Transformer Neural Network for Early Battery Capacity Prediction Based on Electrochemical Impedance Spectroscopy

Early prediction of the decaying capacity of lithiumion batteries is significant for e.g. early detection of abnormal cells in a battery pack. Due to the influence of the historical aging path on battery health, traditional battery capacity prediction methods rely on a large dataset of measured capacity data, which make them unsuitable for early battery capacity prediction and real-time application. To improve this situation, a multivariate and direct multi-step prediction transformer model based on electrochemical impedance spectroscopy (EIS) is proposed. To allow for real-time application, first, a robust battery capacity estimator is trained by using a convolution neural network based on EIS. Subsequently, by using the estimated capacity and impedance data, the proposed transformer model is trained for capacity prediction. Compared with five baseline models and published models, the proposed model has the best performance with prediction errors that are reduced by around 30 %. Compared with five published methods, the proposed model can indeed achieve earlier capacity prediction: only 5 cycles of data are needed for making a 15 cycle capacity prediction.