Electric Vehicle User Behavior Prediction Using Gaussian Mixture Models and Soft Information

Intelligent scheduling algorithms pave the way to vehicular mobility electrification. Besides enabling charging grid providers to upscale the charging grid infrastructure without costly changes, they enable optimal green or price-scheduled charging. However, scheduling always requires a priori user behavior parameter knowledge, like the arrival time, departure time, and energy demand. Recent comparisons indicate that machine learning-based predictions are more accurate than direct user prediction. Therefore, we exploit a time-series measurement dataset to predict the departure time and energy demand for intelligent scheduling. The collected data belongs to a charging grid used by nurses working at an elderly home in Denmark. Our data analysis reveals a clustered distribution for the departure time versus arrival time and the energy demand versus arrival time. Targeting on arrival prediction, we propose a sub-clustering strategy employing Gaussian mixture models in combination with the expectation-maximization or the variational Bayesian method to learn the departure times and energy demand. We demonstrate that the investigated method inherently determines different soft information metrics enabling us to propose different prediction strategies adaptable to the users' desired reliability degree for charge scheduling. Moreover, we show that our proposed clustering approach outperforms the assessed regression-based machine learning approaches.