Predictive Modeling of Electric Vehicle Charging Behavior Using Ensemble Learning and Temporal Pattern Recognition

The use of electric vehicles (EV) is being increased in the world as a way of shifting to more sustainable transportation systems. Nevertheless, the irregularity of EV charging patterns is a leading problem to the management of power grid, energy distribution planning, and optimization of charging infrastructure. Proper forecasting of EV charging behavior such as the energy consumption, time spent charging and the cost is important to grid operators, charging station providers and energy policy makers. The conventional forecasting techniques have a hard time to model the complicated, non-linear dependencies among time-related attributes, user behavior trends, car properties, and environmental attributes. This study introduces a detailed machine learning model that compares the 13 state-of-the-art algorithms of predicting EV charging energy consumption. We build a superior temporal feature engineering model that learns our hour-based, day-based, weekly, and seasonal features in terms of charging station data. Our suggested ensemble method will be a combination of gradient boosting machines (XGBoost, LightGBM, CatBoost) and classical algorithms (Linear Regression, Ridge, Lasso, Decision Trees), ensemble models (Random Forest, AdaBoost, Gradient Boosting), and neural networks (KNN, SVR, Neural Networks). By means of intensive testing on real world charging station data consisting of 1,320 charging sessions on various vehicle models, locations and user types, we can show that ensemble learning methods are much more effective than traditional statistical techniques. Results can be used as practical intelligence on how to maximize the use of charging infrastructure, dynamic prices, grid balancing loads, and user experience improvements within the fastchanging EV ecosystem.