TEMDI: A Temporal Enhanced Multisource Data Integration model for accurate PM2.5 concentration forecasting

Abstract

Accurate forecasting of PM2.5 concentration is crucial for implementing effective protective measures and mitigating the adverse health impacts of air pollution. To address the complex spatial propagation dynamics and temporal variations of PM2.5, we developed the Temporal Enhanced Multisource Data Integration (TEMDI) model. This innovative approach combines spatial modeling by a Graph Neural Network (GNN) to capture the intricate spatial propagation patterns based on multi-source data fusion, and a novel Time Series Enhancement (TSE) module that includes Ensemble Empirical Mode Decomposition (EEMD), Gated Recurrent Units (GRUs), and a self-attention mechanism to adequately manage the time series’ short-term and long-term trends. Our results demonstrate TEMDI’s superior performance, achieving exceptionally high Probability of Detection (POD) rates of 96.15%, 80.28%, and 71.86% for forecast horizons of 3, 36, and 72 h, respectively. Furthermore, our feature importance analysis reveals that multi-scale features extracted by the EEMD component become increasingly crucial as the prediction horizon extends. The TEMDI model’s ability to provide accurate, reliable PM2.5 forecasts and its enhanced interpretability position it as a valuable tool for guiding environmental policy and management decisions to safeguard public health.