A Spatiotemporal PM2.5 Prediction Model Based on Multi-Layer Clustering Node Partitioning Strategy and Multi-Criteria Node Importance Evaluation System

Accurate spatiotemporal prediction of PM2.5 concentrations provides technical support for urban air pollution warning and prevention. To address the insufficient prediction accuracy and high spatiotemporal training cost in existing approaches to large-scale PM2.5 prediction, this study applies a multi-criteria node importance evaluation system and a hierarchical greedy K-Means clustering algorithm to PM2.5 spatiotemporal modeling. First, air pollutant and meteorological data from 559 monitoring stations in southeastern China are selected for modeling, and feature enhancement is performed based on Improved Seasonal and Trend decomposition using Loess (ISTL). Second, the clustering algorithm divides stations into multiple station clusters. Subsequently, the node importance of each station cluster is calculated, with a topographic factor introduced for station importance modeling, and spatiotemporal prediction uses MCN-NIE-TCNTRANS (Multi-layer Clustering Node Partition and Node Importance Evaluation System-based TCN-Transformer model). Finally, cross-model comparison and ablation experiments are designed to evaluate model performance. Results show that the multi-criteria node importance evaluation system effectively improves the prediction performance of this model. Without a significant increase in spatiotemporal cost, it outperforms the suboptimal model, achieving a 1.92% increase in average R2 and a 3.07% reduction in MAE. This work provides a quantitative assessment and decision support for sustainable urban ecological development.