A service demand forecasting model for one-way electric car-sharing systems combining long short-term memory networks with Granger causality test

As an emerging green and sustainable transportation system, electric car-sharing helps reduce environmental pollution, carbon emissions, and traffic congestion in cities. An accurate service demand forecasting model is essential for system operators to improve vehicle relocation efficiency so as to satisfy uncertain user trip demand. Therefore, this paper studies relevant indicators affecting car-sharing service demand at the operational level and constructs a micro demand forecasting model for one-way electric car-sharing systems. Taking the city of Shanghai as a case study, firstly, 11 raw indicators are hypothesized to be influential when predicting the order volume of car-sharing service and four 11-variate vector autoregression systems under different space-time conditions are established and estimated. Then significant indicators are achieved from 11 raw indicators by conducting Granger causality test. After that, the exact effect mechanism between each indicator and order volume is analyzed by impulse function and the contribution of each indicator to the error of prediction is calculated by variance decomposition. Finally, long short-term memory network is adopted to train historical data of significant indicators to get an accurate prediction of service demand in different stations and time periods. Major findings of this research are (1) Indicators including 'Driving distance', 'Weekday', 'Daily highest temperature', 'Daily lowest temperature', and 'Fee per kilometer' have no 'predictive power' for the demand forecasting model; (2) 'Car pick-up interval', Trip time', 'Workday or not', 'Daily weather condition', 'Fee per minute', and 'Order volume' itself are significant indicators but the effect of 'Workday or not' and 'Daily weather condition' are negligible; (3) The forecasting model based on long short-term network has better prediction performance than other forecasting models; (4) It would be better to exclude irrelevant indicators before constructing the forecasting model. (C) 2019 Elsevier Ltd. All rights reserved.