A hybrid remaining useful life prediction method for lithium-ion batteries based on transfer learning with CDRSN-BiGRU-AM

Abstract

Abstract The prediction of the remaining useful life (RUL) of widely used lithium-ion batteries (LIBs) is of great importance. Existing techniques struggle to balance prediction accuracy with execution time. To achieve accurate RUL prediction quickly, a hybrid RUL prediction method for LIBs has been developed. This method first employs a channel-wise deep residual shrinkage network to adaptively extract features from input data enhancing important information features and suppressing ineffective ones based on the significance of the feature information. Subsequently, a bidirectional gated recurrent unit is used to extract bidirectional temporal features from the processed data, and an attention mechanism is introduced to maximize the extraction of significant temporal mutual information. Finally, a fully connected layer transfer strategy is applied to transition the model from offline training to online prediction, which avoids unstable predictions due to random model initialization and significantly improves the model’s computational efficiency. The simulation results show that the root mean square error of the proposed method did not exceed 1.77% and the mean absolute error did not exceed 1.44% on the NASA dataset. Consequently, the proposed method can achieve accurate online RUL prediction accuracy for LIBs.