Meng Xu

Hybrid bike-sharing systems (operating both bikes and ebikes) are emerging worldwide for flexibility and sustainability. Excessive rebalancing, while addressing supply-demand imbalances, causes financial burdens and resource waste. Hence, long-term daily operations (e.g., rebalancing at midnights for upcoming days) have practical economic significance, but they remain underexplored due to difficulties in tracking high within-day and between-day demand variability and the interplay between demand and daily inventory management strategies. Besides, cooperative operation of bikes and ebikes expands solution space, given their inherent demand coupling. This study is the first to explore long-term daily inventory strategies for the hybrid system (ebikes get charged at stations), determining spatial-heterogeneous inventory at midnight to maximize profits. To address the demand variability perception issue, we develop a recurrent-attentive neural process (RANP) model to predict hour-to-hour demand of the upcoming day. The RANP is integrated into a long-term optimization model, referred to as the generative-models-informed Markov decision process (GMI-MDP), where two cooperative intelligent agents determine the bike–ebike allocation based on demand perception and system rewards. A suite of numerical experiments utilizing a real-world dataset from New York is carried out, and various strategies are compared. The proposed method, through online and offline tests, demonstrates superiority over other MDP-based and rule-based methods, achieving a faster solution-seeking process and more stable rewards than MDP with exact upcoming demand. By comparing inventory volatility, we offer insights into managerial operations.

Fitting and modeling spatial-temporal processes are essential research topics in transportation studies. Recently, due to the analytically tractable formulation and good fitting accuracy, the Gaussian Process (GP) is becoming increasingly preferable in fitting transportation processes. However, conventional GPs are inapplicable for large-scale problems due to computational issues. Moreover, they dismiss physics laws when conducting multi-output fitting tasks. This paper proposes a physics regularized multi-output grid Gaussian Process Model (PRMGGP) model for fast and multi-output fitting of large-scale spatial-temporal processes in transportation systems. The PRMGGP model adopts a grid input structure to capture inherent spatial-temporal correlations in the fitting process, takes advantage of the Kronecker algebra to notably accelerate the computation speed, and utilizes a shadow GP to incorporate physics laws of the process. Model training and predictive algorithms are developed coordinately and are tested via synthetic datasets. Furthermore, we apply the proposed model and other widely used machine learning models to fit the numbers of pickups, returns, and idle bikes of a large-scale bike-sharing system based on Citi Bike data from New York City. The results demonstrate the computational efficiency, interpretable results, and the prediction accuracy of the PRMGGP model, which can be a promising methodology for modeling multi-output processes in transportation systems.