An improved bidirectional long short-term memory hybrid neural network with Gaussian filtering for multi-temperature state of charge estimation of lithium-ion batteries

The new energy revolution is fundamentally reshaping the global energy structure. Power lithium batteries face issues such as charge-discharge imbalance and limited endurance. To enhance the performance and economic efficiency of power lithium batteries, an improved bidirectional long short-term memory neural network (BiLSTM) with Gaussian filtering for multi-temperature state of charge (SOC) estimation of lithium-ion batteries is proposed, named TCN-BiLSTM-SA. The algorithm employs Gaussian filtering to smooth the data, which is combined with the original data as input for data augmentation. Subsequently, the data is trained in a hybrid neural network composed of temporal convolutional networks (TCN) and BiLSTM. A self-attention mechanism (SA) is incorporated to adjust feature weights, enabling accurate prediction of the SOC for lithium-ion batteries. The proposed method was validated under various temperatures and operating conditions. The algorithm achieved a root mean square error (RMSE) of less than 1.496%, a mean absolute error (MAE) of less than 1.412%, and an R2 coefficient of determination of no less than 99.6%. These results indicate that the proposed approach exhibits high estimation accuracy and superior predictive performance.