Prediction of Daily Pan Evaporation using Wavelet Neural Networks

Abstract

Prediction of daily evaporation has an important role in reservoir management, regional water planning and evaluation of drinking-water supplies. The main purpose of this study was to assess different types of mother wavelet as activation functions instead of commonly used sigmoid for finding the main differences in the results of daily pan evaporation prediction in the Lar synoptic station. So, using conjunction of wavelet theory and multilayer perceptron (MLP) network, two mother Wavelets named Mexican Hat and polyWOG1 are considered for developing hybrid WNNs. The algorithms were trained and tested using a 6-year data record (1999 daily values) from 2005/01/01 to 2010/09/01. Instead of using common sigmoid activation functions in MLP network, wavelet function was applied to construct the wavelet neural network. Results show that Mexican hat wavelet neural network in the best topology presents 98.35 % accuracy in training phase and 96.04 % in testing and PolyWOG1 wavelet neural network in the best topology presents 95.92 % accuracy in training phase and 91.03 % in testing of model. In the MLP model with standard sigmoid function results were 90.6 % in training and 87.63 % in testing. Comparison of WNN and MLP shows that Mexican hat mother wavelet could have better accuracy in the daily pan evaporation modeling.