Application of artificial neural networks as decision-support for operational and planning purposes in water-resources management

The PAI-OFF method (Process Modelling and Artificial Intelligence for Online Flood Forecasting) combines the reliability of physically based hydrological/hydraulic modelling with the operational benefits of methods of artificial intelligence. Such benefits are extremely low computational effort in forecasting, robustness, and easy handling of the forecasting system. However, the use of the PAI-OFF presupposes at first the establishment of a physically based hydrological model of the catchment under consideration. In the presence of particular hydrodynamic conditions, such as backwater effects, a hydrodynamic flood-routing model of the affected river reach is also included. Both models are then used to simulate the response of the catchment to the full range of potential flood-generating precipitation events. The resulting rainfall-runoff data are processed into a database of corresponding input-output vectors. This database is supplemented by hydrological/meteorological information that characterizes the state of the catchment at the beginning of an event. Then several artificial neural networks are trained with the help of this database; a polynomial neural network (PoNN) to describe the rainfall-runoff function and a multilayer-feedforward network (MLFN) to depict the flood-routing process. The presentation of the theoretical fundamentals is then followed by an application of the PAI-OFF method in flood forecasting in the fast-responding catchment of the River Freiberger Mulde (catchment about 3,000 km super(2)) in the Saxon Ore Mountains. Both the outstanding fastness of the computation and the high quality of the forecasts highlight the potential of the PAI-OFF method for online flood-forecasting in rapidly responding catchments.