Developing a statistical based approach for predicting local air quality in complex terrain area

A statistical framework has been developed for predicting the next-day air quality for Hong Kong. In this approach, Generalized Additive Models (GAMs) linking air pollutant concentration with meteorological data were first constructed, based on observations during the December 1997 to November 2009 period. GAMs were used for forecasting local air quality with weather predictions from (1) the Global Forecast System (GFS) model; (2) dynamical downscaling of GFS predictions using WRF model (GFS-WRF), and (3) bias corrected GFS-WRF (Improved GFS-WRF). The system was verified by carrying out retrospective daily air quality predictions in this one-year period (December 2009 to November 2010). Even with the uncertainties in weather predictions, it was found that, downscaled weather forecasts from Improved GFS-WRF combined with GAMs give better results than those based on GFS and GFS-WRF alone. The statistical model with Improved GFS-WRF inputs performed well in forecasting both urban and sub-urban pollutant concentrations including respirable suspended particulates, O3 and NO2. The Hit Rate (False Alarm Ratio) for categorical forecasts of events with daily air pollution index (API) over 100 given by Improved GFS-WRF is also higher (or possibly lower) than that using GFS and GFS-WRF only. Further, this paper describes the implementation of Improved GFS-WRF to detect the onset of O3 episodes in advance due to the presence of tropical cyclones, based on categorical evaluation and supported by case studies. Our results indicate that the statistical model can be a useful tool for air quality prediction for urban and sub-urban sites in the complex terrain area, like Hong Kong.