Analysis of neural networks for air quality forecasting in populated areas

In the modern world, steps must be taken to counter the harmful effects of air pollution on public health and the environment, including asthma, bronchitis, and chronic obstructive pulmonary disease (COPD). In this regard, machine learning has proved most promising in predicting the Air Quality Index (AQI). This study applies AQI forecasting to Chennai, Tamil Nadu, India, analyzing the temporal variability of PM2.5, PM10, SO2 and NO2 using Tamil Nadu Pollution Control Board (TNPCB) data from Traffic, Residential and Commercial areas (January 2021 - November 2024). Time Series models like LSTM, GRU, SVM and N-BEATS (Neural Basis Expansion Analysis for Time Series) with ADAM Optimization and Bayesian Optimization. N-BEATS performed best, with an MAE of 0.3714 and an MSE of 0.2068 on average, whereas LSTM performed worst, with an MAE of 0.4043 and an MSE of 0.2281 on average. Thus, N-BEATS, combined with these techniques, demonstrates potential for short-term AQI forecasting in Chennai, with performance varying by season and pollutant composition. This study provides a clear, season-specific comparison of four forecasting methods, along with a reproducible tuning process and deployment information, linking model performance to pollutant composition and temporal variability and supports Sustainable Development Goal 3: Good Health and Well-Being.