Prediction of horizontal visibility by training feedforward network with resilient backpropagation algorithm

Meteorological phenomena are complex systems with different parts that are in contact with each other as well as their surroundings. The purpose of this research is to demonstrate the efficiency of neural networks in predicting meteorological variables. For this purpose, the prediction of horizontal visibility that is widely used in meteorology and aviation especially at airports has been selected for analysis. The data of this study are a compilation of Metar and Synop reports of Bandar Abbas synoptic station in the period from 1 to 30 March 2014.
To implement this network, at first, the whole data were randomly divided into three categories with proportions of 75, 15 and 15 percent for learning, testing and validation of network and saved in other files. The seven variables for inputs )temperature, dew point temperature, atmospheric pressure, sky cloud coverage, wind speed and wind direction) of the network with 28 various composites tested with a feedforward network and their correlation with the output and amount of root mean square (RMS) error of network have been studied. The results show, the compositions that containing the present air phenomena are most correlated with the horizontal visibility. Besides, the dew point temperature, atmospheric pressure and the amount of cloud cover are variables that alone do not have an affect on the horizontal visibility.
In this research, a network which works with training neural networks by resilient backpropagation algorithm is used. This is a learning heuristic for supervised learning in feedforward artificial neural networks, which only the sign of the partial derivative is used to determine the direction of the bias and weight updates and the magnitude of their derivative has no effect on their updates. Of course, the size of their change (increment and reduce rates) is determined by a separate update value. This network with eight neurons and sigmoid transfer function in the hidden layer and the linear transfer function in the output layer is used for predicting of horizontal visibility. This network was performed with two standardization data sets between intervals 0.0-1.0 and 0.1-0.9; also, different learning rates, incremental and reduced rates for weights and biases. The results show that the normalization is not appropriate between zero and one. The appropriate amounts of learning rate, incremental and reduced rates for this network are 0.0001, 1.2 and 0.35, respectively.
The values of the coefficient of determination for training, test and validation data for a running network with all variables were 0.9972, 0.9866 and 0.9839, respectively. These values show that nearly 99 percent of the measured horizontal visibility is affected by these independent variables and the rest of its variations are dependent on other factors.