Comparative Analysis of Machine Learning Techniques for an Hour-Ahead Forecasting of Electric Vehicle States

Abstract

This thesis presents a method for forecasting the state of electric vehicle (EV) batteries (Vehicle-to-grid, Grid-to-vehicle) for the next hour towards leveraging updated data as the grid's auxiliary power source. The aim is to optimize grid power utilization and mitigate concerns such as voltage and frequency variations, power loss, and harmonic distortion. The proposed prediction model incorporates various distributed energy resources, including demand-side management strategies and energy storage devices, to support the grid. Multiple machine learning (ML) techniques, including logistic regression, artificial neural networks (ANN), naive Bayes, K-nearest neighbour (KNN), and Support Vector Machines (SVM), are employed to predict V2G and G2V conditions. The model is refined by including vital features significantly influencing the outcomes and applying feature selection techniques. Furthermore, correlation time and Gaussian correlated noise are employed to assess feature dependency and evaluate the methods' robustness against noise. The accuracy analysis revealed that the ANN technique performed best among the tested ML techniques. It demonstrated superior accuracy in predicting the state of EV batteries. Additionally, feature selection, correlation time, and Gaussian correlated noise enhance the model's performance and evaluate its robustness.