Can artificial neural networks estimate potential evapotranspiration in Peruvian highlands?

Abstract

Evapotranspiration (ETo) is one of the most important variables of the water cycle when water requirements for irrigation, water resource planning or hydrological applications are analyzed. In this context, models based on artificial neural networks (ANN) of the retro-propagation type can be an alternative method to estimate ETo in highland regions using a number of input variables limited. The objective of this study is to develop ANN models to estimate ETo for the Peruvian highlands using input variables such as maximum air temperature (Tmax), minimum air temperature (Tmin), hours of sunshine (Sh), relative humidity (Rh) and wind speed (Wv), as an alternative method to FAO Penman–Monteith method (FAO-PM56) and Hargreaves–Samani (HS). Daily climatic datasets recorded at 12 meteorological stations between 1963 and 2015 were selected in this study. For evaluation reason, the ETo calculated using the FAO-PM56 was also considered. The main input variable to ANN modeling is Tmax, followed by Sh and Wv or combinations between them. Hargreaves–Samani (HS) showed a poor performance in the estimation of the ETo in the Peruvian highlands compared to the 13 ANN models. Additionally, it was determined that in stations with lower thermal amplitude (< 14.2 °C) the lowest performance levels are presented in the estimation of the ETo with HS equation, which does not occur markedly with the ANN models that they estimate adequately ETo. Therefore, ANN models represent a great option to replace the FAO-PM56 and HS method, when ETo data series are scarce.