以深度學習探討共享機車之即時需求預測與調度

Abstract

共享電動機車服務作為新興的共享微型交通工具，逐漸成為實現永續都市的重要手段。然而其營運仍面臨供需在時空上長期失衡的挑戰。為解決此問題，本論文提出結合深度學習與強化學習之方法，實現即時需求預測與動態調度。 本研究基於共享滑板車與電動機車的歷史資料，提出一種改進的時空序圖神經網路架構-SpDCGRU，透過整合都市建成環境因子、天氣因素與週期性等特徵，有效捕捉複雜的時空需求變化。實驗結果顯示，本模型在需求預測準確度上具顯著優勢，於平均絕對誤差較第二佳之基準模型提升4.75%。此外，本研究亦透過模擬實驗的結果突顯出以未滿足需求為調度依據的重要性，相較於既有研究以取車需求為主的策略更具成效。進一步地，為優化即時調度策略，本研究建構一基於Soft Actor-Critic (SAC)的深度強化學習框架，將預訓練的需求預測模型納入模擬環境中，進行迭代學習以考量長期動態變化。該方法能同時指派多輛運送貨車來往於多個指定區域取放車，提升實務應用性。透過引入動作限制以降低動作空間的複雜度並提升學習效率，所提出模型在減少未滿足需求方面，優於人工策略的成效達450%。 綜上所述，本研究推動了共享電動機車與類似服務於即時需求導向的發展，對營運業者與政策制定者在提升服務能力與都市永續等相關方面，給予具體可行的方案與建議。

Shared electric moped scooter (E-moped) services have emerged as a promising mode of shared micro-mobility for sustainable urban transportation. However, their operations are challenged by the persistent spatio-temporal imbalance between supply and demand. To tackle it, this dissertation proposes the real-time demand prediction and dynamic relocation through a series of deep learning (DL) and reinforcement learning (RL) approaches. Based on the shared electric scooter (E-scooter) and E-moped data, this study proposes a novel graph neural network (GNN) architecture, the Sparse Diffusion Convolutional Gated Recurrent Unit (SpDCGRU), which effectively captures complex spatio-temporal dependencies by incorporating built environment, weather, and periodic features. Experimental results demonstrate the superiority of the proposed model in the demand prediction, achieving a 4.75% improvement in mean absolute error (MAE) over the second-best baseline. The findings also underscore the significance of using the unmet demand, rather than pick-up demand, as the basis for relocation strategies. Building on this foundation, a Soft Actor-Critic (SAC) framework is developed to optimize real-time relocation, incorporating long-term demand dynamics with pre-trained prediction models within an iterative simulation environment. This method highlights its applicability in real-world operations by enabling the simultaneous coordination of multiple delivery trucks across multiple designated regions. With the introduction of action constraints to reduce complexity and improve learning efficiency, the proposed model achieves an average of 450% greater performance in reducing unmet demand than the manual strategy. Overall, this study contributes to the advancement of real-time, demand-responsive strategies in shared E-moped and analogous mobility services. The findings provide actionable insights for service operators and city policymakers seeking to improve operational reliability and promote urban sustainability.