Programme Poster session 2 abstract 617
Rainfall-runoff modeling for flood forecasting: application of global
methodologies to a medium-size basin in Brazil
Author(s): Bruno Rabelo Versiani (1), Marcus Felipe Matos Cruz (2)
, Alan Fábio de Bastos (2)
(1) Corresponding author: Associated Professor - Universidade Federal de Minas Gerais - Departamento de
Engenharia Hidráulica e Recursos Hídricos - Av. do Contorno, 842 - 30110-060 – Belo Horizonte – Minas Gerais
– Brazil – Phone: ++ 55 31 34994821 – Fax: ++ 55 31 34994823 – Email: versiani@ehr.ufmg.br
(2)
Universidade Federal de Minas Gerais - Departamento de Engenharia Hidráulica e Recursos Hídricos - Av. do
Contorno, 842 - 30110-060 – Belo Horizonte – Minas Gerais – Brazil.
Keyword(s): rainfall
-runoff modeling, DPFT methodology, GR3 model, Artificial Neural Network model
Session: Poster session 2
Abstract This paper deals with the
hydrologic modeling of rainfall-runoff relationship, employing a global approach, applied to a medium-size basin in
Brazil. In general the Brazilian hydrologic gauging stations network display only data for precipitation and runoff and,
in the majority of cases, only daily data is available. Due to the difficulty in the calibration of the large number of
parameters that conceptual models generally require, global empirical methods are currently used in studies and
technological applications.
It was used the DPFT ( Différences Premières de la Fonction de Transfert - First
Differences of the Transfer Function) methodology to identify the Unit Hydrograph and effective precipitation.
Otherwise, a non-linear Artificial Neural Network ( ANN ) model approach was used to compare with the DPFT
methodology.
Basically, the Unit Hydrograph method proposes that, for a given basin, runoff is the result of a loss
function and of a linear transfer function. The loss function is strongly non-linear, and transforms the total measured
precipitation, either weighted or arithmetic mean, into effective precipitation, which produces the surface runoff. The
linear transfer function increases over time the effective rainfall, so as to obtain the surface runoff. Classically, a loss
function model (assumed to be the most appropriate for the basin) is required, a priori, to obtain the effective rainfall
for each event. As opposed to this classic approach, the DPFT methodology, proceeding iteratively from an array of
events of total rainfall – runoff, establishes the Transfer Function, and the effective rainfall for each event. This
distinctiveness permits a comparison and choice, a posteriori, of the best loss function for a given hydrographic
basin. Two simple models of loss function were studied and calibrated: a Reservoir model (with three parameters)
and the GR3 model (with one parameter)
An Artificial Neural Network (ANN) is a flexible mathematical
structure which is capable of identifying complex non-linear relationships between input and output data sets, as
rainfall-runoff processes. The employed ANN model is based on the class of higher order feedforward neural
networks, the Non-linear Sigmoidal Regression Blocks Networks (NSRBN), which deploy a constructive
incremental learning algorithm that is responsible for the network´s definition.
The DPFT methodology was
applied to the Rio das Velhas basin, at Honório Bicalho cross section ( 1655 km2 ), located in the State of Minas
Gerais – Brazil, using 15 significant events of daily rainfall and runoff data. The results were analyzed and compared
with regards to their efficiency. When analyzing the performance of the Reservoir and GR3 models in the calibration
phase, it was established that the loss function models produced similar results However, in the validation stage, a
superior performance of the GR3 model was observed.
With respect to the ANN approach, the results suggest
that it may provide a superior alternative to the other used models for developing input-output simulation and
forecasting, in situations that do not require modeling of the internal structure of the basin. Nevertheless, the ANN
approach presented does not provide models that have physically realistic components and parameters.