Abstract
Short-term streamflow forecasting is crucial for many activities including flood risk mitigation, multi-use water management and production planning for hydropower plants. Several studies have shown that data-driven approaches are more suitable for operational forecasting than physics-based models due to the higher flexibility and reproducibility of the former. Thus, weather data becomes the key information to be fed to machine learning methods for a reliable streamflow prediction. At this aim, we propose a data-driven framework for the assessment of one day ahead streamflow forecasting using as input the data provided by the Deutscher Wetterdienst (DWD) ICON climate forecasting model. The suitability of the ICON-D2-EPS ensemble regional climate model outputs is tested in the challenging problem of hourly forecasting in a complex alpine environment. The proposed data-driven methodology consists of two main steps, which are the bias correction of the weather forecast data followed by the streamflow machine learning algorithm. Long historical time series of precipitation and temperature provided by weather stations have been used for training the machine learning algorithm, while ICON prediction outputs have been adopted as testing. The ICON data were adjusted through an hourly bias correcton algorithm to make these data consistent with historical training data. Instead, the prediction of the hourly flow rate is calculated by an algorithm based on support vector regression. The performance of ICON data into this hydrological forecasting framework shows promising results for 24-hours streamflow prediction even in the context of small catchments with complex orography.