Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Civil-Comp Proceedings
ISSN 1759-3433
CCP: 97
Edited by: Y. Tsompanakis, B.H.V. Topping
Paper 24

A Two-Stage Neural Network Architecture for Forecasting River Flows

M. Cisty

Department of Land and Water Resources Management, Faculty of Civil Engineering, Slovak University of Technology Bratislava, Slovak Republic

Full Bibliographic Reference for this paper
M. Cisty, "A Two-Stage Neural Network Architecture for Forecasting River Flows", in Y. Tsompanakis, B.H.V. Topping, (Editors), "Proceedings of the Second International Conference on Soft Computing Technology in Civil, Structural and Environmental Engineering", Civil-Comp Press, Stirlingshire, UK, Paper 24, 2011. doi:10.4203/ccp.97.24
Keywords: hydrologic prediction, river flows, multi-layer perceptron, self-organising map, hybrid model.

Accurate streamflow forecasts are an important component of watershed planning and sustainable water resource management, which has continually attracted the interest of scientists and engineers. They can help in predicting agricultural water supplies, predicting potential flood damage, organizing land use, estimating loads on bridges, etc.

A hydrological system is influenced by many factors, such as weather, land with vegetation cover, infiltration, evaporation and transpiration, so it includes a good deal of stochastically dependent components, multi-time scales and highly nonlinear characteristics.

In the modelling of a hydrological time series, one of the key problems is noise, which could be different under various hydrological conditions which could occur in a given watershed. In general, it is hard for a single model to capture a dynamic input-output relationship incorporated in the hydrologic data. Furthermore, using a single model to learn the data could introduce some inaccuracies as there are obviously various noise levels in different input regions: when the single model is extracting knowledge from some input region (e.g., in a small flow regime), it could induce local over-fitting in this region and potentially bring a loosening in the degree of precision in another region (e.g., during a condition when peak flows are predicted).

A potential solution to the above problems is to use a combination of data-driven models or an ensemble of data-driven models. In order to improve the accuracy of the prediction of a single multi-layer perceptron (MLP) model, this paper proposes a hybrid learning architecture which combines a modular set of MLP predictors with a self-organizing feature map. Specifically, clustering is used to find different input regions to which the whole input space can be decomposed. After clustering using a self-organising map (SOM), the individual prediction models are trained for each cluster or each class with the aim of obtaining a solution of the problem for the input data included in a particular cluster.

The feasibility of this proposed model is evaluated using actual river flow data. The daily river flows and other data used to construct the models were taken from the flow measurement station at the upper part of the River Hron, and the meteorological data were taken from the Telgart measuring station (Slovakia).

The statistical evaluation of the results and comparison with a standard global model shows that an adequate distinction of the hydro-meteorological conditions characterizing the basin may improve the rainfall-runoff modelling performance, because all the statistical characteristics are best for a proposed modular model with four MLP predictors.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description