Applying science-guided AI to dam safety operations

Credit: WaterArchives.org / Aerial view of the Teton dam failure on June 5th, 1976, Newdale, Idaho.
Oct 19, 2023
Table of contents

Safe operations of our world’s dams depend on sizing and constructing the infrastructure adequately, maintaining that infrastructure, and managing the dams with the best information available. Each of these components carry risk, and oftentimes the consequences of failure are catastrophic for life and property. The number of high hazard dams has doubled over the last 20 years, meaning that more dams are at risk of events/failures. The cost of these failures are devastating, with damages from dam failures in the millions, sometimes billions, and concern that failure incidents will increase due to aging infrastructure of dams. Recently, the high profile Michigan dam failures in 2020 caused an estimated $175 million in downstream damages and will take $120-200 million to repair and restore the lakes. 

In this blog post, we’ll explore why dams fail, and discuss how non-structural – i.e. digital – tools can help with keeping dams safe, especially in times of uncertain climate and extremes.

Why do dams fail?

So why do dams fail, and what can we do to protect against failures? According to the Association of State Dam Safety Officials Incident database, 42% of failure and non-failure incidents reported were triggered by extreme weather - this is by far the most common cause of an incident. Heavy precipitation and <span class="term" data-def="Flooding that develops more gradually, usually from prolonged and persistent moderate to heavy rainfall">areal flooding</span> can lead to overtopping; in fact, from 2010-2019, about half (180 out of ~350) failures in the U.S. were hydrologic/flooding related. Other types of failures are geotechnical, sometimes due to seepage events and, in some cases, seismic activity can lead to instability and failure.  

Extreme weather is not preventable - we can’t sprinkle dust on a hurricane or thunderstorm cloud and make it disappear (yet?) - but preparation for extreme weather and flooding impacts is within our purview. Current processes vary: some use advanced models and early warning systems, while others use weather forecasts and on-the-ground rain, and mentally translate how the rivers will respond. 

Ultimately, data and forecasts are key ingredients in operating safely, and the more reliable and automatic the workflows, the more effective they are at protecting during unexpected events. This is where HydroForecast, a modeling tool that uses AI and science to predict river flow 10-days ahead can be critical in safe operations.

Let’s see how.

Digital infrastructure for dam safety

When a tropical storm from the Gulf of Mexico made its way up through Texas and the Rocky Mountain West, it dropped an unusual pattern of wide-spread rainfall for the late summer period in Colorado. This time of year is common for conducting in-stream maintenance, as lower flows and normally dry weather make it more safe, reliable and predictable. However, unexpected events like these require the help of forecasts that are accurate and allow for as much preparation time as possible. 

This is where the help of digital infrastructure, or streamflow forecasts supported by science and AI, can better equip organizations, and their workers and equipment on-site, for what’s to come.

So let's look at the data. Here’s what the streamflow observation (black line in top plot) and rainfall rates (bottom plot) looked like over the late August event for a specific inflow event into a dam in Colorado.

View of the large inflow into a dam in the Rocky Mountains, Colorado, following remnants of a hurricane's path up from the Gulf of Mexico. The black line is the observed streamflow, and the dotted line is the long-term average. Source: HydroForecast
Observed precipitation rate over the event period. Source: HydroForecast

Next, we’ll walk through the forecasts leading up to the event. 

On August 21, five days ahead of the observed peak flow on August 25th, HydroForecast’s probabilistic distribution captured a rise ahead. The NWS-River Forecast Center’s (RFC) forecast showed a large peak on August 24th. The GFS, GEFS, and ECMWF-HRES precipitation forecasts were all predicting between 1.5 - 2” of rain over the upcoming 10 day period. 

This type of early warning system for larger unexpected off season events can give enough time to move people and equipment out of the way, and alert any downstream users of a potential unscheduled release.

HydroForecast’s mean lines up with observations at the start of the forecast, and picks up the future rise. The National Weather Service's River Forecast Center (red dashed line) predicted an early peak. Source: HydroForecast

On August 23rd, 3 days ahead of the observed peak, HydroForecast’s peak timing became more pronounced, though the mean is slightly late. The NWS-RFC’s forecast dropped its prediction of any rise in flows, likely due to reduced precipitation totals in the GFS and GEFS. ECMWF-HRES increased its precipitation forecast from ~2” to ~3” which, importantly, HydroForecast incorporates into its prediction and therefore stayed more consistent.

HydroForecast’s probability distribution captures the peak within its 50% confidence bounds. Source: HydroForecast

One day ahead of the event, the NWS-RFC predicted a significant, erroneous peak. This type of forecast error can be difficult to manage for flooding and safety, as it could trigger actions that turn out to be false alarms. 

HydroForecast, however, remained consistent, shifting to update rainfall from the 12-hours before. 

The NWS-RFC adjusted to predict a very large peak one day before the event, making an operator’s decision very difficult. Source: HydroForecast

Removing the NWS-RFC’s forecast, HydroForecast’s mean and 50% confidence interval captured the observed peak and recession with marginal error

Just like Goldilocks, forecasts that are too high or too low, or plain inconsistent, are problematic for reservoir operators. HydroForecast intends to be a reliable and consistent accurate forecast for safe dam management. 

HydroForecast’s mean captured the event with minimal error, helping operators with day ahead decisions. Source: HydroForecast


So, what are we left with? Here are the takeaways:

  • Consistency and reliability matters – a forecast that does not fluctuate day to day can help dam operators be more confident to make critical decisions
  • Weather forecasts significantly influence hydrological models, so having methods (e.g. in HydroForecast’s machine learning model) to use multiple weather forecast sources effectively can hedge against these errors and produce a better prediction
  • Extreme events happen, and are unpredictable: industry-proven forecasts that utilize the latest data and hydrologic methods are a way to help dam operators prepare for the unexpected. Even just a few days ahead can prevent damages and unnecessary downstream effects that can lead to regulatory challenges. 

If you’d like to learn more about HydroForecast, please get in touch with Head of Product laura@upstream.tech or email our team more broadly at team@hydroforecast.com to learn how HydroForecast can make a difference when it comes to dam safety.

Read more

No items found.