Design of Artificial Neural Networks for Flood Forecasting

Abstract

This chapter provides an overview of currently available neural network design with the purpose of a timely warning for operational flood risk management, considering the need to evaluate the uncertainty of the forecast. Neural network models are very effective with regard to their computational requirements and provide new options for operational scenario analysis and ensemble forecasts. Here the widely used “multi layer feed forward network” is compared to an alternative, the “polynomial neural network”. A new training strategy permits to discriminate between input vectors. This method opens a way to reflect physical facts by means of input vectors in neural models, i.e. the neural model is portraying the rainfall runoff process on the basis of process understanding and physical boundary conditions of the considered catchment.