Forecasting of Reference Evapotranspiration using MLP, RBF, and SVM Neural Networks

Abstract

Estimation of reference crop evapotranspiration is one of the most important elements in optimizing agricultural water consumption and in management of water resources. Forecasting the daily and weekly reference evapotranspiration can be used in predicting of crop water requirements and in short range planning irrigation. The aim of this study was to evaluate the performance of three types of artificial neural networks: MLP (multilayer perceptron), RBF (radial basis function network), and SVM (support vector machine) in forecasting the daily and weekly reference evapotranspiration at Tabriz synoptic stations. For this purpose, the meteorological data of 39-year period (1971-2009) were used. To train the neural network, 80 percent of time series data was selected randomly and 20 percent of data was used to validate the different models. To create the time series of daily and weekly reference evapotranspiration in a given period, the standard Penman-Monteith FAO 56 equation was used. Different combinations of input data (different delays) were used to evaluate the models. The results of daily forecasting of reference evapotranspiration showed that SVM with RBF kernel with input set of M5, RMSE=0.51 mm/day and R2=0.92 had the best performance. Moreover, the results of weekly forecasting of reference evapotranspiration showed that SVM with polynomial kernel with inputs set of M8, RMSE=3.88 mm/week and R2=0.95 had the best performance.

Keywords

• artificial neural networks
• crop water requirement
• time series
• penman-monteith
• tabriz