Machine Learning-Based Electric Vehicle Charging Demand Prediction Using Origin-Destination Data: A UAE Case Study

Abstract

Optimal prediction and coordination of the energy demand of electric vehicles (EVs) is essential to address the energy availability and range anxiety concerns of current and potential EV users. As a result, different EV demand predictors are developed in the literature based on traffic simulators and/or locally-generated EV charging datasets, to provide the required inputs for EV demand management programs. These predictors, however, may not reliably scale to model the EV energy requirements in different regions, particularly with the scarcity of real-world data on EV driving patterns. This work proposes a data-driven, machine learning (ML)-based EV demand predictor based on vehicular traffic flow data between different origin-destination (OD) pairs. The proposed model incorporates the driving patterns in the regions under consideration to determine the corresponding EV energy consumption and hence, the minimum EV energy requirements per trip. The data used in this work is obtained from TomTom Move O/D Analysis portal for the cities of Dubai and Sharjah, UAE. Different ML models are trained on the dataset to develop the EV demand predictor, namely random forests (RF), extreme gradient boosting (XGBoost), multilayer perceptron (MLP) and linear regression models. Results reveal that the MLP offers a superior performance to all other models, with an $R^{2}\ >\ 0.8$ and a symmetric mean absolute percentage error of ≈ 20% on both the training and testing data subsets, and a significantly lower training time compared to RF and XGBoost. This makes it suitable for EV demand predictions to incorporate regular updates in vehicular traffic flow data for further model tuning.