AI promises to supercharge weather forecasts

Posted: April 04, 2025

On April 1, 1960, NASA sent the world’s first weather satellite into orbit. Coming less than three years after Sputnik ushered in the Space Age, it sounds somewhat underwhelming with hindsight: a solar-powered cylinder on spindly legs, carrying two TV cameras to beam videos of cloud formations back to Earth. 

But more than six decades later, satellites like it remain the most crucial input for weather forecasting and climate monitoring—even if they’ve gotten substantially more sophisticated. 

The European Centre for Medium-Range Weather Forecasts, or ECMWF, which produces predictions for 35 states, now bases its forecasts on around 800 million observations every day, the vast majority from satellites (alongside measurements from planes, boats, sea buoys and other stations).

But, thanks to AI, the way these observations are processed, and the detail and quality of the resulting forecasts, is now rapidly evolving.

Why weather forecasts matter for power companies and grid operators 

Reliable weather forecasts are crucial far beyond deciding whether to wear sunscreen or leave the house with an umbrella. 

Many industries are deeply dependent on the elements, from shipping to construction to agriculture. But perhaps few are as exposed as utilities. Power plants and transmission lines can be found in virtually every geography and are highly sensitive to extreme weather, from droughts and heatwaves to wildfires, hurricanes and floods. 

Some of those threats are now exacerbated by climate change, raising the stakes for each individual disaster. Total damages from Hurricane Helene, which tore through the southeastern U.S. last year, alone have reached close to $80 billion.

Getting better and earlier warning of storms and other weather events before they strike could allow for better preparation, including of power infrastructure—especially as global warming makes extreme weather more unpredictable.  

Knowing more accurately when a storm will hit, and how severe it will be, can allow utilities to ready maintenance crews to repair knocked-out power lines, for example. Power producers can also make sure they have enough backup reserves to jump into action when individual plants are affected. 

Better awareness of changing weather patterns can also spur more long-term measures to protect critical infrastructure.  

Analysis of Winter Storm Uri, which caused widespread power outages in Texas in 2021, found that a principal cause of the blackout was natural gas producers' failure to properly winterize production and distribution facilities, for instance. 

Even absent a disaster, weather forecasts are essential for both power companies and grid operators charged with predicting energy demand on a daily basis and navigating dips and peaks in consumption. And with the rise of renewables, robust predictions for energy availability will only become more urgent since wind, solar and hydropower all operate at the whims of nature. 

How machine learning and AI are improving weather forecasts 

Enter AI.

PG&E Corp., the San Francisco-based utility, already uses machine learning to analyze wind speed, humidity and temperature to forecast the short-term chances of wildfires, which are frequently caused by power lines in dry, windy conditions.  

Spotting the signs of an impending blaze by combining historical weather data and satellite images means the company can make a more informed call on whether to pre-emptively shut down power lines in high-risk areas, says Andy Abranches, the PG&E’s senior director for wildfire preparedness and operations. 

“Our system runs four times a day about a trillion data points in the Amazon cloud to tell us the weather and fire potential,” Abranches said during a panel discussion by the United States Energy Association last year. “It allows us to make those decisions from an operational perspective regarding how we should run the grid.” 

Now, both Google and the ECMWF say they have made breakthroughs in using AI to produce much more detailed and accurate weather models for broader applications. 

In December, Google announced a new AI model, GenCast, that it says delivers faster and more accurate forecasts up to 15 days ahead, besting ECMWF in predicting both day-to-day weather and extreme events. 

A diffusion model adapted to the spherical geometry of the Earth, GenCast uses historical weather data to produce an ensemble of 50 or more predictions—a big difference from the company’s previous deterministic model that provided a single best estimate. 

Google says GenCast, along with the company’s other next-generation AI-based weather models, will improve forecasting of precipitation, wildfires, flooding and extreme heat. It also says that tests have shown the new model can accurately forecast power generation from wind farms around the world. 

AI transformations in weather science and forecasting 

Not to be outdone, the ECMWF two months later unveiled its own AI forecasting system, which it says is about 20% more accurate on key predictions, such as the trajectory of tropical cyclones, than the best conventional methods. 

Like Google, ECMWF specifically emphasized the value of its system for renewable energy forecasting. On top of the standard temperature, precipitation and wind, the model forecasts solar radiation and wind speeds at 100 meters—the height of a typical wind turbine. Alongside the power generation industry, the insurance, security and shipping sectors could also benefit. 

While other tech companies such as Huawei and Nvidia, as well as meteorological offices around the world, are also applying AI to the weather, the ECMWF crucially makes its predictions freely available to everyone. “This milestone will transform weather science and predictions,” Florence Rabier, the organization’s director-general, said in announcing the news. 

For now, the model is limited to a deterministic forecast, although the ECMWF also plans to eventually expand this to a 50-strong collection of scenarios. 

In the meantime, a whole host of companies now also specialize in providing granular regional weather data for power companies, combining advanced machine learning models and intelligent sensors. 

Harkening back to the early days of satellite launches, some are even working on improving the physical infrastructure for collecting meteorological data. Closer to Earth, Swiss company Meteomatics is even selling drones to replace traditional weather balloons, which are still used by the hundreds. According to the company, the drones can provide additional measurements and address gaps in existing networks, making weather models even more reliable. 

It’s a long way from a single satellite with a video camera.