Using artificial neural networks as a basis for decision-making in operational and planning water management measures

We develop, test and apply a new methodology for providing constantly up-datable, real time operational flood forecast models for online application. The methodology is based on complex detailed rainfall-runoff models coupled with full hydrodynamic flood-wave propagation models. These models are then run through an artificial network which subsequently provides a reliable flood forecast.

Abstract

The frequent occurance of floods in recent years indicates the importance of reliable flood forceasting with respect to economic and ecologic problems. Today, operationally used models are based on simple empirical approaches which are usually calibrated on historical records. However, these simple approaches may fail when predicting extreme flood events because adequate historical records are rarely available.  Furthermore, as a consequence of the considerable simplifications, empirical models cannot cope with the great complexity of the formation of floods in a catchment. For this reason, empirical modelling relies on adjustments by means of measured water levels, which has the disadvantage of restricting the prediction horizont.

Another concept uses physically based rainfall-runoff modelling. Here, physical catchment-parameters are determined as the basis for the computation. This also allows more realistic calculations  for unobserved (future) events and enables the estimation of effects caused by changes in the hydrologic system (changes in landuse, relocation of dams, etc.). The disadvantages of these physically based approaches are their complex parametrisation and excessive computational efforts. The latter is especially the case for a realistic and spatially high resoluted modelling of large catchments and long-term simulations for assessing catchment states. In this respect, artificial neural networks (ANN) are recommended as an efficient and robust tool operational flood forcasting. In order to maintain the advantages of physically based models the ANN are trained on the results of these models. Furthermore each application of the trained ANN results in an increase in its predictive capability as each application continously enlarges the datapool with newly observed timeseries and, thus,  is actually able to adjust the ANNs behavior to gradual changes in catchment characteristics.

The objective of this research is the development and programming of a practically oriented water management tool based on artificial neural networks. The new tool serves as an integrative (multipurpose) rainfall-runoff model which provides a simple and very straightforward runoff prediction. This provides a decision support for operational and planning water management measures. A first application of the new tool focuses on the catchment of the river Freiberger Mulde.

Duration:          Aug. 2002 - Dez. 2006
sponsored by:  BMBF