A Pulse-Test-Based Method for Predicting Lithium Battery DRT Curves and State Estimation

Distribution of relaxation times (DRT) curves for lithium batteries facilitates the analysis of internal dynamics and polarization characteristics, offering substantial model interpretability. Furthermore, they serve as valuable tools for estimating battery states. However, current DRT curve calculation methods based on pulse data are constrained by device sampling rates and shelving times, limiting their ability to produce DRT curves corresponding to impedances in the middle- and low-frequency ranges. This paper employs a data-driven methodology to derive battery DRT curves, thereby circumventing the constraints of sampling frequency and shelving time on the scope of the DRT prediction. The accuracy of DRT curve predictions is enhanced through the development of an encoder-decoder network that integrates the benefits of convolutional neural networks (CNNs), long-short-term memory (LSTM) networks, and multihead self-attention (MSA) mechanisms, thereby improving the network's capacity to capture both temporal and spatial features, as well as to identify the weights and long-term dependencies of sequential data. The proposed method is validated using experimental test data sets and publicly available data sets. The findings indicate that a broader prediction range of DRT curves can be achieved, and the developed neural network enhances the forecast accuracy of the DRT curves. Moreover, the states of charge (SOC) estimation is corroborated by the predicted DRT curves, revealing that the SOC estimation aligns closely with results derived from actual DRT curves, demonstrating a high predictive accuracy. This method is beneficial for analyzing the correlation between polarization reactions and the SOC across various cells.