Advisor: Anne Thompson

Abstract:

Currently there are a number of tools available for surface ozone and PM_2.5 forecasting, but not many of these tools provide an inherent uncertainty in forecasts they produce.  Thus, forecasters tend to rely on multiple tools to ascertain forecast uncertainty.  In this talk, a new regional statistical approach is proposed that integrates uncertainty of surface ozone and PM_2.5 prediction. The model combines self-organizing map (SOM), a type of artificial neural network, with a weighted stepwise quadratic regression, which uses meteorological variables as predictors for surface ozone and PM_2.5.  The SOM method allows identification of different meteorological regimes and groups them according to similarity.  In this way, when a regression is developed for a specific regime, data from all the other regimes is also used, with weights based on their similarity to this specific regime.  This approach yields a distinct model for each regime while still taking into account all training cases when building each regime’s model.  In addition, because the prior day pollutant concentration is a powerful but potentially biased predictor, all of the models are trained twice: with and without prior day concentration predictor.  This process doubles the amount of total available models.  For prediction, all of the models are available to generate an ensemble of anywhere from several to a few hundred forecasts for a specific air quality station, allowing forecasters to identify a range of scenarios that are likely to occur.  The model is evaluated based on its surface ozone forecasts for San Joaquin Valley, CA.