Comparison of static MLP and dynamic NARX neural networks for forecasting of atmospheric PM10 and SO2 concentrations in an industrial site of Turkey

Abstract

This study aims to compare performances of two static and one dynamic neural networks used for prediction of hourly ambient air quality concentrations in an industrial site of Turkey. Two air pollutants (PM10 and SO2) and three meteorological parameters (ambient air temperature, relative humidity, and wind speed) were used as input variables. The predictions of the dynamic nonlinear autoregressive exogenous (NARX) model were compared with the predictions of the static multilayer perceptron (MLP) neural network model. The results showed that the predictions of the NARX neural network were obviously better than the predictions of MLP networks. The coefficient of determination (R2), index of agreement and efficiency between the observed and predicted air pollutant concentrations by the NARX model were 0.9773, 0.994, and 0.977 for PM10, respectively while the same parameters were 0.9984, ≈1, and ≈1 for SO2. The MBEs (mean bias errors) were also approximately zero for both pollutants that indicate the adequacy of the model. The values of RMSE (root mean squared error) were also fractional as 0.0191 and 0.0087 for both pollutants. The NARX model predicted SO2 concentrations better than PM10 concentrations. In comparison with MLP network structures, NARX network exhibits faster convergence. The model suggested in this study could be used to support and improve air quality management practices.