Streamflow Prediction

Background

Scientists and engineers in RAL’s Hydrometeorological Applications Program at the National Center for Atmospheric Research are undertaking research to facilitate the transition of advances in land surface modeling, weather and climate prediction and downscaling, data assimilation, post-processing and other areas into operational streamflow forecasting practice in the US.  Projects supporting this overall objective are described below. 

Assessing the Viability of Over-the-Loop Streamflow Forecasting to Support Risk-based Water, Energy, and Hazard Management

The U.S. Army Corps of Engineers and Bureau of Reclamation have jointly sponsored a multi-year project to evaluate and provide a real-time demonstration of the viability of new science-based techniques and strategies for real-time hydrologic flood and drought forecasting in support of real-time water decisions.  NCAR has collaborated with both agencies and the University of Washington to develop and run a fully automated streamflow forecast system called SHARP (System for Hydromet Analysis, Research, and Prediction). The system produces real-time ensemble streamflow predictions for lead times of days to seasons, using a variety of weather and climate forecast datasets, hydrologic models, statistical methods and other tools.  The project is focused on exploring ways to enhance the physical realism of real-time watershed monitoring and prediction, while still maintaining the computational agility needed to depict uncertainties (e.g., using ensemble techniques). The research will help us demonstrate the tradeoffs of making components of our nation’s hydrologic monitoring and prediction workflows more automated and objective, opening the door to advances such as more complex watershed models and uncertainty-aware products. The overarching goal is to strengthen our nation’s scientific foundation for operational hydrologic prediction to better manage resources and risks in the face of changing weather and climate extremes.

FY2017 Accomplishments

  • Running the SHARP system to explore new techniques for hydrologic data assimilation (DA), and in particular, the use of a particle filter to assimilate streamflow for improved hydrologic state estimation, and a novel analog-particle-filter approach to avoid DA challenges such as filter collapse. 
  • Running the SHARP system to explore new techniques for streamflow forecast post-processing, and in particular the use of a particle filter to assimilate streamflow for improved hydrologic state estimation, and a novel analog-particle-filter approach to avoid DA challenges such as filter collapse. 
  • Generating reservoir inflow hindcasts for the Klamath River basin and the Crooked River watersheds of Oregon to support a Bureau of Reclamation effort to develop forecast-informed reservoir operating rules.
  • Developing distributed regional watershed forecasting models using an implementation of the SUMMA model on the USGS Geospatial Fabric in the Pacific Northwest.
  • Initial explorations of the ability of the pseudo-dynamical Intermediate Complexity Atmospheric Research (ICAR) model for downscaling regional weather predictions. 
  • Expansion of test watersheds to include pilot basins in California.
  • Implementation of SUMMA-based SHARP outputs in an online display via the NCAR HydroInspector tool.

 

Publications

Mendoza, PA, AW Wood, EA Clark, E Rothwell, MP Clark, B Nijssen, LD Brekke, and JR Arnold, 2017, An intercomparison of approaches for improving predictability in operational seasonal streamflow forecasting, Hydrol. Earth Syst. Sci., 21, 3915–3935, 2017

Arnal, L., AW Wood, E. Stephens, H.L. Cloke, and F. Pappenberger, 2017: An Efficient Approach for Estimating Streamflow Forecast Skill Elasticity. J. Hydrometeor., 18, 1715–1729, https://doi.org/10.1175/JHM-D-16-0259.1

Clark, EA, AW Wood, and B Nijssen, 2017, Assessing ensemble particle filters for the estimation of model states for streamflow forecasting, Wat. Res. Rsrch. (in review).

Clark, E. A., AW Wood, B. Nijssen, and M. P. Clark. Implications of streamflow data assimilation via particle filter on streamflow forecasts in basins with seasonal snow, 2018. Hydrology and Earth System Sciences (in prep).

Mendoza, PA, AW Wood, EA Clark, N Voisin, B Nijssen, MH Ramos, 2017, An assessment of streamflow post-processing techniques for short-range ensemble streamflow forecasts, in prep.

