Towards automated seasonal river discharge ensemble forecasts on a federated compute and data infrastructure

The wealth of available global datasets at high spatial and temporal resolutions opens many opportunities for hydrological modelling and forecasting. It is now possible to provide high-resolution hydrological simulations for a river basin anywhere in the world, even in basins without in situ observations. Combined with the strength of global parameter estimations of the wflow_sbm concept (Imhoff et al, 2021), this allows us to build and run hydrological models without calibration. These models can ultimately be used to provide discharge forecasts for the seasonal time scale. Here we present a workflow that tests the interoperability, scalability, and performance of combining cloud and high-throughput compute and data resources. The workflow combines open source technologies, including containerization, to provide automated monthly river discharge forecasts for practically every basin on the globe on cloud, HTC and in the future HPC platforms. We leverage the global ERA5 and SEAS5 products from the Copernicus Climate Data store as input for the wflow hydrological model. The workflow automatically downloads the required input data for the model domain, resamples the data to the required model grids, and runs the simulations. The workflow is automatically triggered every month when new SEA5 forecasts become available. Prior to running the forecasts, the ERA5 files are used to update the hindcast model states in preparation for the forecast. Next, 50 wflow ensemble members, forced using the SEAS5 forecasts, are run in parallel to provide estimates on the probability of discharge events. The workflow is currently set up and running for the Rhine basin on SURF’s high-throughput computing platform, but can easily be deployed on different infrastructures and for different river basins.