Lithium batteries health state prediction method based on TCN-GRU-Attention fusion model with multi-source charging information

Lithium-ion batteries inevitably experience a decline in State of Health (SOH) due to prolonged use, and continued operation increases safety risks. Therefore, it is essential to develop models that can accurately predict SOH. Cyclic aging experiments are initially conducted on lithium batteries using a self-built experimental platform to collect data on charging voltage and temperature aging. A multi-channel temporal convolutional neural network is employed to perform feature extraction on the multi-source data, preserving the dependencies of the features over time. The multi-channel input enables the neural network to capture degradation features simultaneously, enhancing its ability to characterize the degradation of lithium batteries at any moment. SOH prediction is then carried out using a combination of Gated Recurrent Unit (GRU) and Self-Attention (SA) mechanism. The SA ensures prediction accuracy by calculating the weight distribution of the multi-channel features, allowing the GRU to focus on the most significant aspects of the data. Finally, the prediction model proposed in this study is compared with the traditional Long Short-Term Memory model, the encoder model, and the fusion model. The results show that although the accuracy of the encoder model is similar to that of some fusion models, the prediction accuracy is still lower than that of the model proposed in this study. Compared with other fusion models, the mean absolute error of the proposed prediction model is reduced by more than 29% on average, and the root mean square error is reduced by at least 20% on average.