Driven by increasingly severe weather events, power outages are costing the U.S. economy between
$50 billion and $150 billion annually in direct losses and lost productivity according to the U.S. Department of Energy.
The state of New York is one of the most highly affected states in the Union, with more than
18.3 million customer outages recorded over the last 10 years from severe weather blackouts.
With this in mind, researchers Sukanta Basu, Xin Li and June Wang from UAlbany’s Atmospheric Science Research Center have partnered with NASA for a two-year project to build an AI that can predict power outages across the state and country. The project will pair the researchers with peers at the NASA Marshall Space Flight Center and the University of Alabama in Huntsville.
WHY THE PARTNERSHIP?
There are already prediction tools that governments and organizations use, like the
Outage Prediction Model from The Weather Company, an IBM Business, which utilities including Hydro One rely on today. It combines historical weather and outage records with hyperlocal forecasts, atmospheric pressure, soil moisture, and foliage data to give utilities up to a 72-hour warning window, refreshed in real time as storms evolve, so crews can be pre-positioned in the areas expected to be hit hardest.
However, by IBM’s own count, only about 23 percent of utilities actually use any outage prediction model at all. Truthfully, most still fall back on plain local weather forecasts. And the tools that are in use are generally trained on a single utility’s own historical outage logs using conventional machine-learning techniques, not satellite remote sensing or large pretrained AI models. Satellite imagery in today’s tools is typically limited to mapping vegetation or flood risk, not feeding the core prediction itself.
That’s the gap UAlbany and NASA are aiming to close. Instead of a model trained narrowly on one utility’s own outage history, the potential advantage of this prediction tool is that it can use NASA’s satellite remote sensing data that contains outage and weather datasets to sharpen forecasts and train the predictor, layered on top of pretrained AI foundation models designed to generalize across regions rather than be rebuilt from scratch for each service territory. The AI would have a specific focus on where severe weather is most likely to damage the infrastructure delivering energy in the most vulnerable areas.
“Power outage prediction is becoming an increasingly challenging problem,” said Professor Sukanta Basu. “By combining NASA’s cutting-edge AI models with satellite data and input from utility partners, we are developing a high-powered tool that goes far beyond the machine learning approaches being used today.”
As of late 2026, NASA’s Earth Observing System Data and Information System (EOSDIS) holds approximately
178.7 Petabytes (PB) of data. To try and make that understandable for our human brains, 1 Petabyte = 1,000,000 Gigabytes, your average phone holds about 256 gigabytes of data. You would need over 700,000 phones to store 178.7 PB, which if each phone was stacked on top of each other, would be 18000 ft high!
WHAT'S THE NEED?
The goal for Basu, Li, and Wang is to solve a long-standing challenge for utility companies: forecasting exactly where and when severe weather will damage electric infrastructure, up to hours to days before it happens.
To achieve this, the team is applying two of NASA’s newest AI foundation models:
- Prithvi Weather and Climate (Prithvi-WxC)
- Prithvi-Earth Observations (Prithvi-EO)
These models incorporate critical environmental variables such as tree canopy density, soil moisture, and seasonal conditions. Factoring in ground conditions is crucial because high-intensity winds are the single largest driver of weather-related power outages nationwide, responsible for an estimated 80 to 90 percent of storm-related grid damage in the U.S.
Hyperlocal accuracy is central to the project. “There is growing demand for increasingly large volumes of weather data to develop, train and evaluate AI-based weather forecasting models, particularly at the local level,” noted June Wang, director of the
New York State Mesonet at UAlbany, the nation’s largest and most advanced weather observation network. The project relies on surface observations from mesonet networks like NYS Mesonet to train and validate its AI prediction models.
Computer science Professor Xin Li added, “This is an excellent opportunity for us to explore the potential of AI in climate science, a critical application that can greatly benefit from closer collaboration between computer scientists and AI researchers.”
REAL-WORLD TESTING & OPEN ACCESSIBILITY
The tool will initially focus on three key geographic regions: the Northeast, the South, and California, with functionality designed to scale nationwide. Four major U.S. utility companies are serving as industry collaborators to guide development and test real-world applications.
This project builds directly on UAlbany’s existing momentum in grid resilience, complementing the North American Forecasting Weather, Outage, Load & Damage Initiative launched alongside UConn earlier this year. Both efforts share the broader goal of giving utilities better predictive capabilities to safeguard energy grids and reduce outage costs for consumers.
NOT EVERYONE IS SOLD
The enthusiasm from the research team isn’t universally shared. A 2026 study by Yamil Essus, Ranga Raju Vatsavai and Benjamin Rachunok, which tested power outage prediction models against several years of East Coast utility data, found that many of the strong results reported in this field don’t hold up under realistic conditions. Evaluated from a random split of the same dataset, the models perform well.
However, when tested on states or storm events the model hadn’t seen before, “predictive accuracy degrades substantially, with models often failing to outperform a simple null baseline,” the authors wrote. Notably, the study specifically tested Prithvi-WxC, one of the same NASA foundation models UAlbany’s team is building on, and found it produced only “limited and inconsistent improvements,” concluding that “publicly trained outage prediction models offer limited and uncertain operational value” given current data and evaluation practices.
Broader skepticism about AI weather forecasting echoes some of that caution. Mike Pritchard, a climate AI researcher at Nvidia, has said “we don’t know the limits of the predictability of the system” when it comes to AI-driven weather models, and has called it “a fool’s errand” to over-tune these tools for a single region when forecasting days out, a caution worth noting for a project that plans to scale from three pilot regions to the whole country.
Julian Greene of the weather-AI firm Brightband puts the current state of the art in more modest terms, too: even leading AI models are only “10% or 20% more accurate” than traditional forecasting methods today, a gain that can amount to differences as small as half a degree. And Ryan Abernathy of Earthmover points to a more basic bottleneck underneath all of it, these models are only as good as the observational data that trains them, and “data collection, at least in the U.S., is actually at risk” amid funding pressure on satellite missions and the National Weather Service, the same kind of ground-truth data (like NYS Mesonet readings) this project leans on.
None of that means the UAlbany-NASA project is bound to repeat those failures — testing the tool against four real utility partners before wider release is itself a direct response to the “does this actually generalize?” critique researchers have raised. But it’s a useful reminder that “the AI predicted it” and “the AI predicted it reliably, out of sample, before the storm hit” are two different claims. Closing that gap is exactly what skeptics say the field hasn’t done yet.
LOOKING FORWARD & UPDATES
The project’s end product will be hosted on GitHub, allowing utilities and researchers to apply their own meteorological and satellite data to predict outages. The researchers will also host workshops, webinars and a virtual course to train utility partners, students, and the broader community.
As the project is more developed and there is more peer testing of their conclusions, we’ll be sure to write about that here at PagPol and our Albany and NYS newsletter.
