Evaluating the predictability of PM10 grades in Seoul, Korea using a neural network model based on synoptic patterns

As of November 2014, the Korean Ministry of Environment (KME) has been forecasting the concentration of particulate matter with diameters ≤ 10 μm (PM10) classified into four grades: low (PM10 ≤ 30 μg m−3), moderate (30 < PM10 ≤ 80 μg m−3), high (80 < PM10 ≤ 150 μg m−3), and very high (PM10 > 150 μg m−3). The KME operational center generates PM10 forecasts using statistical and chemistry-transport models, but the overall performance and the hit rate for the four PM10 grades has not previously been evaluated. To provide a statistical reference for the current air quality forecasting system, we have developed a neural network model based on the synoptic patterns of several meteorological fields such as geopotential height, air temperature, relative humidity, and wind. Hindcast of the four PM10 grades in Seoul, Korea was performed for the cold seasons (October–March) of 2001–2014 when the high and very high PM10 grades are frequently observed. Because synoptic patterns of the meteorological fields are distinctive for each PM10 grade, these fields were adopted and quantified as predictors in the form of cosine similarities to train the neural network model. Using these predictors in conjunction with the PM10 concentration in Seoul from the day before prediction as an additional predictor, an overall hit rate of 69% was achieved; the hit rates for the low, moderate, high, and very high PM10 grades were 33%, 83%, 45%, and 33%, respectively. Our findings also suggest that the synoptic patterns of meteorological variables are reliable predictors for the identification of the favorable conditions for each PM10 grade, as well as for the transboundary transport of PM10 from China. This evaluation of PM10 predictability can be reliably used as a statistical reference and further, complement to the current air quality forecasting system.