Prediction of daily maximum ozone concentrations from meteorological conditions using a two-stage neural network

Meteorological conditions exert large impacts on ozone concentrations. Predicting ozone concentrations from meteorological conditions is a very important issue in air pollution. A self-organizing map (SOM) neural network is suitable for clustering data because of its visualization property. A multilayer perceptron (MLP) neural network was widely used recently in predicting air pollutant concentrations since MLP can capture the complex nonlinear concentration–meteorology relationship. In this work, a two-stage neural network (model I) was developed and used to predict ozone concentrations from meteorological conditions. The two-stage neural network first utilized an unsupervised neural network (two-level clustering approach: SOM followed by K-means clustering) to cluster meteorological conditions into different meteorological regimes. It was found that ozone concentrations within most meteorological regimes exhibited significantly different concentration characteristics. Then a supervised MLP neural network was used to simulate the nonlinear ozone-meteorology relationship within each meteorological regime. The results showed that meteorological conditions can explain at least 60% variance of ozone concentrations by the two-stage neural network. In addition, three other models (model II: multiple linear regressions (MLR), model III: two-level clustering approach followed by MLR and model IV: MLP) were also utilized to predict ozone concentrations, and were compared with model I. The sequence of predicted accuracy was model I > model IV > model III > model II, suggesting that the two-stage neural network had the best prediction performance among the four models and can elucidate better the dependence of ozone on meteorology than other models.