Abstract
The aims of the research project can be summarized in three points: to validate different type of data as input of the hydrological models, to develop a near-operative pipeline for the short-term streamflow forecasting and the analysis of uncertainties through the pipeline up to the hydrological outcomes. The first part is dedicated to data analysis, both meteorological data and hydrological data. Ground stations local data were considered as reference after statistical analysis and imputation to clean the sensors issues. The meteorological analysis for precipitation and temperature have been carried out using a gridded dataset obtained as interpolation of the local ground stations covering South Tyrol. The reanalysis data (ERA5-Land) and the forecast data (DWD ICON) have been compared with the interpolated local data, identifying the biases. A literature review of bias correction methods applied in hydrology were crucial to understand and apply a reduction of uncertainty on meteorological data. The bias correction methodology were used in different case studies and on different data types, to underline the quality of this tools. Indeed, reanalysis data and forecast data were corrected against interpolated ground stations gridded datasets, using both univariate and multivariate methods. To check the quality of the correction, these data were used as forcing for the hydrological models, physical or data-driven depending on the application. The second part is instead dedicated to the data-driven model approach. The fully data-driven model were applied to the entire South Tyrol region. The lumped approach is showing promising results and is based on the Support Vector Regression models (SVR), resulting as the best algorithm in the streamflow forecasting in the short-term against Long Short Term Memory (LSTM) and Feed Forward (FF) models. SVR combined with the bias corrected forecast data (DWD ICON-D2) as meteorological inputs, is a fully operational tool for the streamflow forecasting up to 48 hours of lead time. This operative chain download and extract the data with the 20 ensemble members, it applies bias correction to align the trained data-driven model to operative data, and finally it runs the models in forecasting to obtain the streamflow at the specified point. The outcome of this process are 20 different simulations, that represents the meteorological uncertainty. On the other hand, to introduce the models uncertainty this setup have been investigated better. Indeed, the use of different models in the chain allows the analysis of models uncertainties, carrying out some considerations applying the approach to two different basins in South Tyrol. The first result of this PhD project is the creation of the meteorological-hydrological pipeline, allowing extensive research applying in unknown points the approach to obtain a streamflow short-term forecasting at the hourly time resolution. This is possible using different types of data, raw or bias corrected, different types of data-driven approaches (SVR or FF), and the use of only precipitation and temperature it is demonstrated to reach already good results. Furthermore, the models as developed in the last version allow the use of more meteorological input points, that could lead to increase the performance of the pipeline, reducing the uncertainties. The fully datadriven approach developed in this PhD project is feasible in practical applications, such as hydropower energy producers in energy field or local protection agencies to predict floods for safety reasons. These decision-making process is supported by the uncertainties propagated through the pipeline, allowing the explainability of how the hydrological outcomes are influenced by the different errors.