A Data-driven Rebalancing Optimization System for Bike-sharing Systems

Description

The bike-sharing systems have been widely deployed for first- and last-mile transportation in urban cities. Due to the geographical and temporal imbalance of bike demand, bikes need to be reallocated among stations to maintain a high service level of the system. There are two challenges for optimizing the bike rebalancing operations. (1) to accurately predict bike pickup and dropoff demand at the station level considering spatial-temporal demand substitution, and (2) to efficiently optimize the routing of multiple rebalancing vehicles for large-scale bike-sharing systems with the existence of outlier and substitute stations.This project proposes an end-to-end solution to a data-driven rebalancing optimization system for bike-sharing systems considering spatial-temporal demand substitution to address the challenge mentioned above. First, we propose a graph convolution network (GCN) to capture the bike inventory distribution in a station transition network as a topological structure for spatial demand substitution modeling. A sequence-to-sequence (Seq2seq) model with Long-short-term-memory (LSTM) unit is used to model the temporal substitution dynamics for accurate prediction. To address the computational challenge of optimizing the routing problem of rebalancing vehicles, we propose a splitting-clustering-aggregation algorithm that decomposes the multi-vehicle routing problem into small and tractable single-vehicle routing problems. The splitting-aggregation feature also allows a single station to be covered by multiple vehicles when its rebalancing target exceeds the vehicle capacity. Finally, extensive numerical experiments using real data (from New York City Citi Bike in Stage I, II, III system expansions and Beijing Mobike), will be conducted to study the operational impact of the spatial-temporal demand substitution on bike demand prediction accuracy and the performance of the proposed data-driven optimization system.This project will contribute to the literature by incorporating multi-source bike system data, meteorology conditions, and the spatial-temporal dynamics of demand substitution into a deep learning-based paradigm to enhance the bike demand prediction accuracy at the station level. The data-driven optimization system will provide a proactive prediction-optimization framework that integrates the demand prediction and rebalancing optimization under the heterogeneity of the environment. As such, we can solve large-scale rebalancing problems efficiently and effectively. We will provide an alternative data-driven network decomposition approach to the traditional mathematical decomposition for large-scale vehicle routing optimization problems.