3 min readForecasting

Local weather prediction: why terrain needs a kilometer-scale model

A regional forecast can be right and still be wrong for your site. How grid spacing and terrain shape local weather prediction, with a 2021 California example and current options.

Diagram of a global weather model grid. A cutaway globe shows the atmosphere divided into latitude and longitude columns with vertical levels, and an inset shows the physical processes computed in each column: radiation, advection, and exchanges of momentum, heat and water.

First published November 10, 2021 · Updated November 16, 2026

A regional forecast can be right and still be wrong for you. It can call for moderate rain across a county while one mountain valley floods, or for a breezy afternoon while one ridge line takes damaging gusts. For anyone responsible for a particular place, such as a plant, a wind farm, a road or a town, the gap between the regional picture and the local one is where the losses happen. Closing that gap is mostly a question of model resolution, and resolution is mostly a question of computing.

What a grid cell can and cannot see

A numerical weather prediction model divides the atmosphere into a three-dimensional grid of cells. For each cell it solves the equations of motion, heat and moisture, step by step through time. The diagram at the top of this page shows the idea.

The limit is built in. Each cell holds one temperature, one wind and one humidity. Anything smaller than a cell cannot appear in the forecast directly and has to be estimated. Regional forecasts often run on 9 to 12 km grids, coarser than a storm cloud.

Terrain suffers the same averaging. A 12 km cell blends a ridge and the valley beside it into one gentle slope. The model's mountains are lower and smoother than the real ones, and so is the weather they produce.

Where terrain changes the weather

Three local effects matter most for operations.

  • Upslope rain. When moist air meets a mountain range it is forced upward. Rising air cools, its moisture condenses, and rain falls on the slope that faces the wind. The far side stays drier. Meteorologists call this orographic lifting.
  • Wind over ridges and through gaps. Air speeds up as it squeezes over a crest or through a pass, while sheltered valleys nearby stay calm.
  • Water and coast. Sea breezes, lake-effect snow and fog form along the sharp boundary between land and water.

A model can produce these effects only if its grid is fine enough to hold the terrain that causes them. In practice that means a grid spacing of a few kilometers or less.

A 2021 example from northern California

On October 24 and 25, 2021, a strong low-pressure system brought flooding rain and high winds to northern California. TempoQuest ran AceCAST at 1 km on the morning of October 24.

The 12-hour rainfall forecast showed 2 to 7 inches in an area northeast of Yuba City, around Grass Valley, with the heaviest totals close to the local mountain ranges. National Weather Service reports at the time put more than a foot of rain at Grass Valley over the 72 hours that included that period.

The wind forecast showed the same local contrast. The map below is the 10 m wind forecast for 5:00 PM local time on October 24. The speeds labeled on it range from about 7 mph to about 49 mph across neighboring communities.

Forecast map of 10 m wind between Folsom and South Lake Tahoe in northern California, with labeled speeds from 6.8 to 49.2 mph.

Fine-grid detail is not proof of forecast skill. See how to score a forecast, including a bounded historical comparison with lidar observations.

What a kilometer-scale forecast costs

Finer grids need more computing. The cost depends on your domain, physics and hardware.

See the benchmark and cost example.

There are three ways to get a kilometer-scale forecast from TempoQuest today.

  • Order one for a site. Aerava runs storm and hail forecasts at 1 km over a 500 km square in the US, up to 12 hours ahead. The price is shown before you run. Aerava states that its storm forecasts are not yet scored and that hail size is experimental guidance.
  • Run the model yourself. AceCAST runs on Linux systems with NVIDIA GPUs, on premises or in the cloud.
  • Have the data built. Custom Weather Data projects are typically requested at around 1 km or 3 km resolution. Pricing follows the work.

See AceCAST 5, available now, for its physics, outputs and upgrade guidance.

What to do next

Price a forecast, or request custom data.

Your forecast

Have a forecast question of your own?

Tell us the decision it has to serve. One short form reaches the team, and we review every request.