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.
Join us on an adventure to North Central Idaho where elk roam the relatively uninhabited forests and lands, and native trout species meander in the tributary streams fed by the Bitterroot mountains. Waters from Clearwater, Elk River, and other tributaries collect in Dworshak reservoir, and eventually flow into the mighty Snake River.
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.
The most interesting and highest inflow period of the year at Dworshak is the spring freshet (snowmelt). That period of the live competition is yet to come so we’ve evaluated our model on previous years to understand its performance by conducting what we refer to as a reforecast. We conduct a reforecast by providing the model with data inputs from remote-sensing, meteorology and streamflow observations from prior years as if it was seeing it in real-time, run the model, then compare its performance with the actual observations during that period.
Below, we highlight the active spring melt period from April to July, and show the model’s performance at the 24-hr and 48-hr lead times, respectively. These plots show the realization of our ensemble model, where we have connected all the forecasts at a certain lead time (e.g. 24-hour) over a longer time period. The plots show the observations (black), reforecasted predictions (green) and the model’s predicted confidence bounds (dark blue - 50% confidence interval, light blue - 90% confidence interval).
These plots show how HydroForecast is able to capture each rise and fall highly accurately. They also show the noise (spikes) in the observations. These spikes are common, since reservoir inflows are generally calculated (not measured) from known reservoir elevation and outflow measurements.
Over the life of the CEATI competition thus far (10/1/2020 to 2/1/2021), the "by lead time" statistics indicate Upstream Tech is outperforming the other models at Dworshak, no matter how you slice and dice the 0-10 day window. In the 0-5 day lead-times for example, we have a normalized Root Mean Squared Error of 0.21, which is 60% lower than the next closest competitor. We have the only positive NSE value in the competition, and are tracking an overall bias of only 2.52%.
During the snow accumulation period in the winter, flows can run an order of magnitude lower than during the active spring melt. These low flows in the winter can be challenging to model: as temperatures hover around freezing, the model needs to know whether to translate any forecasted precipitation into snow (to accumulate) or into rain (to runoff). The plots below illustrate the 10/20/2020 - 11/20/2020 period, where flows are at about 1/10th their spring volumes. We show the observations in black and two model versions: v2 (green) -- a first iteration of our hourly model and v6 (orange) - our current production model as of writing. Notice the three small spikes (under the grey bars) and the differences in how the model responds. We have tuned v6 to capture peaks more accurately at these 24-hr and 48-hr lead times.
The following panel of plots show the precipitation and temperature over the same period. Check out the oscillation in temperatures around the freezing mark (horizontal line). The three spikes highlighted above in the streamflow correspond with precipitation events shown below. A couple of things we notice:
We’ll be watching this basin as snow continues to accumulate and then melt through July. At the time of this post, we’re pleased with how the model is tracking the precipitation and snow patterns. Our near perfect NSEs during validation and low RMSEs and percent bias during this initial forecast period are a positive indication that HydroForecast is capturing the hydrologic dynamics of the Dworshak basin.