Self-supervised online learning algorithm for electric vehicle charging station demand and event prediction

Abstract

With the increasing popularity of electric vehicles (EVs), countries are setting up new charging stations to meet up the rising demand. Therefore, accurately forecasting charging demand and charging events is highly significant. Historical data are crucial for developing a quality forecasting model, but countries or locations with recently installed EV stations suffer from data inadequacy. Delayed data accumulation for forecasting model creation impedes EV's optimal operation, and an offline or fixed-sized data-based learning model may not perform optimally due to the future uncertainties of input variables. Therefore, it is required to create an online forecasting model that can learn right from the beginning of the operation of charging stations, forecast, and relearn, when necessary, by considering the impact of input/external variables. For optimal model development, impactful input variables should be chosen online using appropriate feature engineering. In this research, a unique feature engineering considering multi-level correlation with multicollinearity and simultaneous online learning General Regression Neural Network (GRNN) based on has been suggested. Also due to the discrete and asynchronous nature of the charging event a detailed data handling method has been developed to create meaningful time series data. It is interestingly realized that the proposed model outperforms general Artificial Neural Networks (ANN), various sophisticated models, such as Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Bi-LSTM, Gated Recurrent Unit (GRU), and the Deep Neural Network (DNN) model when the appropriate inputs and their delayed variables are used.