Evaluating Different Functions of Artificial Neural Networks for Predicting the Hourly Variations of Horizontal Visibility under Dry and Humid Conditions (Case Study: Zabol City)

Abstract

The present research was conducted to compare different functions of two artificial neural networks (ANNs) including the multilayer perceptron (MLP) and radial basis function (RBF) in order to forecast the Horizontal Visibility (HV<1km) in Zabol city under dry and humid weather conditions. For this purpose, hourly data of horizontal visibility (HV), wind speed, relative humidity, temperature, and atmospheric pressure were used. Before importing these data to the ANNs, they were normalized and multicollinearity impact between the climatic variables was calculated using the variance inflation factor. In this study, 70% of data were used for data training and 30% for data testing. Accuracy of the models was estimated using the mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and the correlation coefficient (R) between observed and predicted values of HV. The sensitivity of the output data was determined based on the most accurate model. The results showed that according to function MLP4, the prediction accuracy of HV was more than the accuracy of other functions of neural networks (ANNs) for both dry and humid climates. The mentioned error values were estimated at less than 0.5. Pearson correlation between observed and predicted values was estimated according to training data and testing data as 0.66 and 0.7, respectively. These coefficients were calculated 0.9 and 0.99 for humid and dry weather, respectively. Moreover, the wind speed and air temperature for dry and humid climate were identified as the most important factors effective on HV at the time of dust storm occurrence.