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. This month, we turn to the Little Tallapoosa River in the Southeastern United States. The recent tragic flooding in Tennessee underscores the importance of improving decision makers’ ability to understand how much water will be coming into an area as accurately and as far in advance as possible. While inflow forecasting is only one piece of the toolkit to identify and mitigate deadly extreme weather events, the CEATI rodeo provides a unique opportunity to directly evaluate potential improvements that leaders can use going forward.
With this in mind, our case study on the Little Tallapoosa offers two big picture lessons which we will unpack in this article:
With clean source headwaters in the Piedmont region of Georgia, the Little Tallapoosa River is an important drinking water resource in its upper watershed, a recreation haven in its meandering parts downstream, and a reliable inflow source for the power-producing R.L. Harris hydroelectric dam.
The forecast basin for the CEATI Rodeo competition is defined by the USGS gauge 02413300 Little Tallapoosa near Newell, Alabama, upstream of the R.L. Harris Reservoir (also known as Lake Wedowee - see map below). Together the Little and Upper Tallapoosa Rivers drain into Harris Reservoir to form the Tallapoosa River, one of Alabama’s largest and most important river systems.
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 snow-fed Taylor River that we highlighted last month, active periods in the Little Tallapoosa are driven by rain. The most significant rain events include convective thunderstorms in the spring and summer and frontal systems in the winter months. The Little Tallapoosa’s flow is thus more variable throughout the year. Climate change is expected to intensify the hydrologic cycle in the Tallapoosa River, increasing the frequency and severity of high and low flows. Both have consequences for water management, but often can be mitigated and better planned for through the integration of forecasts.
As we enter the final months of the live CEATI competition, we reflect on how this year compares with others and how well HydroForecast has been able to predict the wet and dry periods in this basin. To understand the performance of the model over the recent years, we evaluated it from March 2019 to June 2021.
The hydrographs below highlight this two year period. They show the 24-hour (top) and 48-hour (bottom) ahead mean predictions (orange), 50% and 90% confidence intervals (blue bands), long term median (grey), 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 forecast history in March 2019 to August 25, 2021. 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 a bias of -43%, NSE = 0 and KGE = -0.03, indicating much lower predictive skill than HydroForecast.
HydroForecast holds high predictive power, especially at the 0-3 day window, and maintains a very reasonable bias (<7%) across the 10-days. HydroForecast has demonstrated the ability to predict both high and low flow periods (illustrated in the hydrographs above) and in both wet and dry years (inter-year variability).
Over the life of the CEATI competition thus far (10/1/2020 to 8/25/2021), HydroForecast is the most accurate forecast available for the Little Tallapoosa River near Newell. Out of all the metrics and lead time windows tracked by the competition (NSE, normalized root mean squared error, correlation coefficient and bias), HydroForecast is in first place in 92% of these, including first place in all metrics at the day 1 and day 3 horizons. HydroForecast only trails in the day 6 bias category (in 2nd place). This holds across various splices of the lead time windows: 1, 3 and 6 days.
The early months of 2020 are memorable to the world as the covid-19 pandemic spread rapidly throughout the globe and across the U.S. Amid this turmoil, severe weather took hold in the Southeastern U.S. with multiple devastating wind and rain events from January to April, 2020, in the Tallapoosa basin and surrounding areas. One such storm in April was responsible for at least two deaths and widespread power losses, another reminder that severe weather often hits vulnerable populations the hardest.
Foresight on the short-term horizon (1-10 day) of incoming severe weather and the expected flows from these events can help emergency managers plan shelters and evacuations with better information and with longer lead-times. Below, we investigate the period from November 15, 2019 - July 1, 2020 to understand the predictive power that HydroForecast had during this critical time period. We pieced together the 24-hour ahead forecasts into a continuous time series below:
From this set of reforecasts, we note that HydroForecast (orange line) predicted the onset of each of these events throughout the winter and spring. Overall the model had a NSE value of 0.97 and a negative bias of 4%, underpredicting a few of the peaks but capturing them within the 50% confidence intervals. Note the long term median in grey dots. Historically, this period has low and steady flows. Clearly, using the past as a predictor of the future can result in misses, whereas forecasts can use the latest information to understand what is ahead.
Next, we see how HydroForecast predicted these events three days ahead:
Comparing the day 3 reforecasts, the NSE was 0.81, and the model is a bit more biased, with negative 9% over this period (the long term median had a 65% negative bias). HydroForecast predicted the rise of each event and the magnitudes are within the 50% confidence bounds for most of the peaks, and well within the 90% for all peaks. The ability to see out three days ahead can not only help emergency management, but also signal to the dam operators of Lake Wedowee that it may be necessary to lower the lake level to prepare for unexpected elevated inflows. This process can help ensure no water is ‘lost’ downstream in power generation as well as maintain the safety of downstream communities during higher flow periods.
What drove HydroForecast in making these predictions? A critical piece of any hydrologic forecast is the expected precipitation. We use two main sources for precipitation forecasts: the ECMWF and NOAA GFS models. HydroForecast is set up to ingest both of these forecasts and then learn which one to trust more at different locations, knowing that when the weather forecasts disagree, the hydrologic forecast is more uncertain.
We see this learning reflected in the 2020 events. The time series plot below illustrates the ECMWF (orange) and GFS (green) precipitation over the same period as the hydrographs above. For example, for the event around 4/12 - 4/18, GFS predicted 25 mm/hr and ECMWF had > 5 mm/hr rainfalls. The model correctly interpreted this to reflect a medium size runoff event with wider confidence bounds, but did not trust the GFS enough to forecast what would likely have been a gross over prediction.
We are pleased by HydroForecast’s overall performance in the Little Tallapoosa basin. In a difficult-to-forecast part of the country, HydroForecast is not only ahead in the CEATI competition at this site but was also able to predict higher and wetter than normal flows during the winter and spring 2020. With only one month left in the CEATI competition, be sure to check out our final posts in September and October as we wrap up this invaluable experience. Stay tuned!