Clark, E. A., A. W. Wood, B. Nijssen, and M. P. Clark, Analog Resampling For Particle Filter Data Assimilation In Hydrologic State Estimation, Hydrol. Earth Syst. Sci (in prep)

Figure 1 – Examples of SHARP forecast system elements and outputs.  Top left: An ensemble member of daily precipitation at the USGS Geospatial Fabric resolution using the SUMMA modeling implementation.  Top right: An over-the-loop ensemble forecast from the SHARP system for the Hungry Horse Reservoir, compared with a forecast from the River Forecast Center.  Bottom: A forecast display using the HydroInspector to show routed streamflow along the Geospatial Fabric.
Figure 1 – Examples of SHARP forecast system elements and outputs.  Top left: An ensemble member of daily precipitation at the USGS Geospatial Fabric resolution using the SUMMA modeling implementation.  Top right: An over-the-loop ensemble forecast from the SHARP system for the Hungry Horse Reservoir, compared with a forecast from the River Forecast Center.  Bottom: A forecast display using the HydroInspector to show routed streamflow along the Geospatial Fabric.

FY2018 Plans

  • Build out the SHARP real-time medium range forecast evaluation domain to the western US (Reclamation service area), including real-time 1/16th degree ensemble forcings. 
  • Apply forecasting techniques developed in pilot watersheds to regional modeling efforts using SUMMA.
  • Demonstrate the value of NCAR modeling and prediction tools via SHARP, including medium-range forecasts based on ICAR downscaling of GEFS, SUMMA and Mizuroute modeling and routing, and model calibration through the Multiscale Parameter Regionalization.
  • Continue to raise awareness and promote discussion about tradeoffs in modeling and streamflow forecasting approaches within the research, operational and management communities.

 

Improving Operational Flood Forecasts in the US Northern Plains Region through Assimilation of Ponded Water Retrievals

Better accounting for ponding within the land-surface water balance can improve predictability for runoff and streamflow, with consequent benefits for society through improved river forecasting and decision-making in water management and emergency response to flooding events.  There are currently no quantitative estimates of the volume of water detained on the landscape during such ponding events, and we lack a sufficiently comprehensive understanding of the phenomenon for practical enhancement of operational forecasting.   The upper Midwestern US terrain and land-use characteristics provide an ideal setting to demonstrate and achieve practical application of NASA remote sensing in the operational flood forecasting context. 

Figure 2: Schematic of the components of SHARP, a research system now running in real-time at NCAR to produce daily-updating ensemble flood forecasts (7-day lead times) and bi-weekly seasonal streamflow predictions for a small number of pilot basins in the Western US.
Figure 2: Schematic of the components of SHARP, a research system now running in real-time at NCAR to produce daily-updating ensemble flood forecasts (7-day lead times) and bi-weekly seasonal streamflow predictions for a small number of pilot basins in the Western US.

NCAR has teamed with scientists from Purdue and NASA, as well as forecasters from the NWRFC, to use satellite data primarily from the MODIS sensors (see below right) and LandSat to derive estimates of surface ponded water volume and extent, and use them to update real-time streamflow forecasting models.  It has upgraded the NASA Land Information System (LIS) -based VIC hydrology model to include surface ponding schemes, and is beginning to evaluate strategies for improving hydrologic simulation and prediction through data assimilation of the ponded water datasets.  

Improving Sub-seasonal to Seasonal Streamflow Predictions in the Lower Colorado and Rio Grande River Basins

Operational streamflow forecasts at daily-to-seasonal lead times are critical to Reclamation’s management of reservoirs in the Colorado River basin that store and allocate water to serve water needs worth billions of dollars annually in seven southwestern states.  These forecasts sometime fail to predict the real-time conditions that are later observed by Reclamation operators, which can lead to sub-optimal outcomes for decisions in water operations, especially during extreme events such as droughts or floods.  Recent studies show that changes in climatic conditions have resulted in changes to temperature and precipitation patterns throughout the West. Anecdotal evidence suggests that differences between streamflow forecasts versus observation is increasing, perhaps due to the fact that existing forecast methodologies must incorporate increasing variability and uncertainty, and extreme weather events.  In addition, some current forecast methods (such as regression-based forecasting) are to some extent dependent on the assumption that climate and weather patterns are stationary over multi-decadal periods, but this is traditional concept has been all but abandoned in the hydrometeorological sciences in recent years. 

