State of Charge Prediction for Electric Vehicles Based on Integrated Model Architecture

To enhance the accuracy of SOC prediction in EVs, which often suffers from significant discrepancies between displayed and actual driving ranges, this study proposes a data-driven model guided by an energy consumption framework. The approach addresses the problem of inaccurate remaining range prediction, improving drivers’ travel planning and vehicle efficiency. A PCA-GA-K-Means-based driving cycle clustering method is introduced, followed by driving style feature extraction using a GMM to capture behavioral differences. A coupled library of twelve typical driving cycle style combinations is constructed to handle complex correlations among driving style, operating conditions, and range. To mitigate multicollinearity and nonlinear feature redundancies, a Pearson-DII-based feature extraction method is proposed. A stacking ensemble model, integrating Random Forest, CatBoost, XGBoost, and SVR as base models with ElasticNet as the meta model, is developed for robust prediction. Validated with real-world vehicle data across −21 °C to 39 °C and four driving cycles, the model significantly improves SOC prediction accuracy, offering a reliable solution for EV range estimation and enhancing user trust in EV technology.