Specialized convolutional transformer networks for estimating battery health via transfer learning

Abstract

Despite continuous advancements, modeling and predicting nonlinear, multiscale, and multiphysics battery systems, which feature inherently inhomogeneous cascades of scales, remains challenging. Deep learning offers a promising alternative by automatically extracting high-dimensional features, enhancing predictions for complex battery systems. However, existing methods fall short in achieving accurate real-time responses due to high training costs and limited generalization. To address this, we developed specialized deep neural networks for health status estimation using transfer learning. Specifically, our method employs a specialized Transformer model for time series prediction with a multi-head probsparse self-attention mechanism to conserve computational resources and improve estimation efficiency. Additionally, one-dimensional (1-D) convolution captures underlying degradation patterns for state of health (SOH) estimation. Transfer learning enables real-time SOH estimation using only charging features and labeled capacities from previous cycles. The proposed method was validated using two datasets, each with different battery chemistries and charge-discharge configurations: 77 lithium iron phosphate (LFP) batteries (nominal capacity 1.1 Ah) and 30 nickel cobalt aluminum (NCA) batteries (nominal capacity 3.5 Ah). We transferred knowledge from 57 LFP batteries to 20 same-batch batteries, achieving an RMSE of 0.247%, an R² of 99.8%, a WMAPE of 0.223%, and an MAE of 0.202%. Additionally, we applied the transfer learning model trained on the LFP batteries to a dataset of 30 NCA batteries, achieving an RMSE of 0.687%, an R² of 96.8%, a WMAPE of 0.443%, and an MAE of 0.523%. Overall, the proposed specialized network architectures have demonstrated remarkable power in achieving accurate predictions, fast training, and enhanced generalization.