Forecasting of Annual Precipitation in Kermanshah and Nouje Hamedan Synoptic Stations by Using Artificial Neural Network

The importance of precipitation forecasting as the most important climatologic element and the basis of all planning, especially in areas where there are significant changes in precipitation regimes, is considerable. The use of artificial neural networks is one of the forecasting methods that have been widely developed in recent years. In this research the data set of some climatologic elements, related to the cold seasons of previous year, were used to forecast the next year precipitations in Kermanshah and Nouje Hamedan synoptic stations. Therefore, time series of seven climatologic elements including mean temperature, precipitation, relative humidity, mixing ratio, vapor pressure, dew point temperature and sea level pressure were entered to the artificial neural networks as input, while next year precipitation was considered as the output of network. Due to nonlinear nature of selected climatologic elements of this research, the multi-layer perceptron (MLP) networks were applied. In fact, they are among the progressive networks with supervising training algorithms and suitable for nonlinear data. Other two categories of training algorithms including BP training algorithms and number normalization algorithm were used for training of networks. Eventually, the combination of these algorithms led to the production of 720 training networks at two stations, and finally the artificial neural network succeeded to proper forecasting of annual precipitation amounts. The best forecast for Kermanshah station is related to the traingd training function with mean and standard deviation normalization algorithm and testerror amount equal to 0.0195 in cold period of the year (overall autumn and winter), and in Nouje station is related to the traingdx training function with pca 0.06 normalization algorithm with testerror amount equal to 0.0047 in the winter.