A MDA-LSTM network for remaining useful life estimation of lithium batteries

Remaining useful life (RUL) of energy storage batteries estimation is of great significance to battery failure warning and battery safety. Previous methods have primarily relied on the battery’s capacity as the sole feature, neglecting the potential information contained within multiple features. To address this limitation and effectively make use of multi-features and temporal information, this paper proposes a novel neural network model called multi-feature fusion and dual attention long short-term memory (MDA-LSTM) network. The MDA-LSTM network comprises a multi-feature fusion (MFF) module and a dual attention module (DAM). The MFF module captures dependencies among multiple features and computes their relationship weights, enhancing the model’s effective use of feature information. DAM includes a local attention module for short-term temporal focus and a global attention module for capturing the overall time series trend. This dual attention mechanism enables the model to consider both short-term and long-term dependencies in the data. Through the combined influence of these two modules, our model can more effectively learn timing information, thereby improving prediction accuracy. We conduct comprehensive experiments on the NASA dataset and CALCE dataset, demonstrating that MDA-LSTM achieves higher prediction accuracy compared to other baseline methods.