A Deep Learning Based Multi-Block Hybrid Model for Bike-Sharing Supply-Demand Prediction

As a new type of short distance commuting, the station-free sharing bike effectively alleviates urban traffic congestion. Thus, they are deployed in a large scale in many cities. However, various complex factors, including spatial, temporal, and other external information, result in serious imbalance of supply and demand between regions, which makes accurate prediction a challenging issue. In this study, our primary objective is to accurately forecast supply and demand by leveraging multi-source datasets. Based on the visual analysis about spatial-temporal characteristics of GPS data in Shanghai, we presented the innovative methods of Area of Interest grading and Traffic Analysis Zone division of bike-sharing, and revealed the distribution characteristics of sharing bike trips. A multi-block hybrid model where three blocks were separately modeled according to different data types was proposed. Moreover, eight state-of-the-art models and two variant models were developed as benchmarks to compare and evaluate the proposed model. The results suggested that MBH outperforms ten baselines with the highest accuracy. In addition, we conducted practical application of prediction results to validate that the proposed model could provide effective information for scheduling and rebalancing of bike-sharing system.