Potential evapotranspiration prediction using nonlinear autoregressive model with exogenous input (NARX) (case study, Yazd Province, Iran)

Potential evapotranspiration (PET) is one of important hydrological cycle parameters which its prediction is useful for water resources planning in the future, changes in the crop water requirements, and drought forecasting. In the case of PET long term prediction, global climate models (GCMs) based on the emission scenarios and then downscaling the outputs are used. For short term forecasting, statistical models are suggested. PET is a nonlinear phenomenon which is affected by temperature, humidity, sunshine and wind speed. In this study, PET was calculated by FAO-Penman-Monteith model for Yazd synoptic satiation during 1965-2010. Recently, NARX has widely used to predict hydrological and climatic parameters. In this research, performance of NARX for PET forecasting was analyzed. NARX is a nonlinear model which in addition to the target parameter, takes exogenous variables that affect target variable into account. Different nonlinear functions can be selected. In this research, a feedforward neural network because of its high ability in nonlinear processes modeling was chosen. The network was trained by GDX algorithm. Then, PET was predicted using nonlinear autoregressive model (NAR) which its input is just PET. Next, the results of NARX and NAR were compared. The results showed adding exogenous variables increases the accuracy of PET prediction, remarkably. R2 correlation coefficient between one month observed and predicted PET by NAR and NARX for 2002 to 2010 was 0.72 and 0.92, respectively.