Data-driven RUL prediction for lithium-ion batteries based on multilayer optimized fusion deep network

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries is crucial to optimize energy storage performance and safety. A data-driven prediction method based on a multilayer optimization fusion deep network was proposed in this investigation. Firstly, interested feature information referred to lithium-ion battery's life was identified by Pearson correlation coefficient analysis. Initial battery capacity data was then decomposed into multiple modal components through variational modal decomposition (VMD), and the particle swarm optimization (PSO) algorithm was adopted to optimize these modal sequences where the minimum average envelope entropy was applied as the fitness function. The extracted features and decomposed, optimized modal sequences were fused and introduced to a multi-channel parallel CNN-BiLSTM deep network for RUL prediction. Additionally, an attention mechanism (AM) was integrated into BiLSTM layer to improve its capacity to capture important information more effectively and efficiently. Finally, the NASA and CALCE datasets were employed to validate the proposed model by experimental study. In comparison with the CNN-BiLSTM-AM and PSO-VMD-CNN-BiLSTM models, the proposed model demonstrated superior precision. Furthermore, by comparison with general models such as CNN, LSTM, and GRU, the proposed model supplied robust generalization capabilities, which showed high potential to extended application in engineering practice.