Machine Learning-integrated Optimization Approach For PM2.5 Air Pollution Control
Date of Award
5-2026
Degree Name
Master of Science in Engineering
Department
Industrial and Entrepreneurial Engineering and Engineering Management
First Advisor
Ilgin Acar, Ph.D.
Second Advisor
Lee Wells, Ph.D.
Third Advisor
Sang Kang, Ph.D.
Keywords
PM2.5 forecasting, machine learning ensemble, linear programming, emission control optimization, cost–stringency trade-off, wildfire smoke, Wayne County Michigan
Access Setting
Masters Thesis-Abstract Only
Restricted to Campus until
5-1-2036
Abstract
Fine particulate matter (particles 2.5 micrometers or smaller in aerodynamic diameter PM2.5) is among the most serious environmental health risks in the United States, with prolonged exposure linked to cardiovascular disease, respiratory illness, and premature mortality. Current emission control systems are largely reactive: violations surface only after exceedance events occur, at which point the response involves penalties, emergency retrofits, and health costs that earlier action could have avoided.
This thesis proposes a two-stage decision-support approach that pairs machine learning-based PM2.5 forecasting with linear programming optimization to determine minimum-cost emission control strategies before violations occur. The core argument is economic. Near-future forecasts (e.g., the following 24 hours) give factory managers and regulators time to schedule operational curtailments and adjust scrubber loads in advance, rather than scrambling after the fact.
The first stage runs a grid search over 876 weighted ensemble combinations of seven machine learning architectures. LR(0.85) + LGB(0.15) emerges as the best-performing model on EPA AQS hourly data for Wayne County, Michigan (August 2018 – August 2025), with a validation RMSE of 8.776 μg/m³, a validation MAE of 3.947 μg/m³, and an independent test MAE of 5.604 μg/m³.
The second stage feeds those 24-hour-ahead forecasts into a single-objective linear program tested across five compliance thresholds, from the EPA NAAQS standard (35 μg/m³) down to the WHO guideline (15 μg/m³), over a 242-day test period. Tightening to the WHO limit requires 382 control-hours at roughly 19 times the cost of the 25 μg/m³ scenario.
Recommended Citation
Demirci, Yavuzalp, "Machine Learning-integrated Optimization Approach For PM2.5 Air Pollution Control" (2026). Masters Theses. 5505.
https://scholarworks.wmich.edu/masters_theses/5505