Application of artificial neural networks for the hydrological modeling of watersheds
Name: LAISI BELLON CESCONETTO
Publication date: 30/08/2021
Advisor:
Name |
Role |
|---|---|
| ROBERTO AVELINO CECÍLIO | Co-advisor * |
| SIDNEY SARA ZANETTI | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| ROBERTO AVELINO CECÍLIO | Co advisor * |
| SIDNEY SARA ZANETTI | Advisor * |
Summary: Water is an essential natural resource for the maintenance of life. It is an essential element for the development of agriculture, ecosystem maintenance, forest development, among others. Population growth and, consequently, the increase in water demand, has promoted the development of studies on techniques that positively contribute to the management of water resources. Currently there are many hydrological models, but some models model a large amount of variables. However, computational advances, coupled with artificial intelligence, have provided better models for estimating hydrological variables. Artificiais neurais networks (ANN´s) are efficient, multivariate and nonlinear tools, and that use the multilayer perceptron method (MLP) has been the most used in the modeling of water resources. Thus, the objective of this work was to develop a method of simulation of total flows through the use of ANN and to evaluate an applicability to estimate reference flows. Synthetic flow data series were generated from source data and other variables tested (area and month). The model choice process was based on statistical indices. The RNA´s used in this work were of the MLP type containing three layers. Initially, a general model was proposed for the entire region applied in the case study (Espírito Santo state, Southeast region of Brazil) and the training of networks occurred simultaneously for an entire database, but always leaving out the data of a fluviometric station, representing a test sample. In addition, models were also tested in clusters and in pairs of stations. For the general model, considering the input variables proposed by Vilanova, Zanetti and Cecílio (2019), it was observed that the values of NSE and NSElog were higher than 0.30, with the exception of the tests carried out in the hydrographic basins of the São Mateus, Jucu and Santa Joana Rivers, WHERE the results were worse. It found that the inclusion of accumulated rainfall from previous days, the number of the month and the drainage area, simultaneously, resulted in a more expressive improvement in the results. As for the models in clusters, there was no improvement only for cluster 1. For the model applied to pairs of stations, in general, there was no improvement in the results. Regarding the reference flows, there was a tendency to overestimate the q7,10 evaluated by general models, by grouping and by pairs and in relation to q7,10 estimated with the recorded data. However, for q90 there was a significant improvement in the results obtained when the model was applied in group 2. As for the average flows, it was observed that there is a tendency to overestimate the qmld of the recorded data; however, the average percentage errors of the average flow simulated by the models in the Metropolitan region were lower than those estimated by the traditional method of regionalization of flows. Thus, it can be considered that the proposed ANN model is viable in the estimation of flows and the choice of the spatial amplitude of the method application will depend on the availability of recorded data and the objective of the work.
Keywords: hydrological modeling, artificial neural network, rainfall-runoff model, flow simulation.
