ABSTRACT
The models aiming at forecasting or projection of water temperature in natural streams located in cold climate zones, where the seasonality plays important role, are of great importance, as stream temperature is still frequently not measured on site and some tools are needed to evaluate water temperature values for future climatic conditions based on simple hydro-meteorological variables. In several papers Artificial Neural Networks (ANN) were proposed to stream temperature forecasting. However, it is still not clear which hydrological and meteorological variables, among the ones that could be available from Global Circulation Models, are the most significant as ANN model inputs. It is well known that using model ensembles may significantly improve the forecasting accuracy, also in the case of ANN models. However, the impact of ANN ensemble size and of the ensemble aggregation approach on the forecasting accuracy has been rarely studied so far. The present paper aims at both, the problem of the choice of proper ANN input variables, ensemble size and ensemble aggregation approach, at an example of Biala Tarnowska river catchment, located in mountainous part of southern Poland. The meteorological data include declination of the sun, mean, minimum and maximum air daily temperature, which are available from two stations, in addition to the river runoff measured in a single gauging station. The river freezing and melting processes that occur during winter months in the catchment pose a major problem for stream temperature forecasting.
