Weather models explained

How a forecast model works

The model divides the atmosphere into a three-dimensional grid of boxes, many kilometres across and stacked dozens of layers high. Into each box goes the best estimate of the current temperature, pressure, wind and moisture, built from millions of observations: satellites, weather balloons, aircraft, ships, buoys and ground stations.

The model then solves the equations of fluid motion and thermodynamics to step forward in time, a few minutes at a time, out to days or weeks ahead. Processes smaller than a grid box, such as individual clouds or turbulence near the ground, cannot be calculated directly and are represented by approximations. Those approximations are one reason models differ.

Global and regional models

Global models must cover the whole Earth, because weather arriving in the UK in five days is over North America or the Atlantic today. They run with grid boxes of roughly 9 to 25 km. Regional models cover a smaller area, such as the UK, with much finer grids, around 1.5 km for the Met Office's UK model, which lets them represent hills, coasts and showers far better, but only for a few days, because weather from outside their area soon arrives.

The best forecasts combine them: a high-resolution regional model for the first couple of days, global models after that. Our forecast data comes from Open-Meteo, which blends the best available models for each location in this way.

The main models you will hear about

ECMWF, from the European Centre for Medium-Range Weather Forecasts based in Reading, has long ranked at or near the top of the global models in published verification scores. Its ensemble of 51 runs drives our 15-day confidence and its extended and seasonal systems drive our long range pages.

GFS, the Global Forecast System from the US National Weather Service, runs four times a day and its data is freely available, which is why many weather websites use it. Our wind and colour maps use it.

The Met Office runs its own global model and a high-resolution model for the UK. ICON from Germany's weather service and GEM from Canada are other widely used global models. In recent years machine-learning models trained on decades of past weather have joined them and are now run operationally alongside the traditional physics-based models.

Why models disagree, and what to do about it

Two models can start from slightly different analyses of today's weather, use different grid sizes and represent clouds, rain and the land surface differently. A few days on, those differences can mean a rain band arriving twelve hours apart, or a depression tracking 200 km further north.

Disagreement between models, or between the runs of one ensemble, is information. It tells you the weather is genuinely hard to call. Rather than picking the model you like best, look at the range. Below is the week-by-week outlook from the ECMWF extended-range system for London, which compares 51 runs with the model's own climate.

Deterministic versus ensemble

A deterministic forecast is a single run at the highest resolution: the model's best single guess. An ensemble is many runs, usually at slightly lower resolution, each from a slightly different start. The single run gives you detail; the ensemble gives you the odds. For the first few days the two usually agree. Beyond that, the ensemble is the better guide, and its average often beats any single run.