AI weather forecasting enters the energy market as Google targets grid operators with WeatherNext 3

Google’s latest AI climate forecasting mannequin predicts wind pace at 100 metres above the bottom, roughly the peak of a contemporary wind turbine. It additionally forecasts cloud cowl and the way a lot daylight reaches the floor, and it updates each hour. Power merchants, grid operators and wind and photo voltaic builders already pay different firms for that knowledge. The introduction of WeatherNext 3 now places Google of their market.

Google DeepMind and Google Analysis released the mannequin on September 3. It produces a worldwide forecast each hour at as much as five-kilometre decision for floor variables comparable to temperature and moisture. The earlier model, WeatherNext 2, labored on a 25-kilometre grid and refreshed each six hours. Google says the brand new power variables are supposed to assist grid operators and builders predict how a lot energy their wind and photo voltaic property will generate, then match that in opposition to demand.

The buyer aspect of the launch has had many of the consideration. WeatherNext 3 now powers climate ends in Google Search, the Gemini app, Google Maps and the Google Maps Platform Climate API. Behind it sits an enterprise layer that issues extra commercially. The identical forecast knowledge might be queried in BigQuery and Earth Engine or downloaded in bulk from Google Cloud Storage, with no mannequin setup required by the client.

Why the power sector is shopping for AI climate forecasting

Grid operators are working a system that has turn into tougher to foretell at each ends. On the era aspect, renewables now account for many new capability. S&P International Market Intelligence’s US Grid Outlook 2026 tasks photo voltaic and power storage as the first sources of latest capability this 12 months, at 51.2GW and 25.7GW respectively out of greater than 90GW of deliberate additions. 

Photo voltaic and wind generate in line with the climate reasonably than demand, so every gigawatt added makes a short-term forecast extra correct.

On the consumption aspect, the brand new load is coming from AI. S&P International identifies the unfold of information centres throughout North America as a main driver of the latest surge in electrical energy demand, forcing utilities to revise their load forecasts upward. Deloitte’s 2026 Power and Utilities Industry Outlook tasks peak demand rising by roughly 26% by 2035, with knowledge centre demand alone probably reaching 176GW, 5 occasions its 2024 stage.

The price of getting a forecast mistaken is simple. If an operator underestimates how a lot wind energy will arrive, it has to purchase alternative electrical energy at quick discover, often from fuel vegetation saved on costly standby. If it overestimates, wind and photo voltaic farms find yourself being paid to modify off as a result of the grid can’t soak up what they’re producing. Each outcomes are costly, and each are forecasting failures.

The market Google is coming into

Promoting climate forecasts to the power sector is a longtime enterprise. Vaisala, Solcast, DNV’s WindGEMINI and IBM’s HyperWatch all compete in it. So does Jua, a Swiss agency that claims its EPT-2 mannequin beats Microsoft Aurora and DeepMind’s earlier GraphCast on accuracy whereas updating 24 occasions a day, in opposition to what it describes as a typical 4 updates a day amongst rivals.

Google’s benefit is attain. The identical forecast seems as a desk in BigQuery, a layer in Earth Engine, an API in Google Maps Platform and the default reply in Google Search. No specialist vendor has that unfold, and the hourly refresh closes the update-frequency hole these distributors have used to distinguish themselves.

The incumbents have one technical argument left. Jua’s printed place is that physics-based fashions comparable to ECMWF’s HRES nonetheless outperform purely data-driven AI fashions throughout record-breaking excessive climate, as a result of physics fashions encode guidelines about how power and mass transfer by means of the environment, whereas AI fashions study patterns from previous knowledge. 

Jua sells a physics-constrained product, so the declare serves its personal pursuits. It additionally describes the situations grid operators fear about most, when a storm falls outdoors something the mannequin has seen in coaching.

What’s new, and what’s being oversold

The architectural declare behind WeatherNext 3 is that it learns from actual observations as a substitute of from simulations. Most AI climate fashions, WeatherNext 2 included, are educated on output from numerical climate prediction fashions, that are supercomputer-driven physics simulations that carry a six-hour knowledge lag. That lag can introduce bias in fast-changing variables comparable to rain and floor temperature. WeatherNext 3 ingests dwell geostationary satellite tv for pc imagery and trains straight on readings from particular person climate stations.

The shift is actual, although narrower than a lot of the protection has urged. Google’s personal system diagram reveals the mannequin taking in one-hour satellite tv for pc mosaics alongside conventional historic evaluation. DeepMind senior analysis scientist Ilan Value instructed Bloomberg the acquire comes from not ready for the following evaluation and utilizing the latest data obtainable as a substitute. 

Reporting places the remaining knowledge lag at three to 4 hours, down from about seven. Dependence on numerical climate prediction has been lowered, not eliminated.

The accuracy figures want related care. Google studies enhancements of as much as 60% in opposition to NASA’s IMERG satellite tv for pc product, 30% in opposition to MRMS radar and 10% in opposition to rain gauge readings at early lead occasions, measured utilizing an ordinary scoring methodology for chance forecasts. These are three separate baselines, and the odds don’t add collectively. The extensively repeated declare of fifty% higher precipitation forecasting applies particularly to forecasts a day or extra forward. Each determine carries an “as much as” qualifier, which makes each a finest case reasonably than a typical consequence.

Google printed no impartial third-party validation alongside the launch. It factors as a substitute to dwell evaluations by Brightband, whose leaderboard it cites in claiming WeatherNext 3 is essentially the most correct international climate mannequin thus far. A utility contemplating a change away from a paid specialist will care extra about efficiency in its personal service territory, by itself property, than a few international leaderboard place.

Google’s personal stake in the issue

Google is promoting forecasting instruments right into a grid drawback its personal trade helped create. The information centre build-out driving the load development utilities are struggling to serve is led by the hyperscalers, Google amongst them, and Google has signed multi-gigawatt renewable procurement agreements to provide its personal services.

Correct prediction of wind and photo voltaic output is straight helpful to an organization matching massive volumes of fresh power in opposition to a load that’s each rising and variable. That’s industrial logic, and it goes some option to explaining why the power variables shipped on this launch.

Google has not printed pricing for enterprise entry to WeatherNext 3, or mentioned whether or not the BigQuery and Earth Engine knowledge carries normal Cloud question costs or a separate licence. Utilities weighing a transfer away from a paid specialist will need that determine earlier than they weigh any accuracy declare.

2025, greater than 65,000 workers in its Company and Funding Financial institution have been actively utilizing the platform, whereas greater than 90% of its engineers have been utilizing AI coding assistants.

The financial institution additionally mentioned AI-based transaction screening allowed it to evaluate greater than twice the earlier transaction quantity whereas lowering handbook operator checks by half.

Financial institution of America is utilizing a generative AI-enabled system known as EricaAssist with greater than 18,000 customer support workers. The instrument summarises why a buyer is asking, retrieves related data, and recommends potential subsequent steps whereas holding the worker answerable for the interplay.

Financial institution of America mentioned in July 2026 that EricaAssist can ship contextual steerage in below three seconds and has lowered common name occasions by almost one minute. The financial institution plans to increase the system to extra servicing situations and enterprise strains later in 2026.

(Photograph by Google)

See additionally: MIT AI forecasts excessive climate with out historic knowledge

Need to study extra about AI and large knowledge from trade leaders? Take a look at AI & Big Data Expo going down in Amsterdam, California, and London. The excellent occasion is a part of TechEx and is co-located with different main expertise occasions, click on here for extra data.

AI Information is powered by TechForge Media. Discover different upcoming enterprise expertise occasions and webinars here.