Using new methods of effective inputs determination for pan evaporation estimation Introduction

Evaporation is one of most important parameters which are affected by many variables such as rainfall, wind velocity, sunny hours, and relative humidity etc. Evaporation estimation is important for any area with surface water resources because of its effect on dam lakes, precipitation-runoff modelling, river area performance, water management – calculating amount of water that plants need and planning for watering and so on. Evaporation can have significant effect on water balance of a river or a reservoir and it may be cause water level to decrease. Due to hydraulic system complications caused by statistical information imperfection and determining all parameters involved, complete hydraulic system modelling is impossible. At such circumstances using al mathematical modelling system will be considered. Matherials & Method: In this study we tried to estimate pan evaporation using two models including Artificial Neural Network (ANN) and Support Vector Machine (SVM) with data preprocessing (gamma test and principal component analysis) to determine affective inputs into two models. For this matter data gatherd from three synoptic stations at Astara, Kiashahr and Talesh at Guilan province has been used. Synoptic stations data includes evaporation, wind velocity at two meter altitude, temperature (minimum, average and maximum), humidity (minimum, average and maximum), sunny and rainy hours. Statistical period of data for Astara and Talesh synoptic stations were 1384 to 1393 and for Kiashahr were 1385 to 1393. 80 percent of meteorology data were used for calibration and other 20 percent were used for model validation. In this study we used multilayer perceptron artificial neural network with sigmoid tangent function and 1 to 20 neurons for hidden layer and support vector machine with radial based kernel function. Calculations has been made in to section with two data preprocess methods. At first section input variable has been selected by gamma test and pan evaporation estimation was made by both models. At second section modelling has been pulled out by input variables selected by principal component analysis. Discussion of results: At gamma test section pan evaporation estimation parameters were as follows: minimum temperature, maximum humidity, minimum humidity, rainfall and sunny hours for Astara station; maximum temperature, minimum temperature, minimum humidity, rainfall and sunny hours for Kiashahr station and maximum temperature, minimum temperature, maximum humidity, average humidity, rainfall and sunny hours for Talesh station. According to principal component analysis results on Astara, Kiashahr and Talesh stations, five, five and four principal component were used in modeling these stations respectively. At first section input compound determined by gamma test to estimate Pan evaporation of the selected stations were used. Pan Evaporation estimation results shows that at Astara station GT-ANN model has less root mean square error than GT-SVM model and beter performance. Pan Evaporation estimation at Kiashahr station was done suitably with both models. At this station GT-SVM had a better performance with root mean square error of 1.295 compared to GT-ANN model with 1.356. At Talesh station both models had close results but results for GT-SVM were more accurate compared to GT-ANN. Nash Sutcliffe coefficient attained for Astara and talesh stations acknowledges their excellent results and for Kiashahr station shows the satisfactory results. At second section modelling were done by using selected inputs by PCA preprocess method. Accordint to results, PCA-ANN model had better performance estimating pan evaporation at Astara and talesh stations than PCA-SVM model as its root mean square error was lower. Value of Nash Sutcliffe coefficient shows the suitable performance of both models at both stations. PCA-SVM model had better performance estimation pan evaporation than PCA-ANN with lower root mean square error at Kiashahr station. Nash Sutcliffe coefficient of PCA-SVM model was 0.666 and for PCA-ANN model was 0.634 which shows the satisfactory performance of both models. Conclusions: Results shows the good performance of preprocessing methods (principal component analysis and gamma test). Actually performance of GT-ANN, PCA-ANN, GT-SVM and PCA-SVM models performance estimating pan evaporation of each one of the stations are very close to each other. This similarity is caused by performance of gamma test and principal component analysis preprocessing methods. Principal component analysis converts input variables to independent principal component using linear relation between input variables. Actually this method reduces the effect of the variables with similar information by giving them lower factor. But in gamma test method consider to gamma factor attained from various input compounds, variable that has a negative effect on output will be determined and eliminated from final input compound. As we said before, nature of none linear Gamma and linear PCA methods are different but when PCA method decreases the factor that is eliminated in gamma test to a small value, inputs determined by both methods will be close to each other. This can be one of the reasons that close the estimating models results to each other. So we cannot recommend one preprocessing methods better than the other. We can conclude that for estimating pan evaporation at these stations both preprocessing methods are suitable. According to results PCA-ANN for Astara and Talesh and GT-SVM model for Kiashahr station had better performance than others. Although both models had acceptable performance estimating pan evaporation of stations but SVM model results were better than ANN model.