Abstract
The paper is devoted to the development of a modification of the Max Voting method to support the ensemble modeling of neural networks for predicting hydraulic resistance in open river channels, known as the Chézy roughness coefficient. The relevance of the study is driven by the need to develop computational algorithms for obtaining accurate and reliable Chézy coefficient forecasts using artificial neural networks under conditions of limited and incomplete data to support mathematical modeling of river flows for various engineering and environmental tasks. It is established that theapplication of ensemble learning methods for neural networks ensures a significant increase in the accuracy and reliability of predictions. A model of ahomogeneous ensemble of fully connected feedforward artificial neural networks with a sigmoid activation function is proposed for calculating theChézy coefficient based on limited field data regarding the hydro-morphological characteristics of open river channels. The construction and trainingof the neural network ensemble are carried out using the Bagging method. Within this approach, a modification of the Max Voting (Hard Voting)method based on the inverse problem solution is presented to implement the aggregation of neural network ensemble forecasts. The implementation ofalgorithms for ensemble model construction, ensemble learning, and forecast aggregation is performed using Python. The results of the validation of the proposed neural network ensemble using the modified voting method demonstrate the high efficiency of the approach. It was found that within the test case, the relative prediction errors range from 0.5 % to 3.9 %, and the Nash-Sutcliffe model efficiency coefficient is 0,975.