SIMULATION OF EVAPORATION IN NAM BO PLAIN BY ARTIFICIAL NEUTRAL NETWORK, MULTIVARIABLE LINEAR REGRESION IN COMBINATION WITH RANDOM DATA GENERATION TECHNIQUES

Abstract

This study evaluated the impact of generated meteorological data series on the weekly evaporation simulation results using Artificial Neural Network (ANN) and multivariate linear regression at Can Tho and Nha Be meteorological stations in Nam Bo plain, Vietnam. Because most of data series measured for meteorological factors at both stations do not follow the normal distribution, the data series representing different scenarios had been generated using Monte Carlo, Latin Hypercube techniques with 5%, 10% and quartiles steps based on specific statistics of measured data. The analysis results demonstrated that both multivariate linear regression and ANN methods provided high evaporation simulation accuracy (R > 0.93 or R² > 0.87), while the ANN structure with 1 hidden layer having 6 neurons and tansig transfer function is suitable to simulate evaporation for both stations. The variation in mean and standard deviation of evapotranspiration simulation results is more dependent on choosing steps in the probability distribution initiation processes than on initiation technique selection. In several cases, the results of evaporation simulation by ANN have a negative sign while this phenomenon did not occur with the multivariate linear regression method.