Development of a Machine Learning Approach for Local-Scale Ozone Forecasting: Application to Kennewick, WA.

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Chemical transport models (CTMs) are crucial for air quality forecasting but demand significant computational resources and often exhibit biases, leading to missed pollution events.
• Existing CTM-based systems, like AIRPACT for the Pacific Northwest, struggle to accurately predict unhealthy ozone (O3) episodes, particularly in specific locations such as Kennewick, WA.

Purpose of the Study:

• To develop and demonstrate an improved machine learning (ML) based ozone (O3) forecasting system for Kennewick, WA, addressing the limitations of CTMs.
• To evaluate the performance of ML models against a CTM for predicting high O3 events and assess computational efficiency.

Main Methods:

• Utilized 2017-2020 simulated meteorology (from WRF model) and O3 observation data from Kennewick for training.
• Developed a two-model ML system: ML1 (Random Forest classifier and Multiple Linear Regression) for high O3 events, and ML2 (two-phase Random Forest regression) for less elevated O3.
• Employed 10-time, 10-fold, and walk-forward cross-validation to prevent overfitting and ensure robust evaluation.

Main Results:

• ML1 successfully captured 5 out of 10 unhealthy O3 events, significantly outperforming AIRPACT and ML2, which missed all such events.
• ML2 demonstrated better skill in predicting less elevated O3 concentrations.
• The combined ML framework achieved reliable forecasting with substantially reduced computational resources (single processor for minutes vs. >100 processors for hours).

Conclusions:

• The developed ML modeling framework provides a reliable and computationally efficient alternative for O3 forecasting, particularly for capturing unhealthy air pollution events.
• The ML system has been operational since May 2019, providing daily 72-h O3 forecasts to clean air agencies and the public via a web portal.
• This research highlights the potential of ML in enhancing air quality forecasting accuracy and accessibility.