Online Car-Hailing Dispatch: Deep Supply-Demand Gap Forecast on Spark

摘要

For online car-hailing dispatch, we have presented Long Short-Term Memory neural networks (called LSTM) to forecast supply-demand gap. It is a new creative thinking to apply deep networks to model gap volatility, incorporating weather information, traffic condition and point of interest (POI) data, as well as twelve previous returns. As far as we know, this paper is the first attempt to train LSTM with realtime internet data for scheduling online car-hailing. The Spearman rank-order correlation and mutual information have been used to generate data set. The study uses dropout function and shuffle the data set before training to avoid overfitting. The LSTM model has been trained on the distributed memory computing platform as we make primal algorithm distributed. After trained on spark cluster, the model achieves a mean absolute percentage error at 27.3%, much better than autoregressive GARCH and ARIMA benchmarks by at least 36.8%. Our pilot investigation reports applying deep learning model to make the supply-demand forecasting with the data from internet in real time is strong promised, and distributed memory calculation is very effective to improve the training speed and modeling capability of LSTM neural networks in this study.