A Spatio-temporal Network for Demand Prediction of Electric Vehicle Sharing Systems

With the improvement of environmental awareness, one-way electric vehicle sharing systems with stations are gradually known. Vehicle rebalance and station expansion are the two things that system operators care most about. In this paper, we study the problem of forecasting travel demand, which can be used to infer the place to deploy new station and provide suggestions for vehicles scheduling. As an attempt to make use of both spatial and temporal features, we propose a spatio-temporal network based on Convolutional Long Short-Term Memory (ConvLSTM) to predict traveling demand in an area without historical travel records. Convolution networks make sure that when demand in an area is predicted, geographical features of its neighborhoods will also be considered. With LSTM, demand will be treated as time series. Therefore, temporal associations are also considered. Our network improves the prediction accuracy which has been corroborated through experiments on real-life data conducted with other regression methods. In addition, it can be observed from prediction curves, that trend of curve predicted by our method is closer to the real curve. Our work provides a travel demand predicting solution with commercial potential that helps to make business decisions.