Investigation of Artificial Neural Network Based Models and Sensitivity Analysis for Reference Evapotranspiration Estimating

Reference evapotranspiration (ETo) is one of the most critical parameters in proper design of irrigation systems. Accurate estimation of ETo leads to reduction of water losses. Due to the ability of Artificial Neural Networks (ANNs) in computational analysis of complex processes, the main objective of this study was to investigate the sensitivity of the ETo trends to key climatic factors in Tehran province using the artificial neural networks, and compare it with the ETo-calculator software results. The ETo was calculated using meteorological data (10-year data of 12 meteorological stations in Tehran province) using the ETo-calculator software. In order to model ETo, a set of inputs to artificial neural networks including the minimum and maximum air temperature (Tmax and Tmin), the minimum and maximum relative humidity (RHmin and RHmax), sunshine hours (n), and wind speed (U2) were considered. After data tagging, by optimizing the number of hidden layers and network algorithms, output values were estimated. The results indicated that artificial neural network is a suitable technique for ETo analysis(R^2≅98% ). The best model for estimation of ETo is feed-forward Multi-Layer Perceptron (MLP) with two hidden layers in its structure (6-11-14-1), Levenberg–Marquardt training algorithm for both hidden and output layers and Linear Tanh and Tanh transfer functions for hidden and output layers, respectively. The sensitivity analysis of the model for input parameters showed that the optimal artificial neural network model and ETo calculator software have the same trend and the Tmax and n are the most effective and least effective parameters in ETo estimation, respectively. Also, based on PCA analysis results the scenario of using of four parameters (Tmax, Tmin, RHmax and U2) as the only inputs to the selected artificial neural network, can estimate ETo with an acceptable accuracy〖(R〗^2≅94% ).