Uncertainty of Artificial Neural Networks for Daily Evaporation Prediction (Case Study: Rasht and Manjil Stations)

This research uses the multilayer perceptron (MLP) model to predict daily evaporation at two synoptic stations located in Rasht and Manjil, Guilan province, in north-west of Iran. Initially the most important combinations of climatic parameters for both of the stations were identified using the gamma test; and daily evaporation were modeled based on the obtained optimal combination. The results of the artificial neural network- Gamma Test (ANN-GT) model are evaluated using the root mean square errors (RMSE), correlation coefficient and Nash-Sutcliffe (NS) criteria. The results showed that the ANN-GT model for Rasht station with a correlation coefficient 0.86, root mean square error 0.95 and Nash-Sutcliffe criteria 0.74 and for Manjil station with correlation coefficient 0.94, root mean square error1.58 and Nash-Sutcliffe criteria 0.89 has an acceptable performance in predicting daily evaporation. To evaluate the uncertainty, we considered a percentage of data which were included in 95 percent of uncertainty (p-factor) and the average width of the 95ppu band (d-factor). Regarding the uncertainty results, the average with of 95PPU bound were obtained as 0.33 and 0.3 for the Manjil and Rasht stations, respectively. This shows the low uncertainty level of the ANN-GT model for predicting daily evaporation at both of the stations. Furthermore, the percentage of the observed data at 95PPU band was low and equal to %25 and %45 for the Rasht and Manjil stations, respectively. The reason for these low values can be due to low uncertainty in the parameters.