NEURAL NETWORKS IN THE ICE-CORRECTION OF DISCHARGE OBSERVATIONS

We have applied three models, a neural network, a conceptual model and a combination of these two, a hybrid model, to model the backwater ef fect of ice in a river. The neural network is a black-box model. It is based mainly on observed data and it lacks the expert knowledge of th e system. The conceptual model is based on a physical description of t he system. The data is used in optimizing the free parameters of the d escription. In the hybrid model, the neural network is modified so tha t the physical description of the conceptual model can be coded into t he structure of the network. In the beginning of fitting, the hybrid n etwork already performs as well as the conceptual model. During fittin g also the structure of the physical description is optimized, not onl y the parameters of the description. The three models are rather diffe rent in form but in the modeling results there are only slight differe nces. Mean error of the models in ice-correction is 13-15 m(3)/s at an observation station where the mean backwater effect of the ice is 100 m(3)/s. The aim of this work is to develop a model for real time esti mation of corrected discharge, which is used in error correction of a discharge forecast model. For this purpose the error of the best model is acceptable.