Abstract. When applying conceptual hydrological models using a temperature index approach for snowmelt to high alpine areas often accumulation of snow during several years can be observed. Some of the reasons why these "snow towers" do not exist in nature are vertical and lateral transport processes. While snow transport models have been developed using grid cell sizes of tens to hundreds of square metres and have been applied in several catchments, no model exists using coarser cell sizes of 1 km 2 , which is a common resolution for meso-and large-scale hydrologic modelling (hundreds to thousands of square kilometres). In this paper we present an approach that uses only gravity and snow density as a proxy for the age of the snow cover and land-use information to redistribute snow in alpine basins. The results are based on the hydrological modelling of the Austrian Inn Basin in Tyrol, Austria, more specifically the Ötztaler Ache catchment, but the findings hold for other tributaries of the river Inn. This transport model is implemented in the distributed rainfall-runoff model COSERO (Continuous Semidistributed Runoff). The results of both model concepts with and without consideration of lateral snow redistribution are compared against observed discharge and snow-covered areas derived from MODIS satellite images. By means of the snow redistribution concept, snow accumulation over several years can be prevented and the snow depletion curve compared with MODIS (Moderate Resolution Imaging Spectroradiometer) data could be improved, too. In a 7-year period the standard model would lead to snow accumulation of approximately 2900 mm SWE (snow water equivalent) in high elevated regions whereas the updated version of the model does not show accumulation and does also predict discharge with more accuracy leading to a Kling-Gupta efficiency of 0.93 instead of 0.9. A further improvement can be shown in the comparison of MODIS snow cover data and the calculated depletion curve, where the redistribution model increased the efficiency (R 2 ) from 0.70 to 0.78 (calibration) and from 0.66 to 0.74 (validation).

No abstract

ZusammenfassungDa die Leistung eines Laufkraftwerks ohne Schwallbetrieb nicht gesteuert werden kann, sind möglichst präzise Leistungsprognosen nötig, um die generierte elektrische Energie bestmöglich am internationalen Strommarkt verwerten zu können. Derzeit befindet sich beim österreichischen Wasserkraftwerksbetreiber Verbund AG für den Zweck der Leistungsprognose eine Kombination aus hydrologischen und hydrodynamischen Modellen (PW) im operativen Betrieb, welche aber insbesondere bei an- sowie absteigenden Leistungsverläufen noch Defizite aufweist. Deshalb wird in dieser Studie an den Laufkraftwerken Braunau-Simbach, Aschach und Greifenstein das Potenzial von Machine Learning (ML) Verfahren bei der kurzfristigen (bis 4 h) Leistungsprognose in fünf hydrologisch interessanten Zeitfenstern eruiert. Dafür werden gemessene Abfluss- und Leistungswerte von stromauf liegenden Laufkraftwerken und Pegeln als Eingangsdaten herangezogen. Die erzielten Ergebnisse zeigen, dass ML im Anwendungsbereich der kurzfristigen Leistungsprognose innerhalb einer Laufkraftwerkskette sinnvoll eingesetzt werden kann. So konnte beim Grenzkraftwerk Braunau-Simbach der Modellfehler in Form der Wurzel der mittleren quadratischen Abweichung (RMSE) im Vergleich zu PW bei der 4‑Stunden-Prognose sowie über die fünf ausgewählten Zeitfenster um rund 63 % verringert werden. Beim Kraftwerk Aschach wurde eine Reduktion von 30 % erzielt, während beim Kraftwerk Greifenstein der RMSE mit ML um mehr als 50 % reduziert wurde. Es hat sich bei ML zudem gezeigt, dass mit kürzerer Prognosezeit auch die Prognosequalität deutlich verbessert wird, während sich diese bei PW in einem deutlich geringeren Ausmaß mit der Prognosezeit ändert. Es ist daher absehbar, dass ab einer bestimmten Prognosezeit PW gegenüber ML wieder im Vorteil ist. Nichtsdestotrotz könnte bei längerer Vorhersagezeit aber durch die Nachkopplung eines ML-Modells an PW die Prognosequalität weiter verbessert werden.

Abstract. When applying conceptual hydrological models using a temperature index approach for snowmelt to high alpine areas often accumulation of snow during several years can be observed. Some of the reasons why these "snow towers" do not exist in nature are vertical and lateral transport processes. While snow transport models have been developed using grid cell sizes of tens to hundreds of square meters and have been applied in several catchments, no model exists using coarser cell sizes of one km2. In this paper we present an approach that uses only gravity and snow density as a proxy for the age of the snow cover and land-use information to redistribute snow in the catchment of Ötztaler Ache, Austria. This transport model is implemented in the distributed rainfall–runoff model COSERO and a comparison between the standard model without using snow transport and the updated version is done using runoff and MODIS data for model validation. While the signal of snow redistribution can hardly be seen in the binary classification compared with MODIS, snow accumulation over several years can be prevented. In a seven year period the classic model would lead to snow accumulation of approximately 2900 mm SWE in high elevated regions whereas the updated version of the model does not show accumulation and does also predict discharge more precisely leading to a Kling–Gupta-Efficiency of 0.93 instead of 0.9.

Canada 1 Water is a 3-year governmental multi-department-private-sector-academic collaboration to model the groundwater-surface-water of Canada coupled with historic climate and climate scenario input. To address this challenge continental Canada has been allocated to one of 6 large watershed basins of approximately two million km2. The model domains are based on natural watershed boundaries and include approximately 1 million km2 of the United States. In year one (2020-2021) data assembly and validation of some 20 datasets (layers) is the focus of work along with conceptual model development. To support analysis of the entire water balance the modelling framework consists of three distinct components and modelling software. Land Surface modelling with the Community Land Model will support information needed for both the regional climate modelling using the Weather Research & Forecasting model (WRF), and input to HydroGeoSphere for groundwater-surface-water modelling. The inclusion of the transboundary watersheds will provide a first time assessment of water resources in this critical international domain. Modelling is also being integrated with Remote Sensing datasets, notably the Gravity Recovery and Climate Experiment (GRACE). GRACE supports regional scale watershed analysis of total water flux. GRACE along with terrestrial time-series data will serve provide validation datasets for model results to ensure that the final project outputs are representative and reliable. The project has an active engagement and collaborative effort underway to try and maximize the long-term benefit of the framework. Much of the supporting model datasets will be published under open access licence to support broad usage and integration.