To address these issues, RAL is working through a project co-funded by Reclamation and NOAA, “Postdocs Applying Climate Expertise” (PACE), to understand the implications of potential climate change in the US Southwest for subseasonal-to-seasonal (S2S) lead forecasts, and to apply this understanding to improve operation long-lead water supply predictions.  This effort features close collaboration with Reclamation water managers in the Lower Colorado and Rio Grande River basins, as well as operational forecasting staff at the National Water and Climate Center (NWCC).

Figure 3: (a) March-July mean temperature anomalies relative to 1982-2016 from observations, 40 CMIP5 models, and seasonal prediction models (NMME+ECMWF), averaged over the box indicated in Fig. 1a. The red line is the mean across NMME-ECMWF models, the gray line is the mean across CMIP5 models, and the black line is observations. Shading indicates the 5-95% range. (b) Correlation between observed and forecasted temperature for different temperature targets and seasonal prediction models for 1982-2016. Forecasts are initialized at the start of each predicted period. All correlations are significant at 95% confidence.
Figure 3: (a) March-July mean temperature anomalies relative to 1982-2016 from observations, 40 CMIP5 models, and seasonal prediction models (NMME+ECMWF), averaged over the box indicated in Fig. 1a. The red line is the mean across NMME-ECMWF models, the gray line is the mean across CMIP5 models, and the black line is observations. Shading indicates the 5-95% range. (b) Correlation between observed and forecasted temperature for different temperature targets and seasonal prediction models for 1982-2016. Forecasts are initialized at the start of each predicted period. All correlations are significant at 95% confidence.

This project has applied state-of-the-art climate forecasts from the NOAA NMME and the ECWMF System 4 to show that the effects of temperature non-stationarity on seasonal streamflow prediction may be offset by including forecasts that represent the warming as inputs to the streamflow forecast approach.  Work is proceeding to produce improved seasonal streamflow predictions for the Rio Grande basin and to operationalize some of the new forecasting approaches at NWCC and Reclamation offices.

Figure 4: (a) Absolute skill improvement of the temperature-aided forecast relative to the baseline forecast at individual gages for issue date 1 March as an illustrative example. (b) Absolute skill improvement for all gages as a function of issue date. (c) Relative skill improvement for all gages as a function of issue date. Solid lines are the median across (black) all gages and (colors) the three basins. Dashed line is the median across all gages when using observed temperature, mimicking the hypothetical case where the future temperature is known at the time of forecast issue, and dotted line is the median when using only the linear trend of observed temperature.
Figure 4: (a) Absolute skill improvement of the temperature-aided forecast relative to the baseline forecast at individual gages for issue date 1 March as an illustrative example. (b) Absolute skill improvement for all gages as a function of issue date. (c) Relative skill improvement for all gages as a function of issue date. Solid lines are the median across (black) all gages and (colors) the three basins. Dashed line is the median across all gages when using observed temperature, mimicking the hypothetical case where the future temperature is known at the time of forecast issue, and dotted line is the median when using only the linear trend of observed temperature.

Key Findings:

  • Seasonal temperature forecasts from climate prediction models are skillful over the headwaters of the Colorado and Rio Grande River basins
  • Adding temperature information to current operational seasonal streamflow forecasts in snowmelt-driven basins improves forecast skill
  • Temperature forecasts help mitigate impacts of non-stationarity on US Southwest streamflow predictability under increasing temperatures

 

Figure 5: Runoff ratio at the Rio Grande River at Otowi Bridge (colors) as a function of water year precipitation and annual mean temperature from (a) reconstructions, (b) observations, and (c) CESM control simulation (1,800 years total). All time series are relative to their median; in the case of observations, relative to the median of the reconstructions. The colored numbers give the percentage of very low (< 10th percentile), low (< 30th), high (> 70th), and very high (> 90th) runoff ratio years that fall within a given quadrant of precipitation and temperature anomalies.
Figure 5: Runoff ratio at the Rio Grande River at Otowi Bridge (colors) as a function of water year precipitation and annual mean temperature from (a) reconstructions, (b) observations, and (c) CESM control simulation (1,800 years total). All time series are relative to their median; in the case of observations, relative to the median of the reconstructions. The colored numbers give the percentage of very low (< 10th percentile), low (< 30th), high (> 70th), and very high (> 90th) runoff ratio years that fall within a given quadrant of precipitation and temperature anomalies.

