AAAR 37th Annual Conference October 14 - October 18, 2019 Oregon Convention Center Portland, Oregon, USA
Abstract View
PM2.5, PM10, and Ozone Forecasting in Southern California: Determining the Best Forecast Model as a Function of Predicted Meteorology and Emissions
SCOTT A. EPSTEIN, Nico Schulte, Mark Bassett, Elham Baranizadeh, Melissa Sheffer, South Coast Air Quality Management District
Abstract Number: 464 Working Group: Urban Aerosols
Abstract The South Coast Air Quality Management District is responsible for issuing a daily air quality forecast of PM2.5, PM10, ozone, carbon monoxide, and nitrogen dioxide for the counties of Los Angeles, Orange, Riverside, and San Bernardino—a region encompassing approximately 18 million residents and over 32,000 square miles. Daily forecasts are tailored to 45 individual areas for the following day and the day after that, with the addition of hourly air quality index predictions expected in summer of 2019. A combination of statistical and gridded chemical transport forecast models are used to make predictions for each pollutant on a 30 and 64 hour time-horizon. The models have prediction errors that are strongly related to location, emissions, and meteorology; differences in model structure and behavior lead to conditions where one model produces smaller errors than the others. Traditionally, an experienced forecaster identified the most reliable model for a particular day based on recent model errors.
To automate this task and reduce prediction errors, we developed an algorithm using feature selection methods to predict the root-mean-squared (RMS) prediction error of each model on the prediction day based on model performance on days with similar meteorology and emissions. A weighted distance-based regression is used to identify days in the previous year that are most similar to the prediction day. The algorithm then selects the model with the smallest RMS error at each of the 45 forecast areas for PM2.5, PM10, and ozone. Implementation of this algorithm leads to reduced RMS prediction error in all forecast regions throughout the entire year. The algorithm will be an important part of our plan to integrate additional forecast models into the forecasting system to further improve forecast accuracy.