ICON, GFS, ECMWF and UKMO compared

Numerical weather prediction (NWP) models are complex computational programmes that simulate the atmosphere's physics. Several national and international meteorological centres operate these models, each with distinct characteristics and update schedules.

• ICON (ICOsahedral Nonhydrostatic Model): Operated by the Deutscher Wetterdienst (DWD), Germany's national meteorological service. ICON is a relatively new global model, replacing the GME model in 2015. It typically runs four times a day (00, 06, 12, 18 UTC).

• GFS (Global Forecast System): Operated by the National Oceanic and Atmospheric Administration (NOAA) in the United States. GFS is a widely used global model, known for its open data policy. It runs four times a day (00, 06, 12, 18 UTC).

• ECMWF (European Centre for Medium-Range Weather Forecasts): An independent intergovernmental organisation supported by 35 member and co-operating states, including Ireland. Its Integrated Forecast System (IFS) is consistently ranked among the most accurate global models. The high-resolution (HRES) deterministic model runs twice daily (00, 12 UTC), with lower-resolution ensemble runs more frequently.

• UKMO (UK Met Office Unified Model): Operated by the Met Office, the UK's national weather service. The Unified Model is used for both global and regional forecasting. The global deterministic model typically runs four times a day (00, 06, 12, 18 UTC).

Each model ingests vast quantities of observational data from satellites, radar, weather stations, and other sources to initialise its forecast. The frequency of runs means that new information is incorporated regularly, leading to forecast adjustments.

The resolution and forecast length are critical factors in understanding model performance and applicability.

• ICON: The global ICON model operates with a variable resolution, typically around 13 km globally, with finer resolution (e.g., 6.5 km) over specific domains. Its forecast horizon generally extends to about 7 days (168 hours).

• GFS: The GFS model has undergone several upgrades. Currently, its primary global resolution is approximately 13 km, with a forecast horizon extending out to 16 days (384 hours). While it provides long-range guidance, the skill significantly decreases beyond 7–10 days.

• ECMWF (HRES): The ECMWF HRES model is renowned for its high resolution, typically around 9 km globally. It provides forecasts out to 10 days (240 hours). This finer resolution often allows for better representation of smaller-scale atmospheric features.

• UKMO (Global): The Met Office's global Unified Model typically runs at a resolution of approximately 10 km, with a forecast horizon extending to 6 days (144 hours).

It is important to note that these resolutions are for the global models. Many meteorological services also run regional models (e.g., HARMONIE-AROME for Ireland, UKV for the UK) at much finer resolutions (e.g., 2.5 km or less) over limited geographical areas, offering more detailed short-range forecasts for specific regions.

Ireland's weather is predominantly influenced by systems originating over the North Atlantic. Each model exhibits particular strengths in this domain.

• ECMWF: Widely considered a leader in medium-range forecasting (3-10 days), the ECMWF model often demonstrates superior skill in predicting the track and intensity of Atlantic depressions. This is attributed to its advanced data assimilation techniques and sophisticated physical parameterisations, particularly concerning ocean-atmosphere coupling. For example, its consistent performance during named storms such as Ophelia (16 Oct 2017) and Eunice (18 Feb 2022) is commonly cited.

• UKMO: The Met Office model also performs well for Atlantic systems, benefiting from extensive observational data coverage over the North Atlantic and a focus on UK and Irish weather patterns. Its regional models often provide high-resolution detail for immediate impacts.

• ICON: DWD's ICON model has shown competitive performance, particularly in short to medium-range forecasting. Its non-hydrostatic core is theoretically better suited for resolving convection and complex terrain interactions, which can be relevant for rapidly developing Atlantic systems.

• GFS: While generally reliable, the GFS can sometimes exhibit a tendency for 'phasing errors' or less accurate intensity predictions for rapidly deepening Atlantic lows compared to ECMWF, especially at longer lead times. However, its frequent updates and long forecast horizon make it valuable for early trend identification.

For critical decisions, comparing multiple models provides a more robust assessment of uncertainty, particularly for complex or rapidly evolving Atlantic weather.

All models have systematic errors or 'biases' related to their approximations of atmospheric processes. Understanding these biases is crucial for interpreting wind forecasts.

• Gusts: Gusts are sub-grid scale phenomena and are not directly resolved by NWP models. Instead, models use parameterisations to estimate gusts based on the mean wind and atmospheric stability. This often leads to a tendency to underestimate peak gusts, especially in convective conditions or complex terrain. For example, a model might forecast a mean wind of 15 m/s (30 knots) with a gust factor of 1.4, yielding a gust of 21 m/s (41 knots). However, local observations in a showery environment might record gusts of 25 m/s (49 knots) or higher. The Wind Agent's Shear Glass shows the modelled gust at 10 m, which should always be considered a lower bound.

• Coastal Wind: Coastal regions present challenges due to abrupt changes in surface roughness (sea to land) and temperature (sea breeze effects). Models with coarser resolution may struggle to accurately represent these transitions, leading to biases:
• Overestimation of wind over land near the coast: Models may not fully resolve the frictional slowing effect of land immediately onshore, especially with onshore flow.
• Underestimation of sea breeze strength: Local thermal circulations like sea breezes can be poorly represented if the model's resolution is too coarse or its boundary layer schemes are inadequate.
• Complex terrain effects: Headlands, cliffs, and bays create localised acceleration and deceleration zones that global models often cannot resolve, leading to significant local variations not captured in the forecast.

