Frequency-domain temporal transformer networks for state-of-charge estimation of Li-ion batteries

Abstract

Accurate estimation of the state of charge (SOC) in lithium-ion batteries is crucial for ensuring the safe operation of new energy systems. Existing deep learning methods typically employ recurrent neural networks (RNNs) or Transformer-based architectures to encode temporal dependencies for SOC prediction. However, most of these approaches rely on sliding time windows to capture sequential features, which may introduce a time lag and consequently reduce estimation accuracy. To address this issue, this paper proposes a Frequency-Domain Temporal Transformer Network (FDTTN) that extracts features in the frequency domain to establish a direct mapping between input features and SOC. Specifically, the proposed network separates raw data into three features: voltage, current, and temperature. Frequency-domain transformation is applied to extract dominant temporal characteristics. A Transformer encoder then processes these features along the time dimension, followed by an attention-guided feature fusion module that enables interaction among the three feature streams. This enhances the representation for subsequent SOC regression. In the post-processing stage, a median filter and exponential moving average (EMA) filtering are introduced to suppress impulse noise and smooth the output, further improving noise robustness and practical applicability. Experiments are conducted on both public battery datasets and a self-collected lithium-ion battery dataset. Results demonstrate that the proposed method achieves high accuracy and strong robustness across multiple datasets, with improvements in mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) compared to baseline models.