Neurohydrological prediction of water temperature and runoff time series

In this paper we give an overview of experiments with artificial neural networks on the Hungarian reach of the Danube River carried out by the Hungarian Hydrological Forecasting Service. Two areas were selected: rainfall-runoff modelling and water temperature simulation. The statistical machine learning method is a universal interpolation and classification tool, but showed poor performance when applied for correlation in complex hydrological situations. Despite very strong learning skills of neural networks even a conceptual model gave more consistent and superior results through validation, and the statistic method is more sensitive to overlearning than deterministic methods. Despite deterministic models being superior artificial neural networks still provide satisfactory results that confirms their application.