Considered by President Teddy Roosevelt as “one of the most scenic stretches of river in America,” the untamed North Fork Shoshone is a classic mountain-fed, cold water, rapid flow river - a fly fishers’ dream. These headwaters of the Shoshone provide many recreational activities and serve as the foundation for downstream users who rely on the waters for municipal supply and irrigation. After the North Fork connects with the South Fork to form the mainstem Shoshone River in Northwest Wyoming, the river flows north and east into the Big Horn River and then into the Yellowstone River before eventually entering the Missouri River.
The Streamflow Forecast Rodeo is a competition hosted through a partnership between the Bureau of Reclamation and the Centre for Energy Advancement through Technological Innovation (CEATI)'s Hydropower Operations and Planning Interest Group. The Rodeo website summarizes: "[This] challenge seeks to improve the skill of short-term streamflow forecasts (10 days) via a year-long competition." Check out the map below to see the 19 hard-to-forecast sites selected for the competition.
Each month we’ll shine a spotlight on a different site, review HydroForecast’s overall performance, and take a look at interesting events.
Considered by President Teddy Roosevelt as “one of the most scenic stretches of river in America,” the untamed North Fork Shoshone is a classic mountain-fed, cold water, rapid flow river - a fly fishers’ dream. These headwaters of the Shoshone provide many recreational activities and serve as the foundation for downstream users who rely on the waters for municipal supply and irrigation. After the North Fork connects with the South Fork to form the mainstem Shoshone River in Northwest Wyoming, the river flows north and east into the Big Horn River and then into the Yellowstone River before eventually entering the Missouri River.
The forecast point for the Rodeo competition is the USGS station 06279940 at Wapiti, WY, above the Buffalo Bill reservoir and dam. The streamflow into this gauging station is primarily snowmelt driven with very little precipitation the remainder of the year. This attribute makes forecasting peak volumes and timing ever more important to informing water availability for the host of downstream uses. Baseflow in the non-snowmelt season is important for aquatic and riverine habitat as well as continuing to feed the Buffalo Bill storage reservoir with dry season inflow.
HydroForecast provides state-of-the-art, accurate streamflow forecasts using a hybrid approach that combines physical science with artificial intelligence. HydroForecast offers a range of advantages over existing forecasting techniques, and we've joined the CEATI competition in order to exhibit, live, these strengths. Under the hood, every forecast is created by an ensemble of neural networks that are provided different members of meteorological forecast ensembles. HydroForecast is rapid to deploy in a new basin and resilient to basin and climatic changes.
Unlike the James River basin that we highlighted last month, the active period in the Shoshone is solely the snowmelt-driven pulse in the spring. This pulse essentially sets flow expectations through the basin for the remainder of the year. As this spring period of the live competition comes to an end, we reflect on how this year compares with others and how well HydroForecast was able to predict the timing and magnitude of this peak. To understand the performance of the model over the recent years, we evaluated it from 2019-present.
The hydrographs below highlight March 2019-May 2021, showing mean predictions (orange) for the model at 24-hr (top), 48-hr (middle), and 120-hr (bottom) ahead, the 50% and 90% confidence intervals (blue bands), long term median (purple), and USGS observations (black).
We notice a few key patterns from these plots:
While the full CEATI competition runs until October 2021, we computed the statistics so far from the start of our forecasts in March 2019 to present. Note that perfect scores for NSE and KGE are equal to one, while an ideal bias score is zero. As a comparison point, the long-term median (i.e. climatology) has NSE = 0.76 and KGE = 0.67, which is decent but carries much lower predictive skill than HydroForecast.
HydroForecast holds high predictive power even five days ahead, maintaining a very low bias (< 1%) and the ability to model high and low flows (illustrated in the hydrographs above).
