Research on optimization of dynamic bike-sharing repositioning and collection based on spatio-temporal demand prediction

As a vital component of urban transportation systems, the bike-sharing system operates on a time-based billing mode and offers “point-to-point, door-to-door” rental services, enabling users to conveniently pick up and drop off bicycles at their desired locations. At present, bike-sharing platforms encounter operational deficiencies, including inaccurate demand prediction, suboptimal bicycle allocation, and delayed collection of faulty bicycles, resulting in a significant mismatch between supply and demand. To address these challenges, this study investigates a spatio-temporal demand prediction method incorporating multi-task learning and a dynamic shared-bikes repositioning and collection approach. Firstly, a multi-gate mixture-of-experts with a bidirectional long short-term memory network is employed to jointly predict the pick-up and drop-off demands by considering the correlation between the pick-up and drop-off demands corresponding to stations. To alleviate the dependency on long time sequences, an attention mechanism is introduced to enhance the attention given to the crucial information. Furthermore, a collaborative optimization model is proposed to address the dynamic repositioning and faulty bicycle collection in the bike-sharing system, which accounts for charging decisions and mileage constraints associated with vehicles. To meet the time-sensitive requirement of large-scale dynamic repositioning management, a simulated annealing-based adaptive large neighborhood search is customized to solve the model. Finally, a comprehensive case study utilizing bike-sharing data from the New York City Citi Bike is conducted to validate the effectiveness of the proposed approach across various performance metrics: Predictive accuracy, computational efficiency, and operating costs.