Evaluation of homogenization methods for seasonal snow depth data in the Austrian Alps, 1930-2010.

Despite the importance of snow in alpine regions, little attention has been given to the homogenization of snow depth time series. Snow depth time series are generally characterized by high spatial heterogeneity and low correlation among the time series, and the homogenization thereof is therefore challenging. In this work, we present a comparison between two homogenization methods for mean seasonal snow depth time series available for Austria: the standard normal homogeneity test (SNHT) and HOMOP. The results of the two methods are generally in good agreement for high elevation sites. For low elevation sites, HOMOP often identifies suspicious breakpoints (that cannot be confirmed by metadata and only occur in relation to seasons with particularly low mean snow depth), while the SNHT classifies the time series as homogeneous. We therefore suggest applying both methods to verify the reliability of the detected breakpoints. The number of computed anomalies is more sensitive to inhomogeneities than trend analysis performed with the Mann-Kendall test. Nevertheless, the homogenized dataset shows an increased number of stations with negative snow depth trends and characterized by consecutive negative anomalies starting from the late 1980s and early 1990s, which was in agreement with the observations available for several stations in the Alps. In summary, homogenization of snow depth data is possible, relevant and should be carried out prior to performing climatological analysis.

Coupling hydrological modeling and support vector regression to model hydropeaking in alpine catchments.

Water management in the alpine region has an important impact on streamflow. In particular, hydropower production is known to cause hydropeaking i.e., sudden fluctuations in river stage caused by the release or storage of water in artificial reservoirs. Modeling hydropeaking with hydrological models, such as the Soil Water Assessment Tool (SWAT), requires knowledge of reservoir management rules. These data are often not available since they are sensitive information belonging to hydropower production companies. In this short communication, we propose to couple the results of a calibrated hydrological model with a machine learning method to reproduce hydropeaking without requiring the knowledge of the actual reservoir management operation. We trained a support vector machine (SVM) with SWAT model outputs, the day of the week and the energy price. We tested the model for the Upper Adige river basin in North-East Italy. A wavelet analysis showed that energy price has a significant influence on river discharge, and a wavelet coherence analysis demonstrated the improved performance of the SVM model in comparison to the SWAT model alone. The SVM model was also able to capture the fluctuations in streamflow caused by hydropeaking when both energy price and river discharge displayed a complex temporal dynamic.