Multi-Domain Machine Learning for Predictive Electric Vehicle Charging Management: A Framework Integrating Behavioral, Infrastructure, and Market Data

Abstract

The growing popularity of electric vehicles (EVs) is bringing to light several new challenges that the infrastructure of the management of charging and smart grid operation as well as energy demand planning have to cope with. The accuracy of the prediction of the EV charging pattern can significantly increase the charging station utilization, reduce the risk of grid overload and enable the integration of renewable energy sources and also support development of future intelligent transportation systems. However, the majority of the EV charging prediction models being used nowadays are developed on the basis of a single domain dataset, and they are not able to capture or show the complex web of user-charging behavior, charging station features and market-related factors interaction. These limitations prevent such models from being applied in diverse scenarios and also from being deployed for actual charging purposes. With an aim to counter these issues, we present an innovative Multi-Domain Machine Learning Framework for EV Charging Prediction in this work. By combining datasets of different types, it enhances accuracy, robustness, and readiness for deployment of the predictive models. The proposed framework integrates three publicly available datasets describing EV charging behavior, charging infrastructure, and dynamics of the EV market. A preprocessing pipeline was developed to do data cleanup, filling missing data, removing duplicates, converting timestamps into the same time format, categorical encoding of variables, scaling of numeric features, and treatment of class imbalance. It should be noted that LightGBM was used to make an efficient computation and strong performance model on the data that is in structured tables, while TabNet made use of sequential attention mechanisms for carrying out interpretable deep learning on heterogeneous tabular datasets. Accuracy, Weighted F1-score, Macro F1-score, confusion matrix analysis, confidence calibration, computational efficiency, and ensemble weight optimization were used as evaluation metrics. The results from the experiments show that the combination of different information types, including behavioral, infrastructure, temporal, and market-related, leads to a significant improvement in predicting performance compared to datasets when individual datasets are used. XLM-RoBERTa was the model that delivered the highest classification accuracy of 98.76 when tested against the integrated multi-domain dataset, while TabNet and LightGBM came in second and third at 96.89 and 85.67, respectively. It is further revealed from the study that TabNet offers a good trade-off between predictiveness and computational requirements in a system. In contrast, LightGBM is not only the fastest but also requires minimal computing, which makes it a choice for systems under a tight budget. Through the ensemble weight optimization strategy, prediction reliability was enhanced since it assigned each model the best possible contribution. It is clear from this that transformer-based contextual learning can be successful, even under heterogeneous EV charging environments. These results contribute to the development of ML technologies applied to the electric mobility sector and, most importantly, lay a groundwork for future intelligent EV charging solutions and smart city infrastructures.