<![CDATA[The rapid growth of electric vehicle (EV) adoption creates uncertainty for power grid operators in anticipating charging load, particularly sessions that draw a disproportionately large amount of energy. This study applies the CRISP-DM (Cross-Industry Standard Process for Data Mining) framework to a large-scale synthetic EV charging dataset (1,965,239 sessions; four raw attributes: connection time, charging duration, energy delivered, and a sequential day index) to build an early, context-only classifier for high-demand sessions, defined as sessions whose delivered energy falls in the top quartile (≥75th percentile). To avoid label leakage, only information available at the start of a session connection time, planned duration, a peak-hour flag, and a weekend flag was used as predictors. The resulting target is naturally imbalanced (75:25); its effect was investigated through an ablation study comparing five classification algorithms (Logistic Regression, Decision Tree, Random Forest, XGBoost, and K-Nearest Neighbors) with and without SMOTE oversampling, evaluated on accuracy, precision, recall, F1-score, and ROC-AUC. The best configuration, XGBoost combined with SMOTE, achieved F1-score = 0.600 and ROC-AUC = 0.818, improving recall from 0.442 to 0.807 at a moderate precision cost. SHAP-based attribution identified charging duration as the dominant predictor in the trained model, followed by connection time, consistent with the raw feature correlations (r = 0.51 and r = 0.30, respectively). The full pipeline data preparation, modeling, ablation, and explainability were deployed as an interactive, open web application built with Streamlit, allowing non-technical stakeholders to reproduce the analysis and inspect results in real time.]]>
Copyrights © 2026
