TFDSUNet: Time-Frequency Dual-Stream Uncertainty Network for Battery SOH/SOC Prediction

Abstract Accurately estimating the state of health (SOH) and state of charge (SOC) of a battery is crucial for optimizing battery performance, ensuring battery safety, and extending battery life. Existing prediction methods fail to fully exploit the periodic information of battery data in the time domain and the prediction results always lack uncertainty descriptions. To address these issues, this paper proposes a novel time-frequency dual-stream uncertainty network (TFDSUNet). Specifically, TFDSUNet can simultaneously extract the local context information in the time domain and global periodic information in the frequency domain from the battery data. Additionally, TFDSUNet can provide confidence intervals for each prediction by considering both data and model uncertainty. Extensive experiments verify that TFDSUNet can obtain more than 14% performance improvement on the battery benchmark datasets and has better generalization ability than other state-of-the-art methods. Furthermore, TFDSUNet is much lighter, with about 1% parameter number of existing methods, making it more practical for deployment and implementation on battery health management devices.