Air Quality Prediction

Machine learning study for predicting air quality using historical Indian air-quality data from 2015–2020.

Air pollution, and PM2.5 concentration in particular, is a major public health concern in urban India, and reliable prediction from historical monitoring data is central to prevention and pollution-control planning.

Predict PM2.5 concentration from the "Air Quality Data in India (2015–2020)" dataset (Kaggle) and compare a classical machine learning model, Random Forest, against a deep learning model, LSTM, to see which better captures pollution trends over time.

• •Team research project at KIIT University (3 student authors), supervised by Ms. Prachi Vijayeeta, Soft Computing
• •Dataset covering PM2.5, PM10, NO2, CO and O3 concentrations across several Indian cities, 2015–2020
• •Missing values replaced by the mean, outliers removed via the interquartile range method, MinMax normalization applied
• •Random Forest trained as the classical ensemble baseline
• •LSTM trained as the deep-learning, time-series approach
• •Models compared on RMSE and MAE

LSTM outperformed Random Forest on both metrics (RMSE 14.50 vs 15.23, MAE 9.85 vs 10.45), confirming it captures long-term temporal pollution trends better. Random Forest stayed competitive as a faster, more interpretable option, especially for quick, actionable predictions. Main limitations: no meteorological data (temperature, humidity, wind) in the dataset, and results were trained and validated on India-specific data only, not tested for generalization to other regions.

The two models trade off differently: LSTM is smoother and better at long-term trends but underestimates sharp pollution spikes, while Random Forest reacts faster to abrupt changes but is noisier. A hybrid approach (Random Forest for immediate estimates, LSTM for longer-horizon forecasts) looks like the more promising direction than picking one model outright.