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.
The headwaters of the James River originate in prairie ponds in central North Dakota and flow eastward, then south into Arrowwood Lake. The basin is characterized by flat croplands and rangelands that surround the glacially carved channel. As the river flows south, a series of flood control and water supply impoundments serve to protect Jamestown and supply nearby rural communities with municipal and irrigation water. Beyond Jamestown, the river slowly meanders over flat terrain through South Dakota before joining the Missouri River. With an average slope of about 0.2 foot per river mile, it is considered one of the flattest rivers in North America of a similar length.
The forecast point for the Rodeo competition is in the headwaters before the river passes into Arrowwood Lake, marked by the USGS 06468250 James River above Arrowwood streamflow gauge (see the location map below). A Natural Resources Conservation Service report estimates that this portion of the river serves over 700 farms and 10-15 animal feeding facilities. Flows in these headwaters are critical for bird habitat and downstream operations in the Bureau of Reclamation’s Jamestown dam, which was built primarily for flood control but also serves municipal supply, recreation, and irrigation. The basin image below shows the landscape reflecting the heavy human land use activities described above.
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.
Like the Trinity River basin that we highlighted last month, the active periods in the James River are seasonal, with one snowmelt-driven pulse in the spring that is critical for the ecosystem and human activities in the basin. This spring period of the live competition is in full swing, and it is interesting to compare where flows are this year with previous years. 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 the 24-hr (top) and 48-hr (bottom) ahead mean predictions (yellow), 50% and 90% confidence intervals, and 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) had much lower predictive skill with NSE = -0.21 and KGE = -0.52.
HydroForecast holds high predictive power even five days ahead, maintaining a relatively low bias (within 10%) and ability to model high and low flows (illustrated in the plots above).
Over the life of the CEATI competition thus far [10/1/2020 to 5/22/2021], 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 all splices of the lead time windows: 1-3 day, 6-10, and 1-10. Given that the observations at this site are ice-affected from mid-November to mid-March, these statistics were calculated on a limited sample and generally reflect HydroForecast’s ability to predict low flows and the start of spring melt (or lack thereof). Yet this year’s spring is different from many others in recent history, presenting a new challenge for HydroForecast.
Flood control was the primary reason the Jamestown dam was constructed in 1952. Since then, numerous high inflow and threatening events into Arrowwood Lake upstream of the dam and town have occurred. If we examine the 35 years of observed annual peak flows (see graph below) reported at the USGS 06468250, the flood of record (8,470 cfs) on 04/18/2009 was nearly exceeded by the spring flows two years later (7,800 cfs on 04/12/2011). The 2009 flood was partially driven by a much larger than average annual snowfall -- 88” compared to the normal 34” -- which caused downstream flows to overtop Jamestown dam for the first time on record.
The James River is no stranger to droughts either. Note in the plot above that the floods of 2009 followed one of the driest years on record in 2008. Evidence of another low flow period this year is mounting, and with it are expected impacts on agriculture and fish habitat. With flows in James River above Arrowwood hovering between 0.5-2 cfs, the USGS StreamStats monitoring website is reporting daily lows at this site near the time of writing. Years of low flow are particularly challenging for meeting water quality standards, when dilution is not an option but agricultural practices continue to add nutrients to streams.
Just how low are the flows? Take a look at the plots below, zooming in at the 48-hour lead time on this water year so far (10-01-2020 to present). The top graph shows the long term median and the bottom graph does not.
Notice that although the April spike normally expected is absent this year, HydroForecast predicted this low flow pattern quite well.
What kind of information does the model track that helps it predict low flows? The time series below zooms into the 2021 spring period, where we’ve already noted the absence of an important seasonal peak. The plots below show three important inputs into HydroForecast over the spring period: precipitation, air temperature, and Normalized Difference Snow Index (NDSI).
HydroForecast tracks snowpack and the timing of melt by capturing dynamic land surface conditions through its satellite and meteorological inputs. With these near-daily NDSI images, along with temperature, precipitation, and vegetation indices, the model learns how to partition precipitation into rain or snow and determine whether it will lead to infiltration or runoff. The hydrograph below shows streamflow for the same time period.
Combining the input time series with the hydrograph, we can piece together a story to explain why we are seeing low flows.
These factors all contribute to the current dry conditions in the upper James River, captured by HydroForecast. Let’s hope this story continues with a wetter summer on the horizon.
Spring is turning to summer and we are excited about HydroForecast’s performance in the James River so far in the CEATI competition. In basins like these, we are thinking of the impacts HydroForecast can have on water resources management planning and allocation decisions given the many uses of the river for agriculture, habitat and downstream hydropower. Since water quality is a known issue in the James River, having accurate foresight into the near-term hydrologic conditions in historically dry years can help mitigate poor water quality.