Transfer Learning-Based Framework Enhanced by Deep Generative Model for Cold-Start Forecasting of Residential EV Charging Behavior

Abstract

Reliable smart charging requires forecasting the charging behavior of EVs. Deep learning algorithms could present a solution. However, deep neural networks (DNNs) require a large number of samples for training. Although there are available datasets for EV charging, such data is not available for newly committed EVs with limited historical charging records. Therefore, forecasting the charging behavior of a new EV owner suffers from the cold-start forecast problem. This paper aims to forecast two essential parameters of charging events for new EVs: plug-out hour and required energy. To this end, a transfer learning-based framework is proposed to address the cold-start forecast. In the proposed framework, we freeze the middle layers of the pre-trained DNNs and make a shortcut between the output layer and input layer during the backpropagation. Moreover, this paper proposes to use a deep generative model, i.e., generative adversarial network (GAN), to improve the accuracy of forecasting. Plug-out hours of charging events are primarily forecasted by the proposed framework and then the plug-out hour is used as an auxiliary feature to forecast the required energy. Numerical results prove the effectiveness of the proposed framework where forecasting results are improved by more than 31% for the plug-out hour and 34% for the required energy compared to alternative machine learning and deep learning algorithms.