عنوان مقاله [English]
Properly intelligent "input selection" tailored to the target using an appropriate method is the first âstep in the design of Artificial Neural Network (ANN) for prediction. ANN architecture is not predetermined; the weights are determined based on input data during the training process. Therefore, when input data is richer, ANN will be better trained and will have a better performance in predicting meteorological parameters. To solve nonlinear equations governing atmospheric motions, for which no general solutions are known, meteorologists have to use the appropriate approximations for prediction. Using the ability of the ANN to consider nonlinear effects, meteorologists will be able to predict most of meteorological parameters without considering the nonlinear equations governing atmospheric motions. There are two approaches for selecting the appropriate input data. In the first approach, time series of the desired parameter, ANN target, such as temperature, relative humidity, pressure, and wind speed from the previous years are used, while in the second approach, the parameters that have a nonlinear or linear relationship with the ANN target are used. Due to a large volume of input data, errors of measurement, the presence of unusual dataand correlation between input variables, in the case of second approach, error increases and prediction accuracy decreases. In most cases, due to the lack of detailed information concerning the data, the trial and error method has to be used to select a proper combination of input data and elimination of unusual data.The trial and error methodis one of the easiest methods for solving problems. Since in this method the governing relationships among the parameters are not considered, the solutions may not fit in the physical situation. In the present research, to avoid using thetrial and error method, we use Principal Component Analysis (PCA) in order to determine the detailed information concerning the details of the input data. PCA has several abilities such as reduction of dimensions of data, extraction of variability modes of data, eliminating the correlation between raw data, and deletion of unusual data. These abilities can be used in various applications. For example, it is possible to reduce dimension data using PCA when we deal with a large volume of raw data. In fact, we achieve simultaneously three targets by using PCA. The first target is reduction of dimensions of the data; therefore, the training process of ANN performs better than the case when we use raw data. The second and the third targets are the extraction of variability modes and deletion of unusual data; thus ANN does not deviate and overtraining does not occur. In our accompanying research presented at the First Computational Physics Conference in 20â22 January 2014, we demonstrated that these abilities of PCA were very important in properly intelligent input selection tailored to the target. Meteorological parameters associated with temperature to determine appropriate parameters for predicting the average daily temperature in 2009 in Yazd synoptic station in a 29-year period (1980 to 2008) has been analyzed by using PCA. The results showed that using numerous capabilities of PCA, a correct, intelligent input selection appropriate for the ANN target without using the trial and error methods is possible.