Simulation of reference evapotranspiration using Artificial Neural Method and Empirical Methods and comparison with experimental Lysimeter data in cold semi-arid climate of Hamedan

Using two different artificial neural methods [Artificial Neural Network (ANN) and Adaptive Neuro-Fuzzy Inference System (ANFIS)] and two empirical evapotranspiration models [Penman-Montieth FAO56 (PMF56) and Blaney-Cridle (BC)] daily reference evapotranspiration values (ET0) were estimated. Time series of two years observed meteorological data (1997-1998) were used as the input and output for modeling non-linear system of ET0 values. The structures of ANN and ANFIS network were constructed in such a way that in addition to their ability for comparing the performance of different arrays, they were able to show the effect of dynamical behavior of the used networks. The model outputs were validated against a two years (1997-1998) Lysimeter (1*1*2.25m) data, in addition to the daily weather observations (minimum and maximum temperatures, minimum and maximum relative humidity, sunshine duration, wind speed) in cold-semi climate of Hamedan. The results indicate that neural networks perform reliable ET0 estimations in both decreasing and increasing steps, in comparison with the classic methods. In addition, the neural intelligent networks provide more accurate estimates in shorter computer time compared with the results obtained by classic approaches. It is shown that validations against independent data set are reliable (with coefficient of correlation of about 0.95) for such non-linear dynamical processes. The comparisons of ANN and ANFIS results indicate that ANN performs more accurate ET0 estimates. For the climate condition of Hamedan, the ANN method with 6-3-1 arrays and feed-forward back propagation (FFBP) learning rule performed better predictions compared to the ET0 estimated by ANFIS.