Interpolate, simulate, assimilate: operational aspects of improving hydrological forecasts in the Rhine basin

Hydrological forecasts are a useful and cost-effective tool to aid decision making. Hydrological forecasts are based on a set-up consisting of several model and data components, which need to be integrated for an effective forecast. Part of this model train is the hydrological model. Regarding hydrological models, there is a current trend to move towards high-resolution spatially distributed gridded models, both in the wider literature and in the specific case ofthe forecasting system RWSosRivers of Rijkswaterstaat. Concurrently, the forecasting models are used for ensemble forecasts and are further improved with data assimilation techniques.

A number of constraints particular to the operational forecasting chain are identified:

  - Near real-time availability of data;
  - Continuous operation;
  - Processing time;
  - Desicion making under uncertainty.

The combination of the movement towards gridded models, ensemble forecasts and the inclusion of data assimilation leads to new challenges, conceptually, but even more so when intersected with operational practice. The challenges that arise on the intersection of hydrological science disciplines and the constraints above are called in this thesis operational aspects.

The thesis consists of five research chapters that each contribute original research on op- erational aspects of hydrological forecasts, with the Rhine basin as case study.

In Chapter 3 , the operational aspect studied was precipitation interpolation under the constraint of limited data availability. Specifically, limited network density of hourly realtime reporting gauges and the constraint of reliability in the sense that any interpolation method should also work in case of missing data. Collected rain gauge data were spatially interpolated using the genRE method, which uses climatological grids to inform the interpolation about the spatial distribution of precipitation. It was shown that improved hourly interpolation results could be obtained with the operational interpolation method by using climatological grids which were derived from non-operationally available daily data with a higher quality and quantity, thereby successfully using non real-time data to enhance the information available in real-time.

The subject of Chapter 4 was the treatment of potential evaporation (PET) in hydrological forecasting. The operational aspect that was studied was the choice between using average PET climatology based on a long time series of offline data, or the use of near real-time available data to calculate PET in near real-time, including calculating PET from forecast data for use in hydrological forecasts. A 20-year reforecast was done and the resulting skill scores with online and offline PET data were compared. It was shown that for our case there was a negligible difference in discharge forecast skill between using the offline or online PET data for forecasts up to 10 days.

Chapter 5 reported the set-up of the wflow_sbm model concept without calibration. Pa- rameters for the model were derived based on pedotransfer functions found in literature and open access spatial data, such as soil properties. The model parameters were subsequently scaled from the highest data resolution to several coarser model resolutions. It was shown that this resulted in adequate modelling results of discharge throughout the Rhine basin, as well as flux conservation between modelling on different spatial scales. The scalability and uncalibrated properties of the model fit the use in an operational context very well.

Last, Chapters 6 and 7 showed the results of two experiments with state updating. Chapter 6 employed the Ensemble Kalman Filter (EnKF) to investigate if measured lake water levelscouldbeusedtoimprovedownstreamdischargeforecasts. Theresultsshowedthat indeed state updating with lake level measurements can aid in making discharge fore- casts, but that the robustness of this state updating is highly dependent on the quality of the lake modelling.

Chapter 7 applied the Asynchronous Ensemble Kalman Filter (AEnKF) to assimilate dis- charge into the gridded hydrological model at the subbasin level. The results of the as- similation were compared with ARMA postprocessing. It was shown that the ARMA cor- rectionwasverystrongforthefirsttwodays, butthatthe AEnKFprovidedmoreconsistent improvements for longer lead times.

Based on these results, specific advice is given to improve the forecasting system for the Rhine. It is recommended to:

• Apply the methods used in this thesis for the spatial preprocessing of the forcing data and to keep an up-to-date database containing all the data that is needed for a reforecast analysis;

• Implement the data assimilation set-up as used in Chapter 7, with the annotation that this method can be further optimized for individual subbasins;

• Make the decision to base further developments in the hydrological model on the wflow model framework, under the vision of ‘choose a modelling framework first, and hydrological model concept later’;

• Invest in the set-up of an automated benchmarking system. This advice fits in a broader picture on the future of state-of-the-art forecasting systems. Such forecasting systems are likely to:

• Be modular in set-up with interchangeable components;

• Have integrated uncertainty estimation, with coupled models of hydrology, meteorology and other environmental models;

• Include the user as integral part of the forecasting chain and include social elements in the forecast;

• Apply improved integration of commensurable data sources throughout and up and- down the forecasting chain;

• Be multi-purpose and cater to a wide range of clients with different needs;

• Be automated learning systems that utilize standard benchmarks to continuously improve modelling;

• Be operated by consortia of parties that have the necessary expertise and resources.