Lightweight machine learning framework using temporal features for electric vehicle demand response forecasting on edge devices

The popularity of Electric Vehicles (EVs) poses particular challenges for demand-side energy management, especially in low-computation scenarios. This research builds a lightweight machine learning (ML) model to forecast EV consumption load and to maximise Demand Response (DR) strategies using real-world data. The Kaggle dataset used in this research consists of time-series EV charging data, which is preprocessed, down sampled, and augmented with time-based features, including hour, weekday, and month. Seven DR strategies, which are Peak Clipping, Valley Filling, Load Shifting, Load Levelling, Strategic Load Growth, Strategic Conservation, and Flexible Load Shape, are implemented on estimated load profiles. The model experiments consist of five ML models: Linear Regression (LR), Support Vector Regression (SVR), k-Nearest Neighbours (kNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost), and their results for both prediction and computation. Such models are selected since they are suitable for systems that lack resources. To serve all DR cases, the measurement metrics used are MAE, RMSE, and R 2 . Findings suggest that XGBoost is the most error-free technique across most DR approaches, achieving an R2 score of 0.975 in Strategic Conservation and 0.943 in Valley Filling, with the highest efficiency, seconded by Random Forest with an R2 score of 0.91. The results of linear regression and kNN were worse, with R² values never exceeding 0.50 across most DR strategies. The work demonstrates the capacity to deliver successful performance from lightweight ML models for EV load prediction and DR modelling, and that such models can provide scalable information to grid operators and policymakers operating within constrained computational environments.