Smart Air: A Spatiotemporal Attention Based Deep Learning Approach for Accurate PM_2.5 and PM₁₀ Forecasting

With the tremendous growth of industry and automobiles in recent decades, air pollution has emerged as a significant concern in India. Air pollution has reached life-threatening levels in numerous countries worldwide. Several major Indian metropolises, including Delhi, Mumbai, Bengaluru, Chennai, and Kolkata, consistently rank among the most polluted cities worldwide. Air pollution has a serious influence on public health. Substantial research is being undertaken to manage air pollution and anticipate the behaviour of particulate matter, particularly PM_2.5 and PM₁₀. PM₁₀ represents particulate matter with an aerodynamic diameter less than 10 μm and PM_2.5, represents particulate matter with an aerodynamic diameter less than 2.5 μm is causing major health hazards such as cardio-vascular and respiratory disorders. Deep learning algorithms are successfully utilised to forecast results in various domain. This study introduces the Decoder-only Air Transformer Network (D-ATN) model, designed to estimate daily concentrations of PM_2.5, PM₁₀, and AQI (Air Quality Index) in five key Indian cities (Delhi, Mumbai, Bengaluru, Chennai, and Kolkata). This estimation is achieved through the fusion of historical pollutant data with meteorological variables. The effectiveness of the proposed D-ATN model has also been evaluated against other classical and advanced models such as ARIMA, XGBoost, RF, LSTM, and Transformer. This assessment can aid government agencies in the development of better pollution control policies.