Uncertainty analysis of artificial intelligence modeling daily reference evapotranspiration in the northwest end of China

Abstract

Accurate estimation of reference evapotranspiration (ET0) is of utmost importance for hydrological balance, global change and water resource management. However, caused by the insufficient meteorological data and indefinite input combination, uncertainties may exist in the simplified artificial intelligence (AI) models. Thus, determination the uncertainty is significant for accurate ET0 results. In this study, the validity of 29 combination scenarios of maximum and minimum temperature, wind speed, relative humidity, solar radiation, sunshine duration, and atmospheric pressure was examined by applying the artificial neural network (ANN), support vector regression (SVR), and extreme learning machine (ELM) models for the arid Altay Prefecture. Performances of the models were evaluated against the Penman-Monteith equation by coefficient of correlation (R), root mean squared error (RMSE), mean absolute error (MAE), and Nash-Sutcliffe efficiency (NS) in testing period. The Monte Carlo (MC) technique was firstly performed to analyze the sensitivity of the meteorological parameters and the uncertainty of the AI models. The results confirmed the indispensable role of temperature and the predominant function of the aerodynamic part in evapotranspiration process. An input pattern, which is able to reveal the physical mechanism of AI models in ET0 estimation, was proposed innovatively. Both the SVR and ELM models are highly recommended for ET0 estimation due to the comparable simulating ability and lower uncertainty. The findings help to understand the implied intrinsic mechanism of evapotranspiration in AI models and can be regarded as a breakthrough in ET0 modeling.