The spatiotemporal distribution of high-risk areas in urban social security management is difficult to accurately quantify and predict. This study proposes a spatiotemporal graph convolutional attention network (STGCN-ATT) model that integrates urban POI data. By calculating the kernel density of multiple types of POIs and generating functional semantic vectors, multi-source features reflecting the urban functional texture are constructed; This article further designs an encoder that includes spatiotemporal graph convolution, GRU temporal modeling, and spatiotemporal attention mechanism to simultaneously capture spatial dependencies, temporal dynamics, and key spatiotemporal patterns. The experiment is based on the annual criminal police situation and POI data of a mega city in China. The results showed that the RMSE predicted by the model within 24 hours was 0.317, and the spatial overlap remained stable in the range of 76.8% -79.1%. It could clearly distinguish the risk evolution law of "nighttime gathering" in entertainment areas and "morning and evening peaks" in transportation hubs, indicating that integrating POI features can effectively improve the ability to characterize the spatiotemporal heterogeneity of risks and provide reliable decision-making basis for dynamic police force allocation.
Jian Wang (Thu,) studied this question.