Attention enhanced hybrid model for spatiotemporal short-term forecasting of particulate matter concentrations

With ever-increasing global air pollution levels, researchers are exploring ways to forecast air pollutant concentrations to prevent the adverse effects of air pollution on humans. Powered by the data obtained from air pollution monitoring stations, we now have a chance to build sophisticated models to estimate the future concentration of various air pollutants. Previous researches show that deep learning models perform a better task of capturing the complex spatiotemporal dynamics from the data as compared to traditional statistical models. In this paper, a hybrid deep learning framework -AGCTCN (Attention based Graph Convolution and Temporal Convolution Network), based on spatial attention, graph convolution and temporal convolution is presented for short-term forecasting of particulate matter (PM) levels. Specifically, quarter-hourly PM(10 )and PM2.5 data aggregated from 27 ground air pollution monitoring stations in Delhi, India is used for training the AGCTCN model and our model forecasts the concentrations of these pollutants at these stations for a particular horizon in the future. The forecasts generated by our model are compared and contrasted to those generated by other state-of-the-art models like GC-LSTM and ConvLSTM with respect to four widely used evaluation metrics namely RMSE (root mean squared error), MAE (mean absolute error), correlation coefficient, and R-2 (coefficient of determination). Our model shows more than 80% improvement over the GC-LSTM model in terms of the R-2 metric, performs the best task of estimating the peak PM concentrations and shows variability and correlation close to that of the observed data. A carefully conducted ablation study shows the effectiveness of the individual components of our architecture. The graph convolution layer improves the model accuracy by approximately 34% on average on the PM datasets. Compared to other models, the AGCTCN framework shows the best performance in short-term forecasting of particulate matter concentrations.