Remaining Useful Life Prediction of Lithium-Ion Batteries: A Temporal and Differential Guided Dual Attention Neural Network

Accurate remaining useful life (RUL) prediction of lithium-ion (Li-ion) batteries contributes to the safe and reliable operation of batteries and reduces safety risks. The simple pipeline and outstanding performance of the deep learning technology have made it a widespread popular prediction method. However, most existing methods focus on the construction of the network and rarely consider incorporating task and data characteristics to impose preferences on the model, i.e., inductive biases. Therefore, a temporal and differential guided dual attention neural network (TDANet) is proposed in this article to overcome the above limitations. First, the raw and differential capacity degradation data are separately fed into the proposed temporal-guided module (TGM) and differential-guided module (DGM) to capture global time-space correlations and local fluctuations. Then, following these two modules, self-attention and convolutional layers are introduced to extract global features guided by TGM and local features guided by DGM in parallel. Finally, the output layer is applied to fuse the above features for capacity prediction. The average mean absolute error (MAE) on the NASA and CALCE databases can reach 0.0167 and 0.0153 when the prediction starting point is 40%, and the prediction results of TDANet outperform known state-of-the-art methods, indicating its superior performance.