Date of Award

8-2024

Document Type

Thesis

Degree Name

Master of Science (MS)

College/School

College of Science and Mathematics

Department/Program

School of Computing

Thesis Sponsor/Dissertation Chair/Project Chair

Eric Forgoston

Committee Member

Pankaj Lal

Committee Member

Aparna Varde

Abstract

Fine Particulate Matter (PM₂.₅) is defined as microscopic particles suspended in the atmosphere with a diameter of ≤ 2.5 microns. PM₂.₅ pollution can originate from a plethora of anthropogenic sources, such as industrial processes, heat and power generation, vehicle emissions, as well as environmental sources, such as wildfires. Recent research has linked exposure to both long-term and short-term PM₂.₅ pollution to adverse health effects, including cardiovascular and respiratory problems, lung cancer and increase in premature mortality rates, especially in vulnerable communities. As our understanding of public health hazards posed by PM₂.₅ pollution deepens, a renewed urgency is brought to developing PM₂.₅ concentration forecasting tools. In this thesis, we consider two independent approaches to the predictive modeling of PM₂.₅ pollution. The first approach involves the development of a machine learning-based Bidirectional Long Short-Term Memory neural network enhanced by Empirical Mode Decomposition time series processing. The second approach involves the development of a physics-based advection-diffusion model. Our methodology was evaluated using publicly available PM₂.₅ concentration readings reported by several air quality monitoring stations in New Jersey. We observed that the one-dimensional implementation of the advection-diffusion model was efficient in approximating the spread of a polluting event from a Manhattan, New York location to Elizabeth, New Jersey and further to Flemington, New Jersey. The mean absolute errors for both locations were found to be within 2.2 μg/m3 and mean percent errors within 8.8% averaged through the nine-hour time frame. The machine learning model, EMD-BiLSTM, was used to predict next hour PM₂.₅ concentrations and showed smaller errors (1.85 μg/m3 MAE) than a benchmark BiLSTM model, especially on higher concentration peaks. Unlike regular BiLSTM, the EMD-BiLSTM model successfully approximated a concentration spike in the testing set 30 μg/m3 higher than the maximum value in the training set.

File Format

PDF

Available for download on Saturday, November 14, 2026

Share

COinS