These biases highlight the importance of local knowledge and using high-resolution regional models or observations when available for coastal operations.

The availability and licensing of NWP model data vary significantly between meteorological centres, impacting how the data can be used and distributed.

• GFS: NOAA operates an open data policy, making GFS model output freely available to the public and commercial entities. This accessibility has made GFS a cornerstone for many weather applications and independent forecasters globally.

• ICON: DWD also offers a substantial portion of its ICON model output as open data, facilitating its use by a broad community.

• ECMWF: While ECMWF provides some data freely (e.g., for research, education, and WMO essential data exchange), its high-resolution deterministic and ensemble forecast products are primarily available through commercial licensing agreements. This funding model supports its advanced research and operational capabilities.

• UKMO: The Met Office operates a mixed model. Some data is available under open government licenses, but its highest-resolution and most comprehensive products often require commercial licensing, particularly for commercial applications.

The Wind Agent sources data from Open-Meteo, which aggregates and processes various model outputs, including licensed data, to provide a comprehensive view. This approach allows users to benefit from the best available model data without navigating complex licensing directly.

The Wind Agent presents multiple model forecasts side by side in the model_compare chart and the Agreement Spine. This approach is fundamental to understanding forecast uncertainty and making informed decisions.

1. Quantifying Uncertainty: No single model is consistently 'best' in all situations. By comparing different models, users can visually assess the degree of agreement or disagreement among them. When models align closely, confidence in the forecast is higher. When they diverge, it signals greater uncertainty, prompting a more cautious approach.

1. Identifying Biases and Strengths: Over time, users may observe that certain models perform better in specific scenarios or exhibit particular biases for their location. For example, one model might consistently forecast higher gusts, while another might be more accurate for light wind conditions. Side-by-side comparison aids in developing this intuitive understanding.

1. Ensemble Thinking: The comparison of deterministic models is a simplified form of ensemble forecasting. Instead of relying on a single 'best guess', it encourages consideration of a range of possible outcomes. This is particularly valuable for high-stakes operations where exceeding a wind limit could have significant consequences.

1. Early Warning of Change: A sudden divergence between models can be an early indicator of a significant forecast change or the development of a complex weather pattern. Monitoring these shifts allows for proactive adjustments to operational plans.

The goal is not to pick a 'winner' but to synthesise information from multiple sources to build a more complete picture of the likely wind conditions and their associated risks.

The model_compare chart is designed to provide a quick visual assessment of model agreement for wind speed and gust. Each line represents a different model's forecast for the selected location and height.

• Lines and Colours: Each model (ICON, GFS, ECMWF, UKMO) is represented by a distinct coloured line. The primary metric displayed is typically the mean wind speed at 10 m, with gust speed often shown as a shaded band or separate line.

• Agreement and Divergence: Observe how closely the lines track each other. When lines are tightly clustered, it indicates high model agreement and generally higher forecast confidence. When lines spread out, it signifies divergence, meaning models are predicting different outcomes for that period. This divergence is a key indicator of forecast uncertainty.

• Trends and Peaks: Identify overall trends (e.g., increasing or decreasing wind) and specific peaks or troughs. Compare the timing and magnitude of these features across models. For instance, if ECMWF forecasts a peak gust of 20 m/s at 14:00, but GFS forecasts 15 m/s at 16:00, this highlights a timing and intensity disagreement.

• Impact of Lead Time: Model agreement generally decreases with increasing lead time. Forecasts for the next 12-24 hours typically show high agreement, while those 3-5 days out will naturally show more spread.

• Comparing against your limit: Mentally (or physically, if using the interactive tool) overlay your operational wind speed or gust limit onto the chart. Note how many models predict an exceedance and by how much. This qualitative assessment informs your decision-making process.

While global NWP models are developed and run by different organisations, they are not entirely independent entities. A significant degree of interdependency exists, primarily through the sharing of observational data and scientific advancements.

• Global Observation Network: All major NWP centres rely on a global network of observations, including satellite data, radiosondes, aircraft reports, and surface weather stations. Much of this data is exchanged internationally through the World Meteorological Organization (WMO) Global Telecommunication System (GTS). Therefore, all models are initialised with largely the same foundational observational dataset.

• Data Assimilation: Each centre employs its own sophisticated data assimilation system to optimally combine observations with a short-range forecast from the previous model run. Differences in these assimilation schemes, as well as the 'first guess' forecast, contribute to initial differences between models.

• Physics Parameterisations: Models use different parameterisations to represent sub-grid scale physical processes like clouds, precipitation, and boundary layer turbulence. These differences lead to variations in how atmospheric energy and moisture are handled, causing forecasts to diverge over time.

• Scientific Collaboration: There is extensive scientific collaboration and knowledge sharing within the meteorological community. Advances made by one centre (e.g., in satellite data processing or cloud physics) are often adopted or adapted by others, leading to a gradual convergence of model capabilities over time. This means that while they are distinct, they are also part of a larger, evolving scientific endeavour, rather than completely isolated systems.