Artificial neural network models for reference evapotranspiration in an arid area of northwest China

We trained and tested artificial neural network (ANN) models for reference evapotranspiration (ET0) using 50 years’ meteorological data from three stations in northwest China. Multiple linear regressions (MLRs), the Penman equation, and two empirical equations were used to compare the performance of the ANNs. A connection weight method was used to quantify the importance of climate factors in performance. In addition, the error changes of the ANNs with seasons were evaluated according to absolute error, variance, and coefficient of variance. Results showed that in arid and semi-arid areas, the ANNs in which the climate data were used successfully estimated ET0, and the ANNs with five inputs were more accurate than those with four or three. Relative to the MLRs, the Penman equation, and empirical equations, the ANNs exhibited high precision. Maximum air temperature, minimum air temperature, and relative humidity were the most crucial input of ANN-based ET0 estimation for arid and semi-arid areas. In the study area, the importance of these three climate factors accounted respectively for 39.82–46.64%, 28.48–33.46%, and 10.73–26.17% to estimation of ET0. Generally, ANNs underestimated ET0 from January to July and overestimated it from August to December.

► Artificial neural network (ANN) models for ET0 were developed.
► ANN models perform well for ET0 estimation than other emprical equations.
► Precision of ANNs for ET0 estimation are various in different seasons.