Data arrangements to train an artificial neural network within solving the tasks for calculating the Chézy roughness coefficient under uncertainty of parameters determining the hydraulic resistance to flow in river channels

Abstract

Hydraulic calculations and mathematical modelling of open flows in river channels keep still being among the most topical hydro-engineering today’s problems in terms of practice. While solving them, independently on the research topic and purpose, and methods used, a number of simplifications and assumptions are usually accepted and applied. Moreover, there is a range of methodological, structural, and parametric uncertainties, which to be overcome require complex empirical pre-researches. First of all, these uncertainties relate to assessing hydraulic resistances and establishing numerical characteristics of them, which depend on many factors varying spatially and temporally.One of the most frequently used integral empirical characteristics expressing the hydraulic resistance to open flows in river channels is the Chézy roughness coefficient C. However, despite a large number of empirical and semi-empirical formulas and dependencies to calculate the Chézy coefficient, there is no ideal way or method to determine this empirical characteristic unambiguously. On the one hand, while opting for an appropriate formula to calculate the Chézy coefficient, we need to take into account practical experience based on comprehensive options analysis considering different empirical equations used alternatively to represent the hydraulic resistance to open flows. On the other hand, the fullness and comprehensiveness of field researches of numerous hydro-morphological factors and parameters characterizing various aspects of the hydraulic resistance to open flows can also have an essential role. In particular, the accuracy assessment of the Chézy coefficient computing based on field data, despite methods and formulas, indicates that the accuracy of field measurements of the parameters included in selected formulas largely determines the relative error of such calculations.This paper deals with the problem of data arrangements and the development of general rules for the formation of training and test samples of data to train artificial neural networks being elaborated to compute the Chézy coefficient taking into account the parametric uncertainty of data on the hydro-morphological factors and parameters characterizing the hydraulic resistance in river channels. The problem is solved on the example of an artificial neural network of direct propagation with one hidden layer and a sigmoid logistic activation function.