediction errors while reducing parameters by 42.32%. Further, an explainable algorithm quantifies regional data contributions to identify critical voltage intervals. Ultimately, precise predictions are achieved using only key data from the 2.9–3.2V range, fully demonstrating the method's efficiency. This study provides a lightweight and interpretable solution for cross-domain battery RUL prediction under fast-charging conditions.