Time-series prediction of lithium battery health status based on Transformer and real-world vehicle operation data

Accurate time-series prediction of power battery State of Health (SOH) is crucial for enhancing battery safety, optimizing maintenance protocols, and informing replacement strategies. This study proposes a novel SOH time-series prediction method for real-world vehicular lithium batteries, leveraging the Transformer algorithm. Initially, raw data undergoes processing to isolate charging segments characterized by relatively stable current. Subsequently, the ampere-hour integral method is employed to compute the reference capacity, and Generalized Additive Models (GAM) are utilized to extract the nonlinear attenuation trend of capacity, which serves as the model's label. Building upon this, a Transformer-based time-series prediction model is constructed, with its critical parameters optimized using the Particle Swarm Optimization (PSO) algorithm. To validate the model's performance, the proposed Transformer model is benchmarked against four prevalent time-series prediction methods: Temporal Convolutional Network (TCN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) network, and Gated Recurrent Unit (GRU). Experimental results demonstrate that the Transformer model achieves superior prediction accuracy on the test set, exhibiting Mean Absolute Percentage Error (MAPE) of 0.085% and Root Mean Square Error (RMSE) of 0.215 Ah. This superior performance suggests significant potential for its application in future large-scale vehicle battery health status time-series prediction.