Real-world cross-battery state of charge prediction in electric vehicles with machine learning: Data quality analysis, data repair and training data reconstruction

Predicting the state of charge (SOC) of electric vehicle batteries in the upcoming period is essential for estimating remaining mileage and pre-warning battery safety. Although machine learning (ML) models have shown potential, low-quality data and diverse operating conditions of real-world vehicles pose challenges to the reliability and generalizability of their actual applications. This study systematically investigates the performance of ML models for real-world cross-battery SOC prediction. Online datasets from numerous real-world electric vehicles are used to verify the models. The impact of prevalent missing and idle values in online datasets on model performance is quantitatively analyzed. Data repair and training data reconstruction strategies are proposed to enhance dataset quality and model reliability. The results demonstrate that ML models trained on single-vehicle data can be generalized to predict SOCs of other vehicles. However, missing and idle values significantly degrade performance. By using the proposed data repair strategies, the R2 improved from <0.1 to >0.9 for low-quality datasets. All ML models achieved R2 value of >0.99 with reconstructed training data. Moreover, a comparative analysis showed that linear regression is the optimal SOC prediction model in terms of both computational efficiency and cross-battery prediction accuracy.