Forecasting of Daily Reference Crop Evapotranspiration Using Wavelet-Artificial Neural Network Hybrid Model

Reference crop evapotranspiration is one of the most important and effective factors for optimizing agricultural water consumption and water resources management. Forecasting of daily reference evapotranspiration can be used for short-term planning of irrigation water requirements. In recent years the use of artificial neural networks and wavelet-neural hybrid model has become very popular in the forecasting of hydrological parameters. The aim of the present study is to use artificial neural network and wavelet-neural network models to forecast reference evapotranspiration in the range of 1 to 28 days ahead at Tabriz synoptic station. For this purpose, a 10-year period (2000 to 2009), 7 years (2000-2006) for training and 3 years (2007- 2009) to test and validation were considered. To create daily reference evapotranspiration time series at the given period, standard equation Penman-Monteith 56 is used. Different combinations of input data (various delays) and various mother wavelets were used. Results of reference evapotranspiration forecasts for a one day ahead, showed that the wavelet-ANN model (RMSE=0.07 mm/day and R=0.999) compared to the artificial neural network model (RMSE=0.69 mm/day and R=0.964) has higher accuracy in forecasting of reference evapotranspiration. The results showed that the use of time delays of 1 to 7 (M7) and 1 to 6 (M6) days can provide the highest accuracy and fewer delays and delays from one year to two years can reduce the accuracy of the models. Comparison of mother wavelets showed that Meyer wavelet due to greater complexity and similarity to the time series of reference evapotranspiration would increase the accuracy of forecasting. To forecast reference evapotranspiration in 2 to 28 days ahead, the wavelet-neural network with Meyer wavelet model was used. The results showed that the increasing of the forecasting period from 2 to 28 days would decrease the accuracy of models (R is ranged from 0.997 to 0.929 for 2 and 28 days ahead). Also, at forecasting 2 to 12 days, annual delays reduced the accuracy of the model, while at 13 to 28-day annual delay increased the accuracy of models. Finally, to compare the models, statistically, t and F tests were performed to compare the mean and variance. The comparison showed that in all the proposed models, at 99 and 95 percent levels, there was no significant difference between the results and observations. The results of this study can be used in irrigation scheduling at study area.