A Hybrid GCN–PCA–LSTM Framework for Accurate Spatiotemporal Prediction of PM2.5 Concentrations

Abstract

PM2.5 pollution has become a critical environmental issue affecting air quality and public health. Accurate concentration prediction is of great significance for pollution early warning and control. Considering that PM2.5 concentration variations exhibit both temporal dependence and spatial correlation, along with the presence of redundant features in multi-source data, a spatiotemporal prediction model integrating a Graph Convolutional Network (GCN), Principal Component Analysis (PCA), and a Long Short-Term Memory network (LSTM) is proposed. Specifically, the GCN is first employed to characterize the spatial dependencies among air quality monitoring stations and to extract spatial features. Subsequently, PCA is utilized to reduce the dimensionality of high-dimensional features, thereby eliminating redundant information and improving computational efficiency. Finally, the LSTM is adopted to model temporal sequence features, enabling dynamic prediction of PM2.5 concentrations. Based on air pollutant and meteorological data collected from ten monitoring stations in Hefei from 2018 to 2019, sliding time window samples are constructed to predict PM2.5 concentrations for the next hour. LSTM, GCN, and GCN-LSTM models are selected as baseline methods for comparison. Experimental results demonstrate that the proposed GCN-PCA-LSTM model outperforms the comparative models in terms of RMSE, MAE, and R² metrics, achieving an RMSE of 7.94, an MAE of 5.79, and an R2 of 89.36%. The model is capable of more accurately capturing the variation trends of PM2.5 concentrations. Moreover, it maintains strong fitting performance during periods of high pollution and rapid fluctuations, indicating robust spatiotemporal modeling capability and stability. In summary, the integration of spatial feature extraction, feature dimensionality reduction, and temporal sequence modeling effectively enhances PM2.5 prediction performance, providing a feasible approach for urban air quality forecasting and refined environmental management.