AIR QUALITY FORECASTING USING LSTM MODELS: A STUDY IN DELHI

Abstract

Air pollution has become one of the most critical environmental challenges facing major cities in India, particularly Delhi, where PM2.5 concentrations frequently exceed safe thresholds. Accurate forecasting of PM2.5 levels can play a key role in supporting public health policies, early warnings, and sustainable urban planning. This study presents an end-to-end deep learning approach for forecasting daily PM2.5 concentrations in Delhi using Long Short-Term Memory (LSTM) neural networks. Data were collected from the OpenAQ API, covering a nine-year period from January 2016 to October 2024. The dataset was preprocessed to address missing values, temporal irregularities, and outliers, then resampled to daily averages to capture long-term pollution dynamics. The univariate LSTM model was trained using a 30-day look-back window to predict the next day’s PM2.5 value. Evaluation on unseen test data yielded an RMSE of 31.09 µg/m3 and an MAE of 19.24 µg/m3, demonstrating robust performance in capturing temporal dependencies and short-term fluctuations. The results suggest that LSTM-based models can serve as effective tools for air quality forecasting in data-scarce and highly variable urban environments.