Uncertainty-aware probabilistic travel demand forecasting for Mobility-on-Demand services

Abstract

Past work has primarily focused on improving the accuracy of travel demand forecasting, often overlooking the inherent uncertainty in such predictions. To this end, we propose an innovative nonparametric uncertainty-aware probabilistic framework for travel demand forecasting in Mobilityon-Demand services. The framework employs a spatiotemporal graph neural network to learn and extract features from city-level travel demand data. These features are then processed through the designed variational autoencoder, which compresses the information and applies resampling and decoding operations to generate forecast samples. A kernel density estimation transforms these samples into a predictive distribution, producing accurate, confident, and well-calibrated predictions. Comprehensive experiments on a real-world dataset, evaluated across multiple metrics and benchmarked against four baseline models, demonstrate the superior performance of the proposed model in both point forecasting and probabilistic forecasting. This framework offers a robust and extensible tool for quantifying uncertainty in future travel demand.