Weather Models

Wildflyer uses data from multiple numerical weather prediction (NWP) models to provide fire weather forecasts. Different models have different strengths — high-resolution local models capture fine wind patterns and convection, while global models provide extended outlooks.

Your organisation editor can choose which models to use in Settings > Weather.

How it works

Weather models solve the equations of atmospheric physics on a grid covering part or all of the Earth. Each grid cell produces forecasts for temperature, relative humidity, wind, precipitation, and other variables. Smaller grid cells (higher resolution) capture more local detail but require more computation and are limited to shorter forecast horizons.

For fire weather, model resolution matters most for:

• Wind — local terrain channelling, valley winds, and sea breezes are only captured by high-resolution models (< 3 km grid spacing)
• Convection — thunderstorm development and precipitation patterns require high resolution to resolve explicitly rather than being approximated
• Terrain effects — slope and aspect create microclimate variations that coarse models cannot represent

High-resolution local models

Best for short-term forecasts (0–48 hours) — fine wind patterns, convection, precipitation detail.

Model	Provider	Coverage	Resolution	Horizon
AROME	Meteo-France	France	~1.3 km	48h
HARMONIE-AROME	KNMI	Netherlands / W. Europe	~2.5 km	Short range
HARMONIE-AROME	DMI	Northern Europe	~2.5 km	Short range
ICON-D2	DWD	Germany + neighbours	~2.2 km	48h

These models capture local phenomena — valley winds, sea breezes, convective cells — that global models miss. For fire weather, this local wind detail can be the difference between a calm day and a dangerous one.

Regional models

Good for medium-range context (1–5 days) at the European scale.

Model	Provider	Coverage	Resolution	Horizon
ICON Europe	DWD	Europe	~7 km	Short/medium range
ARPEGE	Meteo-France	Europe / France	~10 km (EU)	Medium range
ECMWF HRES (IFS)	ECMWF	Europe	~9 km	Medium range

Global models

Useful for extended outlook (5–10+ days) and for comparing multiple independent forecasts.

Model	Provider	Resolution	Horizon
ECMWF IFS 0.25°	ECMWF	~25 km	Medium range
ECMWF AIFS 0.25°	ECMWF (AI-enhanced)	~25 km	Medium range
GFS	NOAA (USA)	13–25 km	Up to 16 days
GEM Global	Canada	~25 km	Medium range
ICON Global	DWD	~13 km	Medium range
UKMO Global	Met Office (UK)	~10 km	Medium range
JMA GSM	Japan	~20 km	Medium range
CMA GRAPES	China	~25 km	Medium range

Reanalysis

For historical analysis and post-event review.

Model	Provider	Resolution	Period
ERA5	ECMWF	~25 km (hourly)	1979–present

ERA5 provides consistent historical data for climatological comparisons and post-event analysis. The EFFIS percentile-based FWI calibration (Vitolo et al., 2020) is built on ERA5 data.

Ensemble forecasts and uncertainty

Most of the models above are deterministic — they produce a single forecast from a single set of initial conditions. Ensemble models take a different approach: the model is run multiple times with slightly different initial conditions, producing a range of possible outcomes.

The ECMWF ENS (Ensemble Prediction System) runs 51 members. The spread between those members is a direct, quantitative measure of forecast uncertainty:

• Tight cluster — ensemble members agree closely → high confidence forecast
• Wide spread — ensemble members diverge significantly → low confidence forecast

For fire services making resource deployment decisions 3–7 days in advance, knowing whether a forecast is high or low confidence is operationally critical. A high-confidence forecast of extreme fire weather warrants early pre-positioning. A low-confidence forecast of the same conditions warrants monitoring but not full commitment.

For fire weather specifically, wind direction and wind speed show the largest spread between ensemble members in complex terrain — which is precisely where the consequences of a wrong forecast are greatest.

Note

When multiple independent models agree on the same outcome (hot, dry, and windy on Thursday), confidence is higher than when a single model shows that outcome. Model agreement across ECMWF, GFS, and ICON is a practical ensemble check even without formal ensemble output.

