Time-Frequency Structured State-Space Neural Network for Battery State-of-Charge Estimation

Abstract

This paper presents a novel deep learning approach for state-of-charge estimation that integrates three complementary methods. First, a discrete wavelet transform is used to extract time-frequency features, mitigating inherent limitations of the short-time Fourier transform, such as the fixed time-frequency resolution trade-off and the reliance on complex-valued representations. Second, a structured state-space model is employed to more effectively handle the sequential nature of time-frequency features. Third, a composite loss function that combines mean squared error and contrastive loss is designed to further enhance performance by encouraging the model to capture the underlying structure of the data. In addition, we employ a modified model architecture that incorporates a classification head tailored to the regression task in order to maximize the benefit of the proposed contrastive loss. The proposed model achieves superior estimation accuracy with fewer learnable parameters than baseline methods reported in previous studies. Comprehensive ablation studies are conducted to quantify the contribution of each design choice, including input feature generation, model architecture and loss function formulation.