Predicting street-level distribution of bike-sharing traffic volume in metro station areas using integrated generative adversarial networks

Shared bicycles are popularized around metropolitan metro stations, accompanied by concentrated mobility flows and built environment characteristics. Traditional predictions of bike-sharing traffic have struggled to analyze the heterogeneity in traffic volume distribution across street sections and the contributing factors, inhibiting targeted urban mobility strategies. To address this problem, we identified the spatial mapping impact of built environments on traffic volume distributions and optimized predictions at the street level using generative adversarial networks (GANs). Through this method, we integrate eight built environment models based on multi-source data from 414 metro station areas in Shanghai, which were clustered into distinct types according to differences in sensitive factors. Integrated GANs achieve higher accuracy (R 2 = 0.881) in predicting travel volume than machine learning models. The results demonstrate: (1) At the street level, bike-sharing traffic concentrates on street sections with dense POI and those intersecting steep housing price transition areas. (2) The heterogeneous correspondence of built environment factors on bike-sharing traffic across different road segments are categorized into five types of metro station areas by pixel error rate ( p ), value-sensitive areas ( p (housing price) = 0.064) with a clustering pattern, land-use-sensitive areas ( p (land use) = 0.063), function-sensitive areas ( p (POI) = 0.065, p (land use) = 0.067) concentrated in the city center, activity-sensitive areas ( p (POI) = 0.058, p (social flow) = 0.060), and mix-sensitive areas influenced by multiple factors. This research realizes efficient street-level predictions of bike-sharing traffic distribution, and helps policy-makers develop well-calibrated traffic management and urban planning policies for different types of metro station areas.