Multi-scale nonlinear temporal-aware enhancement network for lithium-ion battery life prediction

Accurate prediction of lithium-ion battery lifetime is crucial for the health management of electric vehicle batteries. However, current prediction models based on machine learning or traditional neural networks are often limited in their ability to effectively capture global dependencies within long-term sequential data and typically underperform when handling local fluctuations that occur during battery degradation, which consequently constrains their prediction accuracy. To address these limitations, an innovative battery lifetime prediction model named multi-scale nonlinear temporal-aware enhancement network (MNTA-Net) is proposed. This model incorporates a Multi-scale Nonlinear Feature Extraction (MNFE) module. Traditional pooling operations are abandoned in this module in favor of employing non-linear activation functions to preserve detailed features, thereby enabling the effective capture of mutation points and local fluctuations within the capacity sequence. Simultaneously, a Temporal-Aware Enhancement (TAE) module is designed to deeply integrate underlying temporal networks with an attention mechanism. This integration is intended to achieve precise modeling of long-term dependencies and facilitate the dynamic perception of key cycle nodes. Experimental results obtained from several public datasets demonstrate that the proposed method significantly outperforms mainstream benchmark models across evaluation metrics such as RMSE, MAE, and R². These findings validate the effectiveness and generalizability of the multi-scale nonlinear feature extraction and temporal enhancement mechanisms in improving the accuracy of battery lifetime prediction.