Reading a forecast for fire operations

A meteorogram is a time-series plot of multiple weather variables from a specific model grid point over a forecast horizon of 72–120 hours. It is one of the most efficient tools for surveying the fire weather situation at a given location. When reading forecast output for fire operations, the key patterns to identify are:

• Afternoon peak — the timing of maximum temperature and minimum relative humidity confirms the expected peak danger window each day
• Wind evolution — is wind speed increasing or decreasing over the forecast period? Is a wind direction shift forecast, and when?
• Fire index traces — when do FWI, ISI, or HDWI cross critical thresholds? Is it a brief peak or a sustained multi-day event?
• Precipitation — will it provide meaningful fuel moisture recovery, or only surface wetting? Light rain (<5 mm) on dry fuels evaporates quickly and may not lower the FFMC significantly
• Multi-day persistence — a single afternoon of dangerous conditions is a different operational problem from five consecutive days above critical thresholds

Model resolution and terrain

Standard global models like ECMWF IFS operate at approximately 9 km resolution. High-resolution regional models like AROME operate at 1.3–2.5 km. In complex Mediterranean terrain — the Pyrenees, the Apennines, the Massif Central, the Sierra Nevada — a single model grid cell averages over valleys, ridges, and slopes that can differ in temperature by 5–10°C and in wind speed by 50–100%.

This is not a flaw in the models — it is an inherent physical limitation of the scale at which they operate.

The practical implication for field crews:

• Models are reliable for synoptic-scale conditions — will it be generally hot and windy on Tuesday? Is a cold front arriving?
• Models cannot predict the Venturi acceleration in a specific gorge, the exact timing of the afternoon anabatic onset on a specific slope, or the precise moment a sea breeze arrives at a specific coastal point

Local knowledge bridges this gap. A model forecast provides the synoptic context; terrain knowledge and local observation provide the last kilometre of interpretation. The Local and Regional Winds page describes the terrain-driven wind mechanisms that models systematically underrepresent.

How to read it in Wildflyer

When choosing models:

• Day-to-day assessment — use the highest-resolution local model available for your region (AROME for France, ICON-D2 for Germany, HARMONIE for the Netherlands)
• 2–5 day planning — compare the regional model with ECMWF HRES
• Extended outlook — look at multiple global models. When they agree, confidence is higher; when they diverge, uncertainty is high
• Post-event analysis — use ERA5 reanalysis for a consistent, quality-controlled view of what conditions were

Data limitations and disclaimer

Caution

Models are decision-support tools, not ground truth. The indicators displayed in Wildflyer are derived from numerical weather prediction models (GFS, ECMWF, AROME, and others). They are estimates computed at the resolution of a kilometre-scale grid, and they carry the inherent limitations of any model: resolution constraints, uncertain initial conditions, and simplifications in the physics.

Ground-truth observations always take priority. Local weather stations, direct field measurements, and visual reports from crews on the ground are the primary reference. When model output and field observation disagree, trust the observation.

When comparing Wildflyer data with figures from a national fire service, a regional operations centre, or another tool, always verify which model, which wind convention, and which threshold classification is being used. Differences between sources are expected and do not indicate an error — they reflect legitimate methodological choices made for different operational contexts.

Model outputs are probabilistic in nature

Even a deterministic forecast (a single model run producing a single value) carries implicit uncertainty. Key sources:

• Initial condition uncertainty — small errors in the observed state of the atmosphere amplify over time
• Grid resolution — values represent averages over a grid cell, not point observations
• Parameterisation — processes smaller than the grid scale (turbulence, convection) are approximated, not resolved

For operational decisions, treat model-derived indices as indicating a range of likely conditions, not a precise measurement. A forecasted FWI of 42 does not mean conditions will be exactly Very High — it means conditions are likely to be in the High-to-Extreme range, and planning should reflect that uncertainty.

Responsibility

Wildflyer provides data and decision-support tools. Operational decisions — resource deployment, crew positioning, risk acceptance — remain the responsibility of the incident commander and the fire service. No automated index replaces situational awareness, local knowledge, and professional judgment in the field.

Sources

• ECMWF (2023). IFS Documentation. European Centre for Medium-Range Weather Forecasts.
• Vitolo, C., Di Giuseppe, F., Krzeminski, B., & San-Miguel-Ayanz, J. (2020). ERA5-based global meteorological wildfire danger maps. Scientific Data, 7: 216.