Shared micromobility systems have become an essential part of urban transportation, offering flexible and sustainable solutions for short-distance travel. However, their large-scale deployment poses several operational challenges, including disorganized parking in free-floating systems, imbalances between supply and demand, high fleet management costs, increased safety risks, and limited integration with public transit. This thesis explores how artificial intelligence techniques can address these challenges by proposing data-driven approaches to improve the efficiency and safety of shared micromobility systems. First, the problem of disorganized parking in free-floating systems is addressed using a multi-layer clustering approach, which identifies high-demand areas and guides the placement of virtual stations. These dedicated spaces, deployed without physical infrastructure, help reduce improper parking and improve spatial organization.
To ensure the system operates efficiently, a demand forecasting approach is proposed to address the imbalance between supply and demand. It estimates vehicle demand by location and time period, thereby facilitating fleet redistribution operations. This approach is also designed to be adaptable to other configurations, including systems with physical stations and free-floating systems.
To ensure scalability while reducing communication overhead and latency, an edge-based architecture is introduced, enabling decentralized data processing and faster decision-making. Building on these contributions, operational decisions are optimized in real time using deep reinforcement learning. A first framework is dedicated to fleet management, recommending nearby alternative drop-off zones and involving users in the redistribution process. It integrates predicted demand and vehicle battery levels to maintain vehicle availability while reducing operational costs, limiting manual redistribution operations, and facilitating the aggregation of vehicles with low battery levels. A second framework addresses safety by recommending safer drop-off zones, based on urban factors such as traffic conditions, the presence of bike lanes, intersection density, lighting, and congestion. Finally, to improve first- and last-mile accessibility, this thesis extends the demand prediction framework to intermodal contexts, with the aim of improving vehicle availability in areas connected to public transportation.
It combines local demand for public transportation with spatial and contextual factors that influence travel patterns. A hybrid graph-based deep learning framework is proposed to model these relationships and better represent urban mobility dynamics.
All data inputs are integrated into an end-to-end intelligent management system, enabling real-time, user-centered decision-making. Experiments conducted on real-world urban case studies demonstrate the system’s effectiveness, including the accurate identification of virtual stations, reliable demand forecasting, and improved vehicle allocation through the recommendation of drop-off zones. Overall, this work contributes to the development of more efficient, safer, and more sustainable shared micromobility systems, and provides valuable insights for researchers and urban planners seeking to improve mobility services.