Electric vehicle charging station demand prediction model deploying data slotting

Abstract

Accurate prediction of energy requirement at charging station is essential for optimizing infrastructure usage, ensuring grid stability, and minimizing operational cost. Literatures suggest deployment of machine learning techniques to forecast the station demand. One major challenge associated with development of machine learning models is the inherent uncertainty in electric vehicle charging behaviour that includes variations in charging patterns, user preferences, and vehicle types. The conventional pre-processing techniques fail to dislodge nonlinearities and highly random patterns that include very low or zero-charging. Employing such techniques affects the model's forecast accuracy. This article performs data-slotting during pre-processing stage and then selects the best among 1-h, 2-h, 3-h and 4-h slots, to frame the feature vectors. The 4-h data with minimum variance is suggested to frame the dataset. Four distinct datasets, comprising different combination of average and total demands as predictor and response respectively are considered. The created dataset is deployed in Random Forest, Categorical Boosting, Extreme Gradient Boosting and Light Gradient Boosting models. The article recommends Categorical Boosting Regression model with least mean absolute error, mean square error and root mean square error of 0.0726, 0.0112, and 0.1059 respectively. Furthermore, the use of feature vector comprising of aggregated load for prescribed slots and the response representing the aggregated demand is observed to provide the least prediction error by the suggested model. The suggested model fed by the proposed feature vector offers significant advantage to charging station operator by enhancing the operational efficiency while performing resource and cost management with strategic planning.