The aim of this research is forecasting the NOx, NO2 and NO concentration levels with different artificial neural networks structures (ANNs) and determining the best ANNs structure for forecasting emissions. For this aim, it was used one learning function and, six different transfer function pairs with three different neuron numbers. The MATLAB software helped constructing ANNs models. In addition, the air pollutants and meteorological factors were used as input parameters simultaneously at the ANNs. The end of the research, NOx, NO and NO2's concentration levels were modelled with high accurate levels. The R2 values of the NOx, NO and NO2 were calculated as 0.998, 0.995 and 0.997, respectively. The best results were obtained from ANNs structures which used Logarithmic sigmoid - Symmetric sigmoid transfer functions with 20 and 30 neuron number for forecasting of the NOx and NO concentration levels, respectively. In addition, the forecasting of NO2 emission rate, the best results were determined from the ANNs structure used Logarithmic sigmoid - Linear transfer function with 30 neuron number. According to sensitivity analyses and correlation tests, it was concluded that O3, SO2, wind direction, wind speed, and relative humidity inputs were more effective on the NO2, NO and NOx concentrations than the other inputs. Finally, it can be said that with the use of both air pollutants and meteorological factors as input parameters simultaneously the artificial neural network models can be simulated the concentration level of NO, NOx and NO2 with high accuracy.
Read full abstract