Shared e-scooters have become a popular mode of urban transportation since their introduction in 2017, offering flexible, convenient, and low-carbon travel options for short-distance trips. However, their rapid growth has created significant urban management challenges, particularly the problem of disorganized parking. Cities newly introducing e-scooter services face a dilemma: scientifically planning parking stations traditionally requires extensive local data collection, which is both time-consuming and costly.

To address this challenge, a research team from Shanghai University and Tsinghua University developed a cross-city transfer learning framework that enables new cities to plan e-scooter parking infrastructure by leveraging operational experience from established markets. The framework utilizes shared e-scooter data from 25 European cities, combined with multi-source open data including Points of Interest (POI) from OpenStreetMap and population distribution data. Each city is discretized into hexagonal grids using Uber’s H3 library, and rich spatial feature vectors are constructed incorporating neighborhood influence modeling.

A key innovation of this work is the multi-dimensional city similarity matching strategy. Rather than relying on a single factor such as geographic proximity, the framework evaluates inter-city resemblance across three dimensions: socio-economic characteristics, POI distribution features, and spatial structure features. Through grid-search optimization, the researchers determined that spatial structure similarity carries the greatest weight in effective knowledge transfer, followed by economic-population similarity.

The empirical results reveal that economic similarity plays a decisive role in transfer learning performance. The super-high-income city group achieved the highest average F1-score of 0.801, while the Central European geographic group also demonstrated strong transferability with an average F1-score of 0.75. Notably, the TOP-3 similarity matching strategy outperformed conventional grouping methods based on geography, population, or economy alone, demonstrating superior stability and generalization across diverse urban contexts. The study also found that expanding beyond the three most similar source cities actually degraded prediction accuracy, indicating that precise similarity matching is more effective than simply accumulating more source data.

The framework provides city planners and shared-mobility operators with a practical, data-driven tool for deploying e-scooter parking infrastructure efficiently in new markets, significantly reducing the reliance on costly local data collection.

The work titled “Cross-city transfer learning for optimal e-scooter parking station deployment: Evidence from 25 European cities” was published in ENGINEERING Management (published on Mar. 26, 2026).
DOI: 10.1007/s42524-026-5157-8