Enhanced state of charge estimation in lithium-ion batteries based on Time-Frequency-Net with time-domain and frequency-domain features

Developing high-accuracy state of charge (SOC) estimation algorithms is essential for efficiently and safely using lithium-ion batteries. However, for data-driven-based models, adaptation, robustness and generalization must be considered to estimate SOC accurately. For this reason, Time-Frequency-Net (TFN) is introduced to enhance the accuracy and stability of SOC estimation from time-domain and frequency-domain features. TFN is essentially the combination of the Mamba Block and the Signal Decomposition Block (SDB). Mamba Block optimizes long-term dependency modeling in time series and reinforces time-domain feature extraction. Additionally, SDB incorporates Fast Fourier Transform with convolutional neural networks, facilitating frequency-domain feature extraction and representation. This approach allows TFN to comprehensively learn the patterns of periodicity, battery dynamics and potential noise. TFN can synthesize time-domain and frequency-domain information and identify data trends to output reliable SOC estimations, improving the robustness of the model and ensuring the battery's dynamic behaviors learning. The experimental results indicate that TFN with an overall average mean absolute error of 0.71 % outperforms other models on self-collected datasets. Furthermore, it performs superiorly in various conditions, including currents with different signal-to-noise ratios, extreme temperatures, and distinct battery types. The effectiveness across varying discharging rates and battery-aging levels confirms the adaptability and generalizability of TFN.