عنوان مقاله [English]
نویسندگان [English]چکیده [English]
The complete time series of meteorological data with enough length is one of the most fundamental issues in environmental and hydrological studies. However, weather stations often have statistical gaps due to lack of measured data or technical problems (Tardio and Bertie, 2012). Precipitation data are so important for climate studies and frequency analysis of hydrologic events such as floods, droughts and also water resources planning and management. Estimation of missing values has been a subject of interest to meteorologists, hydrologists and environmental experts. Recently, artificial intelligence models support vector machine (SVM) and Artificial Neural Network (ANN) has been used to reconstruct the rainfall data. SVM is a supervised learning method that can be used for classification and regression. This method, developed by Vapnik (1998), is based on the theory of statistical learning. The other method which is widely used to predict the characteristics of non-linear and complex phenomena is artificial neural network. Artificial intelligence and neural network models were proposed by McCulloch and Pitts (1943), for the first time, to predict precipitation. Golabi et al. (2013) compared the performance of different algorithms of artificial neural networks for modeling seasonal rainfall in Khuzestan province. In this study, MLP and RBF networks with applying some changes in the middle layers, neurons, and training algorithms MOM, LM and CG were used to predict seasonal rainfall. The main purpose of this study was to compare the performance of artificial neural network (ANN) and Support Vector Machine (SVM) in reconstructing of precipitation in rain gauge stations of Hamedan Province. For this purpose, the monthly precipitation data of four rain gauge stations, including Aghajanbolaghi, Sarabi, Aghkahriz and Maryanej, were used during the period of 1991 to 2010.
SVM regression model is depended on a functional dependence y to a set of independent variables x estimated. It is assumed that other issues such as regression, the relationship between the dependent and independent variables by a certain function f noise, plus an additional amount are determined. In this study, SVM model was used to reconstruct the missing data by using non-linear kernel function of RBF Statistica10 and with more than 100 replicates. The Kernel function were used because higher accuracy in the reconstruction of the monthly precipitation data. Optimal characteristics of SVM model include ε, C and ϒ should be determined. For this purpose, characteristics C and ? by was calculated by the network search optimization algorithm and variable ? by trial and error. An artificial neural network is composed of neurons. The smallest units of information-processing are neurons or nodes which form the basis of the performance of neural networks (Minhaj, 2005). Each neuron receives input and then processing them, generate an output signal. Thus, the neurons in the network act as a center for processing and distribution of information, its own input and output (Sadorsky, 2006). The study of artificial neural networks Multilayer Perceptron (MLP) is used by the propagation algorithm. The precipitation data of four stations including Aghajanbolaghi, Sarabi, Aghkahriz and Maryanej, were used in this study. These stations were chosen based on their close distance and high correlation in monthly precipitation data. To determine the best input pattern for the network, the various factors that may affect the phenomenon should be considered. To reconstruct rainfall data in neural network, programming in Matlab software is used. In order to evaluate the performance of Support Vector Machine (SVM) and artificial neural networks (ANN) in reconstructing the missing precipitation values, the Coefficient Determination (R2) and root mean square error (RMSE) were calculated for observed and estimated precipitation values.
In this study, the MLP artificial neural networks and RBF Support Vector Machine models were used for reconstruction of monthly precipitation data of rain gauge station of Sarabi in Hamedan province, during the years 1991 to 2010. Three modes were intended as training models. Models for training in the first case used only one station data (Aghajanblaghy), the second used two stations data (Aghajanblaghy and AqhKahriz) and for the third training data from three stations were used (Aghajanblaghy, Aqhkahriz and Maryanaj). Also, Sarabi was considered as reconstruction station. The results of this study indicated that when the data for one and two stations were used for training in modeling, the Support Vector Machine model showed better performance than SVM. While, using the data from three stations in the models, the artificial neural network model was better than the Support Vector machine with a little difference. Reconstructing of monthly precipitation, the best performance was obtained by using data from three stations for training ANN and SVM. Coefficient Determination and RMSE for the models were 0.88 and 12.33 mm for ANN and 0.87 and 12.89 mm for SVM. Reconstructing of seasonal precipitation, the best results was gained by using data from three stations for training ANN and SVM models. Coefficient of determination and mean squares for these models were 0.95 and 26.62 mm for ANN and 0.94 and 18.82 mm for SVM. Although both models in three different educational models performed almost the same in the reconstruction of monthly data (Sarabi), the results showed that when the number of stations in learning more models increased, the models performance improved.