Predicting Automotive Battery Degradation Trajectories Using Transformer With Variational Auto-Encoder Based Data Augmentation

With the widespread adoption of electric vehicles (EVs), accurately predicting the degradation trajectory of batteries has become crucial to ensure the safe and reliable operation of EVs. In this study, a novel battery degradation trajectory prediction method based on actual EVs operational data is proposed. Firstly, the state of health of the battery is estimated using Coulomb counting methodology, refined through analysis of noise sources and charging pattern, and further denoised through Empirical Mode Decomposition. Moreover, a Transformer-based battery degradation trajectory prediction method is introduced, which demonstrates significantly higher prediction accuracy compared to various existing methods. To address the scarcity of real-world aging data, a data augmentation method based on Variational Auto-encoder (VAE) is proposed. Through data augmentation, diverse and representative battery aging data are generated, enhancing the prediction performance of all neural networks. Following data augmentation, the Transformer still exhibits the lowest Mean Absolute Error (MAE) of 0.5813% and Root Mean Square Error (RMSE) of 0.7012%. This indicates that the VAE enhances the accuracy and reliability of battery degradation trajectory prediction. Additionally, the high accuracy of the VAE-Transformer in early prediction scenarios is validated. A comprehensive analysis of the resource consumption of the neural networks further demonstrates the high applicability of the proposed method in practical applications.