A machine learning approach for ozone forecasting and its application for Kennewick, WA

Kai Fan, Brian K. Lamb, Ranil Dhammapala, Ryan Lamastro, Yunha Lee 

Chemical transport models (CTM) are widely used for air quality modeling, but these models miss forecasting some air pollution events, and require a lot of computational power. In Kennewick, WA, elevated O3 episodes can occur during the summer and early fall, but the CTM-based operational forecasting system (AIRPACT) struggles to capture them. This research used the 2015 – 2018 historical archives from the Weather Research and Forecasting (WRF) meteorological model forecasts produced daily by the University of Washington, and O3 observation data at Kennewick to train two machine learning modeling frameworks, ML1 and ML2 for a reliable forecasting system. ML1 used the random forest (RF) classifier and multiple linear regression (MLR) models, and ML2 used a two-phase RF regression model with best-fit weighting factors. Since April 2019, the ML modeling frameworks have been used to produce daily 72-hour O3 forecasts and have provided the forecasts via the web for the agency and public use. For the peak O3 days, AIRPACT showed a large variation, while ML2 underpredicted and ML1 performed the best. In the future, this ML forecast system will be applied to other locations within the Pacific Northwest.