Capacity estimation of lithium-ion batteries based on segment IC curve data dimensionality reduction and reconstruction methods

Abstract

Monitoring and accurately predicting battery capacity are critical to the development of advanced intelligent battery management systems (BMS). Data-driven battery prediction studies rely on the assumption of complete data and stable charge/discharge patterns. Enabling on-board prediction of batteries in non-regular charging and discharging patterns remains a challenging endeavor. To tackle this issue, this study introduces an innovative method for predicting battery capacity using the starting charging segment data to reconstruct incremental capacity (IC) curves. Firstly, the IC curve containing the degradation trend law is obtained by combining the interval point taking method with Locally Weighted Scatterplot Smoothing (LOWESS) smoothing, and the IC curve is dimensionally reduced to decompose into the voltage–time curve and the IC-time curve. Then, the voltage–time curve prediction model and the IC-time curve prediction model are built separately using the initial fragment data, and the complete one-dimensional curves are obtained by building the prediction models. The complete IC curve is then obtained by reconstructing the two curves. Finally, the SSA-SVR estimation model is built, extracting the peak of the complete IC curve and the battery capacity to train the Sparrow Search Algorithm-Support Vector Regression (SSA-SVR) model and perform capacity prediction. The method was validated on the NASA and CALCE datasets, and the root mean squared error (RMSE) of the model predictions was 0.0123 A∙h and 0.026 A∙h, respectively. The experimental results show that the starting fragment IC curve data are downscaled and predict the reconstruction of the complete IC curves. The obtained curves are able to extract valid peak information for accurate prediction of battery capacity. This paper reconstructs the complete IC curve data by reducing the dimensionality of the initial fragment data and then extracts features to predict capacity, simulates missing data or charging interruptions, improves the robustness of the model, and provides methodological support for the complement of irregular charging data.