Prediction of longitudinal dispersion coefficient in natural streams by prediction map

Abstract

Longitudinal dispersion coefficient can be determined by experimental procedures in natural streams. Many theoretical and empirical equations that are based on hydraulic and geometric characteristics have been developed from the field experiments of longitudinal dispersion coefficient. Regression analysis, which carries some restrictive assumptions such as linearity, normality and homoscedasticity, was used to derive some of these equations. Generally speaking, results obtained from regression analyses are not that accurate as these assumptions are often not satisfied completely. In this study, a method called Prediction Map (PM) is developed based on geostatistics to predict longitudinal dispersion coefficient from measured discharge values, shear velocities, and other conventional parameters of the hydraulic variables and normalized velocity with the objective of overcoming the drawbacks indicated above. As part of this method, a new procedure called Iterative Error Training Procedure (IETP) was developed to minimize prediction error. The prediction error level was reduced after implementing the IETP. PM was compared with various regression models by taking analyzed errors (average relative error percentage and root mean square error), coefficient of efficiency, coefficient of determination and Scatter Index as performance evaluation criteria. The results of the study indicate that the PM approach can perform very well in predicting longitudinal dispersion coefficient by applying IETP. The presented approach yielded lower average relative error percentage, root mean square error and Scatter Indices, and higher coefficient of efficiency and coefficient of determination values compared to the regression models. One of the important advantages of the PM method is that valuable interpretations and a prediction map can be extracted from the resulting contour maps, and as a result, more accurate predictions can be obtained compared to regression analysis.