Outlier Removal for Improving the Accuracy of Electric Vehicle Behavior Prediction

Abstract

The concept of supplying electric power from an electric vehicle (EV) to a power system is called Vehicle to Grid (V2G). It is expected to bring about improvement in stability of the grid. As the demand for renewable energy rises in recent years, the stability of the power grid is simultaneously gathering more attention. It is important to predict EV behavior in V2G systems because if the behavior of EVs can be predicted in advance, it enables us to establish an efficient plan of power distribution. However, the irregular behavior of EVs makes prediction difficult. Prediction of EV cars is performed by extracting patterns from past actions, where unusual sudden actions hinder learning of the prediction model and cause a decrease in prediction accuracy. Therefore, to analyze and remove the irregularity of EVs are necessary in order to stabilize the V2G system. In this paper, we analyzed the behavior of EVs using real-world State of Charge (SoC) data. Then, using the technology of outlier removal, we excluded the irregular behaviors from a dataset. Finally, we attempted improvement of prediction accuracy by using the data without the irregularities. As a result, we successfully improved the accuracy of the prediction of EV behaviors. Also, in the process of behavior prediction, we performed a comprehensive analysis of EV SoC data, and introduced the results.