Deep Learning-Based PM2.5 Air Quality Prediction: A Comparative Analysis of LSTM, CNN-LSTM, and Hybrid BiLSTM-CNN Architectures Across Different Pollution Regimes

Mishra, Rishabh (2025) Deep Learning-Based PM2.5 Air Quality Prediction: A Comparative Analysis of LSTM, CNN-LSTM, and Hybrid BiLSTM-CNN Architectures Across Different Pollution Regimes. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (559kB) | Preview

Abstract

This study conducted in order to design and checked the performance of three deep learning models (LSTM, CNN- LSTM, and Hybrid Bi-LSTM-CNN) for PM2.5 weather quality prediction in three different areas: Dublin (average PM2.5 level: 3.69 µg/m3 ), Lahore (182.64 µg/m3 ), and Delhi (176.46 µg/m3 ). This project utilized a very precise temporal validation technique with 60-20-20 train-validation-test splitting and systematic hyperparameter optimization using the hyperband algorithm from keras tuner with 10 trials per model (MAX TRIALS = 10). A combined criteria feature selection technique founded on correlation analysis and Variance Inflation factor resulted in the exclusion of PM10 from Lahore and Delhi, narrowing down the feature list to seven factors: temperature, humidity, wind speed, CO, NO2, O3, and SO2.

The extent of the model’s performance was found to differ from one city to another. In cities with a high level of pollution, Lahore stood out for being the most accurate in predicting the pollution levels. The LSTM model had Lahore as the top-performing city with an R 2 of 0.9303 and an MAE of 31.75 µg/m3 , while Delhi and its LSTM model were right next to it with an R2 of 0.909 and an MAE of 29.54 µg/m3 . However, the case of low- pollution Dublin required more attention from the model’s side. With an R 2 equal to 0.28 and an MAE equal to2.18 µg/m3 , the LSTM model was not much better than the persistence baseline, which had much better numbers. That means the very basic pollution dynamics that have already reached very low levels, and the model could not capture them.

The cross-city combined model showed impressive generalization capability. The overall best performance was achieved by the Hybrid Bi-LSTM-CNN architecture (R2 = 0.92, MAE = 22.70 µg/m3 , RMSE = 39.44 µg/m3 ) over 73,513 training samples that covered three pollution regimes. According to the feature importance analysis, Permutation-based, CO (37.6% relative importance), and NO2 (31.8%) were the main predictors, while meteorological variables (humidity, temperature, wind speed) had a very small impact (combined: 3.3%).

Item Type:	Thesis (Masters)
Supervisors:	Name Email Nagahamulla, Harshani UNSPECIFIED
Uncontrolled Keywords:	PM2.5 forecast; neural network; hybrid Bi-LSTM-CNN; cross-city prediction; factor relevance
Subjects:	G Geography. Anthropology. Recreation > GE Environmental Sciences Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning G Geography. Anthropology. Recreation > GE Environmental Sciences > Earth sciences > Atmospheric science > Meteorology > Weather
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	08 Sep 2026 10:31
Last Modified:	08 Sep 2026 10:31
URI:	https://norma.ncirl.ie/id/eprint/9885

Actions (login required)

View Item