Forecasting PM2.5 Concentration in India Using a Cluster Based Hybrid Graph Neural Network Approach

Abstract

Air pollution modeling and forecasting over a national level scale for a country as large as India is a very challenging task due to the large amount of data involved in a limited spatial frequency. Often the air pollution and pollutant dispersion process depend on underlying meteorological conditions. Recently, Graph Neural Networks emerged as an effective deep learning model for discovering spatial patterns for various classification and regression tasks. This study proposes to employ a cluster-based Local Hybrid-Graph Neural Network (HGNN) methodology instead of using a single global Graph Neural Network for monitoring station-wise multi-step PM_2.5 concentration forecasting across India’s states. This methodology respects sudden changes in PM\(_{2.5}\) concentration due to the local meteorological variations. However, the local Hybrid GNN models consist of two parts: a spatio-temporal unit containing a Graph Neural Network layer along with a Gated Recurrent Unit layer to model the influence of wind speed and other meteorological variables on PM_2.5 concentration. The other part is a station wise feature extraction unit to extract station-wise meteorological feature impact on PM_2.5 concentration, along with the temporal dependency between historical records. The results from the two units are fused in step-wise manner for multi-step PM_2.5 forecasting. The proposed methodology was used to develop separate PM_2.5 concentration forecasting models, +24, +48 and +72 hours ahead. Subsequently, a detailed analysis is carried out to unfold the advantages of the proposed methodology. Results demonstrate the proposed models perform better than the state-of-the-art with significantly lesser computation time.