Comparison of multiple regression and artificial neural network methods in estimation of pan evaporation and determination of most important affecting variables using principal factors analysis

Pan evaporation is a key element in water resource planning, irrigation management and crop production. In most of the weather stations of country long term, homogen pan evaporation (Epan) data are not available. Hence, empirical model are used to estimate this variable.
The objective of this study was to compare the skill of regression and artificial neural network models in estimation of pan evaporation and determination of most affecting weather variables on Epan by principle factors analysis (PCA) approach using historical climatic datasets of four station namely Karaj, Ahvaz, Shiraz and Tabriz during period of 1986 to 2005.Meterological data including Maximum and Minimum temperature, wind speed and sunshine hours were used as predictors. Results of PCA approach revealed that in Ahvaz, Tabriz and Shiraz stations respectively 90%, 91% and 93% of pan evaporation variations can be attributed to wind speed, sunshine hours and Tmax, Tmin. PCA method did not perform well in Karaj. Application of ANN models showed acceptable results in all study stations. The R2 value was 0.81 in Ahwaz, Shiraz and Tabriz stations and 0.88 in Karaj station. Similarly, the RMSE values for Ahwaz, Shiraz and Tabriz stations was 0.2 mm and 0.1 mm for Karaj station. Besides, based on MSE2 index, the proposed model was overestimating in first three stations and underestimating in Karaj climate.MAE values in Karaj, Ahwaz, Shiraz and Tabriz were 3.83, 33.6,21.79 and 15.6 respectively. Results of multiple regressions showed that in Karaj, Tmax, in Tabriz station, wind velocity and in Shiraz station;Tmin are the most significant affecting variables on Epan with P values of 0.03,0.04 and 0.1 respectivelyý.