Google's WeatherNext 3 forecasts hourly at 5km with energy-specific outputs
DeepMind reports a 60% improvement in rain forecasting. The energy-specific predictions are aimed at the grids that AI data centres are straining.
Google DeepMind has released WeatherNext 3, a forecasting system producing hourly predictions at 5km resolution with outputs tailored to energy applications, reporting a 60 percent improvement in rain forecasting, according to TechCrunch.
Machine-learned weather models have been beating conventional numerical weather prediction on headline skill scores for two years. What distinguishes this release is resolution and specificity: 5km hourly is fine enough to matter operationally, and energy-specific outputs are aimed at a named customer rather than at meteorology in general.
Why rain is the hard case
A 60 percent improvement in precipitation forecasting is a larger claim than it sounds, because rain is where forecasting has always been weakest.
Temperature and pressure evolve relatively smoothly and are well handled by physics-based models. Precipitation depends on convection, which is chaotic, operates below the resolution of most conventional models, and produces sharp spatial boundaries — it rains here and not 3km away. Numerical models parameterise convection rather than resolving it, which is a principled approximation and a persistent source of error.
Learned models sidestep the parameterisation by fitting to observed outcomes directly. That works well where the historical record is dense and less well for events that are rare in the training data, which is the standing objection to the whole approach: a model fitted to the last 40 years may handle a genuinely unprecedented event poorly, and unprecedented events are the ones forecasts exist for.
The energy framing is the commercial story
Energy-specific forecasting means predicting the variables that determine grid operation: wind speed at turbine hub height, solar irradiance at panel level, and demand-driving temperature, at the resolution and cadence a grid operator schedules against.
That market has become considerably more valuable in the last year, for reasons that trace directly back to AI.
Five gigawatt-scale AI data centres are expected online this year. Loudoun County in Virginia hosts roughly 250 data centres and now draws more revenue from them than its operating budget requires. Australia has softened renewables rules for new AI data centres against a forecast sevenfold increase in electricity demand. South Korea has committed $919 billion to an AI infrastructure programme targeting 8.4GW by 2029. SpaceX runs a foundry in Bastrop producing gas turbine blades aimed at a backlog constraining data centre power.
Grids absorbing gigawatt-scale, relatively inflexible loads alongside growing renewable generation need better forecasts to balance, and the marginal value of accuracy rises with the stakes.
The circularity is worth stating: AI is straining the grid, and AI is being sold as the tool for managing the strain.
What is not established
Google has not published the evaluation methodology behind the 60 percent figure, which baseline it is measured against, or over what regions and periods. Improvement in precipitation forecasting is highly sensitive to all three, and a figure measured against a weak baseline in a data-rich region is a different claim from one against ECMWF's operational model globally.
Operational meteorology also has a verification culture that will test this independently over the coming months, which is more than can be said for most AI benchmark claims. The forecasts are checked against what actually happens, every day, by people with no commercial stake in the answer.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters