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 hydrologically diverse Trinity River is born in the mountainous, snow-capped Trinity “Alps” region where flows rush through canyons amid steep hillslopes. The river continues downstream to mid-elevations, picks up Coffee Creek, and flows into Trinity Lake and Dam. The lower subwatersheds below the dam are distinct from the headwaters and are characterized by wide, agricultural valleys rich in minerals. The Trinity is the largest tributary into the Klamath River system, though before it reaches the confluence a significant volume is diverted via the Central Valley Project that transfers significant volume into the Sacramento River.
The forecast point for the Rodeo competition is in the upper Trinity River before it meets Coffee Creek, marked by the USGS 11523200 streamflow gauge (see the maps below). This part of the Trinity River forms the main stem into Trinity Lake, and is the second greatest flow contributor to the Central Valley Project.
For a wonderful full history, visit the Trinity River's community website and follow along their timeline.
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 Rouge River basin that we highlighted last month, the spring snowmelt season is the most active and most important for establishing the baselines for the many uses of the Trinity river system. This period of the live competition is just beginning, so we’ve evaluated our model based on previous years to understand its performance. To do this, we produce reforecasts and then quantify the model’s performance against observations that it has not used or seen.
The hydrographs below highlight the active 2019 spring melt period from March through June 30th, showing the 24-hr (top) and 48-hr (bottom) ahead mean predictions (green), 50% and 90% confidence intervals, and observations (black).
While the full CEATI competition continues to run until October 2021, the statistics from a one year reforecast validation period suggest that the model understands hydrologic behavior in the upper Trinity basin. 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 = 0.43 and KGE = 0.26.
Over the life of the CEATI competition thus far [10/1/2020 to 4/27/2021], in three of the four metrics tracked by the competition, HydroForecast is in first place in NSE, normalized root mean squared error, and Correlation Coefficient metrics over the 1-3, 4-10 and 1-10 day lead-time windows. Since the competition began, HydroForecast’s 1-10 day lead time NSE is 0.90 and the Correlation Coefficient is 0.95.
In California, droughts are a part of the hydrologic story. The record setting 2012 - 2016 drought cost California over a billion dollars in economic loss from agriculture and dairy production, and created prime conditions for wildfires and landslides. Read how scientists tied this drought's intensity to human-induced global warming. Unfortunately, evidence suggests that this was not a fluke: climate change is intensifying the magnitudes and durations of droughts. This has especially pronounced consequences for the California economy, which is so tightly linked to access to water (e.g. agriculture, hydropower, ecological preservation, tourism, dairy and even cannabis production).
As we write this blog during the spring season, the big worry is not whether there will be a drought in 2021 but rather the extent of its severity. As of April 26, the California Department of Water Resources reported that Trinity Lake’s water level is only at 66% of its average for this time of year. The 2021 Water Supply Outlook estimates that streamflow in the Trinity River from April - July will be 53-55% of the 30-year average. Below is the hydrograph for the spring season so far. Note the magnitude differences between this season and the 2019 season referenced earlier. At this time in 2019, flows had reached 5,000 cfs and were hovering around a 1,000 cfs baseline. In 2021 to date, the flows have yet to exceed 750 cfs. Some rivers are expected to be so dry and warm that the state will truck millions of young salmon to the Pacific Ocean to save them.
HydroForecast tracks snowpack and the timing of melt by capturing dynamic land surface conditions through its satellite and meteorological inputs. The two satellite images below from the Normalized Difference Vegetation Index (NDVI) depict the transformation of the basin from a wintery, snow-covered scene to well-vegetated forest cover. With these near daily NDVI images, the model learns how snowmelt and green up patterns track with patterns in streamflow.
The time series below show three important inputs into HydroForecast over part of this year’s thaw period: air temperature from two sources, NDVI, and Normalized Difference Snow Index (NDSI). The model receives this data and learns how each of these inputs influence and are influenced by the presence of one another, e.g. the relationship between vegetation cover and snow cover.
From these inputs, we note several key patterns: 1) as air temperatures begin to steadily fluctuate above freezing, the NDSI (snow cover) decreases and NDVI (vegetation) increases; 2) snow cover decreases rapidly once temperatures sustain above freezing - and vegetation growth lags this melting. As vegetation growth ramps up and precipitation falls as rain, the model learns to infiltrate and produce runoff during the spring season.
An interesting question this data can help answer is whether we can detect differences in spring patterns between a dry year and a wet year. We explored this by checking out these same three inputs (temperature, NDVI, NDSI) over the 2019 spring season - a year where snowpack was ~135% of average. Below are the same model inputs over the 2019 spring period. We notice two key points:
Spring is underway and we are excited about HydroForecast’s performance as snow melts in the upper Trinity basin. Top of mind for us is how HydroForecast can help reduce the devastation of drought by predicting the likelihood of low snowpack and drier conditions earlier and more accurately than existing models. This enables the state to have more time and better information for planning as they enact policies and measures to protect freshwater.