Following an October storm that cut power to more than 300,000 customers -- utilities in Connecticut say they want to better predict storm outages. That means tweaking computer models which, by nature, are imperfect.
Say you’re a utility preparing for a storm. You need weather forecasts, maps of where your power lines are, and what’s around them. You need to think about trees --- if they’re trimmed -- or in danger of falling into a line. And then there’s storm severity.
In other words, a lot of stuff to consider.
“That’s the complexity, really, of the problem,” said Emmanouil Anagnostou, a professor of civil and environmental engineering at UConn and director of the Eversource Energy Center. “We really need to have a model, or a system, that can predict that range of possible impacts from different storms.”
Speaking at a conference on UConn’s Storrs campus, which mixed utility workers with academics, Anagnostou explained the statistical model, which is called the “Outage Prediction Model.”
It’s used to predict storm outages and help utility companies decide where to send resources and repair crews ahead of bad weather.
But the model failed to predict the severity of October’s late-month storm, and as a result, it took up to four days to reconnect some customers.
Now, United Illuminating and Eversource say they want to improve it.
Anagnostou said part of that involves building upon existing “machine learning” models. Those self-training systems do really well at building connections between historic and new storm data. But they’re not great at predicting more extreme weather events. Outliers, by their very nature, aren’t rooted in robust observation.
“And that also ties with climate change,” Anagnostou said. “If we want to use the model in a way to understand future storms -- the impact of future storms, you cannot do that unless you have the opportunity to extrapolate.”
So Anagnostou said the thought now is to take a more hands-on “supervised” approach by purposefully tweaking some parameters fed into the system, and pairing that data the with more autonomous “machine learning” models.
The hope is for a forecasting model with a bit more nuance.
“We hope that we will have this capability, if a future storm hits, to provide Eversource with more accurate data -- or at least information -- on the impact,” Anagnostou said.
He said UConn’s outage forecasting team, which includes the work of a lot of graduate students, will meet with power managers soon. To explain the modeling, its inherent uncertainties, and hopefully, allow grid managers to make more informed decisions.
