Variation of ANN Model's Predictive Performance Concerning Short-term (<24 hrs) Concentrations with Prediction Lagging Time

In this study, neural network models (NNMs) were examined as alternatives to dispersion models in predicting the short-term $SO_2$ concentrations in a coastal area because the performances of dispersion models in coastal areas have been found to be unsatisfactory. The NNMs were constructed for various combinations of averaging time and prediction time in advance by using the historical data of meteorological parameters and $SO_2$ concentrations in 2002 in the coastal area of Boryeung, Korea. The NNMs were able to make much more accurate predictions of 1 hr $SO_2$ concentrations at ground level in the morning in coastal area than the atmospheric dispersion models such as fumigation models, ADMS3 and ISCST3 for identical conditions of atmospheric stability, area, and weather. Even when predictions of 24-h $SO_2$ concentrations were made 24 hours in advance, the predictions and measurements were in good accordance(correlation coefficient=0.65 for n=216). This accordance level could be improved by appropriate expansion of training parameters. Thus it may be concluded that the NNMs can be successfully used to predict short-term ground level concentrations averaged over time less than 24 hours even in complex terrain. The prediction performance of ANN models tends to improve as the prediction lagging time approaches the concentration averaging time, but to become worse as the lagging time departs from the averaging time.