Predicting Grid-Level Urban Population Dynamics Based on POI and Road Networks: A Random Forest Approach and SHAP Evaluation

Predicting urban population dynamics encounters challenges like intricate spatiotemporal variations, privacy concerns tied to user data, and a lack of model interpretability, all impeding sustainable urban development. To address these issues, this paper introduces a method for hourly urban population prediction based on Points of Interest (POI) and road network data. By modeling POIs and road networks, the approach reconstructs grid-level counts of accessible POIs across different time intervals. These features are key inputs for a Random Forest algorithm, demonstrating high precision and consistent stability in forecasting population dynamics at an hourly level. Additionally, SHAP analysis assesses the spatiotemporal significance of different POI types and their accessibility concerning the prediction model. Results highlight the substantial influence of office spaces, residential areas, and shopping facilities on the population prediction model, with local POIs and those accessible within a 60-min radius making noteworthy contributions. This research establishes a foundation for urban population forecasting, providing valuable insights for traffic planning, management, and future urban development strategies.