Hybrid Spatio-temporal Deep Learning Framework for Particulate Matter(PM2.5) Concentration Forecasting

With massive urbanization, air pollution has turned out to be a life-threatening factor that requires serious control management. Proper assessment and prediction of outdoor air pollution could significantly warn people about the risk of chronic and acute respiratory diseases including asthma during outdoor exposure. IoT based air quality monitoring through sensors deployed at different locations could be an expedient solution for this. Accurate modeling of the air quality big data collected by the IoT sensors extensively helps in predicting future values for accessing the risk of outdoor exposure. This extraction of knowledge gained could greatly in reducing the deterioration of human health through early warnings and decision-making. In this paper, a hybrid deep learning-based architecture using CNN-LSTM combination is proposed to model Particulate Matter (PM2.5) of a location with the values collected from the IoT air quality sensors. The encoder-decoder based architecture models the time distributed multivariate air pollutant data of 9 locations. The 3D CNN and 1D CNN is exploited to encode both the inter-dependency (spatial autocorrelation) and intra-dependency (heterogeneity) in the spatiotemporal air pollutant data. The CNN based encoder captures all relevant spatiotemporal features for facilitating improved accuracy in predictions. The LSTM learns the temporal dependencies in the encoded air pollutant data for PM2.5 concentration forecasting. Extensive experimentations are performed on real-world IoT City Pulse Pollution dataset. The proposed model is compared with ConvLSTM in terms of root mean square error (RMSE), the mean absolute error (MAE) and the R-squared (R2) and is found to outperform ConvLSTM.