Earlier, the project found, using a paleo analysis, that recent spring runoff declines in the Rio Grande Basin are highly unusual, and that warming temperature trends are contributing to this decline in efficiencies. Additional findings included that:

  • The decreasing runoff efficiency trend from 1986-2015 in the Upper Rio Grande River basin is unprecedented in the last 440 years
  • Very low runoff ratios are 2.5 to 3 times more likely when temperatures are above- normal than when they are below-normal
  • The trend arises primarily from natural variability but runoff sensitivity to temperature implies further declines should warming continue

 

Publications

Lehner, F, ER Wahl, AW Wood, DB Blatchford, and D Llewellyn, 2017, Assessing recent declines in Upper Rio Grande River runoff efficiency from a paleoclimate perspective, Geophys. Res. Lett., 44, doi:10.1002/2017GL073253.

Lehner, F., AW Wood, Llewellyn, D., Blatchford, D. B., Goodbody, A. G. & Pappenberger, F. (2017). Mitigating the impacts of climate non-stationarity on seasonal streamflow predictability in the US Southwest. Geophysical Research Letters, 44. https://doi.org/10.1002/2017GL076043

Sub-seasonal to Seasonal Climate Products for Hydrology and Water Management

This project's overarching goal is to improve the understanding and application of S2S climate forecast products in the hydrology and water management sector, and to create an operational initial product generation capability in this area at the NCEP Climate Prediction Center (CPC).

The potential value of the sub-seasonal to seasonal (S2S) prediction has not yet been fully realized by stakeholders in the water management applications sector.  Hurdles to adoption, in addition to low forecast skill, Aside from situations in which the S2S forecasts have low skill, include:

  • misalignment of forecast products with users' space-time analysis needs, which often follow specific watershed boundaries;
  • use of data formats that users cannot easily input to their analysis tools;
  • systematic biases in products relative to user climatologies; and
  • verification information that is tailored toward forecast producer diagnostic needs but is unusable or not relevant to water sector users.

 

In each of these areas, more work can be done to bridge the gap to potential stakeholders and enhance quality, specificity, and accessibility, and thus usability of S2S predictions. This project is addressing the hurdles described above and also investigating opportunities for skill enhancement through forecast post-processing as well as for developing products describing extremes at S2S time scales.

The project will use S2S reforecasts (including CFSv2 and NMME, and other forecast sources) and real-time forecasts to apply and assess various post-processing approaches that may enhance the skill and reliability of the raw climate outputs. It will develop verification data products characterizing the predictability of surface precipitation and temperature predictions at bi-weekly, monthly, and seasonal time steps over the CONUS domain, at watershed-focused USGS Hydrology Unit Code (HUC)-4 and other HUC-spatial units. The benefits of statistical post-processing will be assessed against benchmarks (or baselines) from raw model outputs. For the prediction of extremes potential, the project will apply spatial extremes models using hierarchical Bayesian framework with climate system covariates. A transition plan leading toward experimental operation of the forecast approaches within NOAA in FY18 has also been developed.

Figure 6:  S2S Climate Outlooks for Watersheds website (http://hydro.rap.ucar.edu/s2s/), showing real-time S2S forecasts based on the CFSv2 and NMME operational products.
Figure 6:  S2S Climate Outlooks for Watersheds website (http://hydro.rap.ucar.edu/s2s/), showing real-time S2S forecasts based on the CFSv2 and NMME operational products.

More information about this project can be found in a recent UCAR AtmosNews article (https://www2.ucar.edu/atmosnews/in-brief/129967/new-climate-forecasts-for-watersheds-and-water-sector).

Support for international streamflow forecast initiatives and programs

In 2017, NCAR personnel supported the following national and international streamflow and hydrologic forecasting efforts and conferences through organizational roles, participation, leadership, and/or advisory board membership.  These include:

  • Hydrologic Ensemble Prediction Experiment (HEPEX:  http://www.hepex.org/) – A. Wood, Co-Chair
  • US CLIVAR Predictability, Predictions, Applications Interface (PPAI) Panel – A. Wood, Co-Chair
  • European Union (EU) Improving Predictions and Management of Hydrological Extremes (IMPREX) – A. Wood, Advisory Board Member
  • ECMWF European Flood Awareness System (EFAS) and Global Flood Awareness System (GLOFAS) Project – A. Wood, Advisory Board Member
  • WMO Hydrological Status and Outlook System (HydroSOS) Initiative – A. Wood, Task Lead