Over the life of the CEATI competition thus far [10/1/2020 to 6/25/2021], HydroForecast is the most accurate forecast available for this portion of the North Fork Shoshone River. HydroForecast is in first place in all four of the metrics tracked by the competition (NSE, normalized root mean squared error, Correlation Coefficient and bias). This holds across various splices of the lead time windows: 1-5 days, 6-10, and 1-10. HydroForecast especially shines at the longer lead time windows (e.g. 5+ days) compared to other forecasts in the competition.
Given that snowmelt is the most important time period for this basin’s hydrology, we will dig into this part of the hydrograph. One distinct signal of snowmelt that appears in the streamflow hydrograph is the diurnal (twice daily) pulsing illustrated in the observed time series (blue line) below.
Scientists have measured and modeled these systems to understand this hydrologic process in snow-fed basins. Within a single day, the radiative energy budget causes snowmelt to peak in the afternoon as the sun’s radiative heat and temperatures are highest. Depending on the size of the snowpack and characteristics, the liquid water percolates and transports through the snow at different rates and therefore reaches the streams at different times. This travel time of liquid water from snowpack to stream is of great interest and importance because it strongly influences stream, soil, and groundwater storage in these snowy headwaters basins. One study found that night-time temperatures control the melt rates during the early spring, when cold night temperatures create a ‘barrier’ to melt the snowpack that must be overcome each new day. With shifts to warmer night-time temperatures occurring faster than day temperature changes, this poses a question as to how streamflow and water storage will be impacted.
Generally, scientists are keeping their eyes on shifts in the diurnal cycle signals within a given basin, because they can indicate underlying changes in snowpack stability and cryospheric processes. To detect these dynamics, it is essential to have updated data and measurements that can monitor the near real-time snowpack and soil conditions.
While daily models are too coarse to detect the diurnal pulse, hourly models can pick up the pattern given the necessary inputs. Below illustrates a short time series of streamflows in the North Fork Shoshone river, where HydroForecast has learned to predict the diurnal signal in its forecasts. HydroForecast traces from 5/29, 5/30, 5/31, and 6/01 are shown in the red, blue, and brown lines. The USGS observations (black line) laid overtop show the agreement in both magnitude and timing of the diurnal pulse. Notice that using the long-term median (purple - dotted line) does not capture this behavior.
One important input that the model receives to help learn the diurnal pulse is the forecasted air temperature from two different weather sources. Below is an image of the temperatures (in Celsius) over the same short time period, illustrating both the clear diurnal signal and also the spread in the predictions between ECMWF and GFS. Traces from 5/29, 5/30, 5/31, and 6/01 are shown. The selected time window is particularly critical as night-time temperatures shift to consistently above freezing (0 degrees). From this point forward, sustained melting occurs during the day and at night.
HydroForecast tracks snowpack and the timing of melt by capturing dynamic land surface conditions through its satellite and meteorological inputs. By feeding HydroForecast with near-real time images of the basin conditions, the model learns to detect changes in the land surface states, most pronounced of which is the transformation from snow covered to vegetated. With these near-daily Normalized Difference Snow Index (NDSI) and Normalized Difference Vegetation Index (NDVI) images along with temperature, precipitation and other weather inputs, the model partitions precipitation into rain or snow and tracks soil moisture and runoff states. Let’s view this melt out and green up through the eyes of the NDVI (left) and NDSI (right) satellite images from Sentinel-2 over the course of the 2021 winter and spring seasons. In the NDVI image, the transition from brown to green denotes vegetation growth; in the NDSI image, white is snow and blue is no snow.
A few highlights to note:
We are excited about HydroForecast’s spring snowmelt season performance in the North Fork Shoshone River. In basins like these, we are thinking of the benefits HydroForecast can have on water resources management planning and allocation decisions given the many uses of the river for downstream agriculture, recreation, and municipal supply. Forecasts are an essential piece of this planning, considering the potential volatility in spring flow magnitudes (in the last two years we saw an order of magnitude difference). Since the hydrology and streamflow volumes are so closely tied to snowpack, we are paying close attention to changes in the melting patterns and rates that may hold the key to the future of this basin’s hydrology.
Stay tuned!