Demand Density Forecasting in Mobility-on-Demand Systems Through Recurrent Mixture Density Networks

Demand forecasting is one of the essential issues in the Mobility-on-Demand (MoD) systems. Most deep learning-based models address this issue by point prediction, which neglects stochasticity in the forecasting result. In this paper, we propose a novel deep learning-based model, the tailored recurrent mixture density network (RMDN), to forecast the demand density in the MoD systems. The tailored RMDN integrates the time-dependent sequence of historical mobility demand information with the temporal features to forecast the short-term demand density. Unlike the point prediction in the existing work, the forecasting result by tailored RMDN benefits from the predicted parameters. Namely, the predicted weights, means, and variances can be utilized to parameterize a Gaussian mixture model that can denote any shape of demand distribution. We then conduct a group of numerical experiments on the New York yellow taxi trip record data. The validation results show that by integrating the temporal features in MoD data, the tailored RMDN model can significantly improve the demand density forecasting results compared to the statistical time-series prediction model ARIMA. In particular, the tailored RMDN model outperforms the ARIMA model up to 51.3% in terms of the log-likelihood values. In addition, we observe that the tailored RMDN is tremendously superior to ARIMA in handling the high volatility in MoD demand.