Document Type
Theses, Ph.D
Disciplines
Computer Sciences
Abstract
Accurate forecasting of near-surface atmospheric air pollutants such as PM2.5, NO2, SO2, CO, and O3 remains a critical scientific and societal challenge. This difficulty arises from several factors, including nonlinear pollutant dynamics, sparse ground monitoring net works, heterogeneous satellite observations, and strong cross-pollutant interdependencies. Substantial advances have been achieved in temporal deep learning, probabilistic modeling, satellite-based estimation, physics-informed methods, and foundation models. However, these paradigms have largely evolved in isolation. As a result, existing systems are often station-dependent or pollutant-specific and optimized for a single forecasting task. This limits their robustness and generalizability across regions and heterogeneous data regimes.
This thesis addresses this fragmentation by proposing a unified and progressive deep learning framework for monitoring and short-term forecasting of atmospheric air pollu tants. The framework is developed through a sequence of interconnected methodological advances. First, a hierarchical temporal architecture, CombineDeepNet, is introduced. It integrates Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Unit (BiGRU) networks to enhance forecasting stability and mitigate recursive error accumulation at individual monitoring stations. Second, the framework advances beyond independent pollutant prediction by introducing system-level multivariate modeling through a Gaussian-mixture Nested Factorial Variational Autoencoder (NF-VAE). This model captures shared latent structures and represents structured interdependencies among interacting pollutants under regime variability and uncertainty. Third, heterogeneous observational sources, including ground measurements, multi-spectral satellite imagery, and meteorological variables, are systematically integrated through a structured multi source fusion mechanism. This improves spatial generalization and robustness. To address spatial-resolution constraints in satellite data, task-driven super-resolution techniques are iv developed, and their impact on downstream forecasting is evaluated. Fourth, atmospheric physical principles are embedded within neural architectures through physics-integrated learning. This enhances interpretability, stability under distribution shift, and physical consistency in open-system environments. Finally, all components are integrated into a unified foundation-model-based forecasting framework called GPT4AP. This framework employs a pre-trained LLM backbone with parameter-efficient Gaussian LoRA adaptation. It supports long-term forecasting, few-shot learning, and zero-shot transfer within a single scalable multi-task architecture.
Extensive experiments across multiple pollutants and monitoring locations demonstrate consistent improvements in predictive accuracy, stability, and cross-domain generalization compared to strong baselines. Collectively, this work establishes a coherent One-for-All paradigm for air-quality forecasting. This paradigm systematically integrates temporal modeling, probabilistic multivariate learning, multi-source fusion, spatial enhancement, physics-consistent learning, and foundation-model generalization within a unified deep learning framework.
DOI
https://doi.org/10.21427/3z3w-1y42
Recommended Citation
Dey, Prasanjit, "Monitoring and Short-term Forecasting of Atmospheric Air Pollutants Using Deep Neural Networks" (2026). Dissertations. 290.
https://arrow.tudublin.ie/scschcomdis/290
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License.