A dynamic clustering and graph neural network-based framework for multiscale pollution source identification and early warning

Aiming at the challenges of low traceability accuracy and disconnection between prediction and diagnosis in air pollution management, this paper proposes a framework that integrates dynamic sequential clustering, multimodal fingerprinting, graph neural network (GRNN) and cross-scale analysis. The dynamic clustering algorithm improves the accuracy of pollution area delineation under complex terrain to 92% through the optimization of the knapsack problem and terrain blocking weights (Jinan case), which solves the problem of insufficient spatial heterogeneity identification of traditional methods; the multimodal fingerprinting framework integrates five types of pollutant concentrations and meteorological parameters to achieve a classification accuracy of 89% in nine types of pollutants, which is an alternative to the high-cost chemical detection; the GRNN model enables a 24-hour future prediction and diagnosis of pollutant pollution through the spatial and temporal map structural modelling. The GRNN model, through spatio-temporal structural modelling, achieves an F1-score of 0.87 (compared with 0.72 for LSTM) for the next 24 hours, and links dynamic clustering to analyse the causes of pollution. The cross-scale decision-making framework reduces Wuhan's emergency response time by 40% through the weighted fusion of three-level architectures (single site → clustered region → city-wide). The system provides an efficient solution for accurate pollution tracing and real-time early warning, and has significant application value in decision-making support for environmental protection departments and industrial source control scenarios.