00 / SHORT ANSWER
A forecast starts with measurements, then admits what the measurements cannot know
A heatwave forecast is built in layers. Observations constrain the atmosphere now; data assimilation turns uneven measurements into a physically consistent three-dimensional analysis; a numerical model advances that analysis; and an ensemble shows how small uncertainties can grow.
The final heat probability is a threshold question asked of the ensemble: how many calculated futures exceed the relevant local temperature criterion? That percentage describes model guidance under stated assumptions. It does not make uncertainty disappear.
01 / THE STARTING STATE
Thousands of partial views become one atmospheric analysis
Stations and radar
Surface instruments anchor conditions near people, while radar tracks the structure and motion of precipitation.
Balloons and aircraft
Profiles of temperature, humidity, pressure, and wind reveal the atmosphere above the screen-height thermometer.
Satellites, ships, and buoys
Broad coverage and ocean observations fill regions where conventional stations are sparse.
Data assimilation
A recent short forecast and quality-controlled observations are combined with their error characteristics to estimate one coherent initial state.
Assimilation matters because measurements differ in time, height, footprint, and precision. It also keeps the reconstructed atmosphere dynamically plausible before the forward calculation begins.
02 / FORWARD PHYSICS
The model advances a gridded atmosphere one time step at a time
Numerical weather prediction solves discretised equations for motion, heat, moisture, and pressure across a three-dimensional grid. Radiation, land–surface exchange, clouds, and turbulence also affect the result; processes smaller than the grid are represented through parameterisations.
A persistent hot spell therefore emerges from an evolving circulation, not from extending yesterday’s temperature line. Sinking air, sunshine, weak ventilation, soil moisture, cloud, and the movement of pressure systems can reinforce or interrupt the heat.
03 / 51 FUTURES
A probability is counted across related, physically calculated forecasts
ECMWF documents a medium-range ensemble covering days 0–15 with 50 perturbed members and 1 control member. The control uses the unperturbed best estimate. The other members sample plausible uncertainties, revealing when a hot signal is robust and when the future is sensitive to a small change at the start.
If 37 of 51 members exceed a stated threshold, the raw ensemble fraction is about 73%. The threshold, calibration, duration, and local warning rules still belong beside that number.
51-future heatwave forecast lab
Watch uncertainty become a probability
Select a lead day and adjust the initial uncertainty. Every orange line is one deterministic member; the dark line is the unperturbed control.
Deterministic educational scenario. The 51-path structure mirrors ECMWF’s 50 perturbed members plus 1 control member for days 0–15; temperatures and the 35°C line are illustrative and are not a live forecast.
04 / READING THE FAN
Lead time changes the shape of a responsible answer
Close to day zero, observations tightly constrain the starting point and the member paths tend to cluster. As lead time grows, nonlinear dynamics amplify small differences. A wider fan means the atmosphere supports a broader range of outcomes; it does not mean every result inside the fan is equally likely.
Heatwave definitions also vary by place and purpose. Some use daily maximum temperature, some include minimum temperature or humidity, and many require several consecutive days. This page uses a single-day 35°C line solely to make the ensemble counting step visible.
05 / SOURCES
Primary operational documentation used for this explainer
NOAA / National Weather Service
About Supercomputers
Explains how satellite, balloon, buoy, radar, and other observations feed numerical models used for extreme-heat prediction.European Centre for Medium-Range Weather Forecasts