Application of artificial neural network in estimating monthly time series reference evapotranspiration with minimum and maximum temperatures

Abstract

There are various methods for computing reference evapotranspiration (ETo) using meteorological data. However, such models tend to perform well for predicting ETo close to the mean, but do not keep accurate performance with extreme observations. It is recognized that the Penman–Monteith (PM) model has the best performance when rich data is available to calculate the ETo, which is not frequently available to a certain extent. In case of poor data, such as prediction of futuristic ETo while investigating climate change effect, although there are models other than PM like Hargreaves–Samani (HGS), the universal sustainability of these models are not quit proved. Accordingly, the calculation of ETo still required numerous research to reach accurate estimation of ETo specially when there is lacking for data to utilize PM method. Recently, methods based on artificial intelligence (AI) have been suggested to provide reliable prediction model for several application in engineering. This manuscript employed artificial neural network (ANN) for predicting daily ETo at Rasht city located northern part of Iran using minimum and maximum daily temperatures collected from 1975 to 1988 of the region. A comprehensive data analysis utilizing the daily time series, minimum and maximum temperatures and solar radiation (T min, T max and R s), as input pattern to predict daily ETo at the current month and for the following month is proposed. The employed ANN model was feed forward backpropagation (FFBP) type with Bayesian regulation backpropagation. The mean square error, mean absolute error, mean absolute relative error and regression coefficient are the statistical performance indices used to evaluate the model accuracy. The results showed that the proposed ANN model could successfully be used to predict daily ETo using only maximum and minimum temperatures with significant level of accuracy. In addition, results show that the proposed ANN model outperforms HGS method.