Song, Jia-CherngJia-CherngSongI-YUN HSIEHCHUIN-SHAN CHEN2025-06-172025-06-172025https://www.scopus.com/record/display.uri?eid=2-s2.0-105003564483&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/730102The growing popularity of shared electric moped scooters (E-mopeds) as a sustainable and convenient transport option faces a fundamental challenge: demand often exceeds supply, necessitating efficient real-time relocation strategies. To address this, we propose an unmet demand-driven relocation policy to improve service efficiency and E-moped utilization. This policy leverages recurrent and diffusion convolutional graph neural networks alongside linear programming. Regional E-moped relocation is suggested to be triggered when its demand falls below a threshold parameter. We demonstrate that the unmet demand strategy consistently outperforms the traditional pick-up method. At a threshold of 3, the unmet demand approach results in 286 relocated E-mopeds compared to 100 using the pick-up strategy, emphasizing the importance of our study. These findings assist operators in implementing more efficient strategies and inform policymakers in refining the maximum fleet sizes and parking space allocation.Graph neural networkReal-time relocationShared electric moped scooter serviceSpatio-temporal demand predictionUnmet demand-driven policy[SDGs]SDG7journal article10.1016/j.jum.2025.04.001