Why forecasts disagree: initial conditions, physics, and local effects

Numerical weather prediction (NWP) models start from the current state of the atmosphere. This 'initial condition' is never perfectly known. Observations from satellites, weather balloons, radar, and surface stations are assimilated into the model's grid, but there are always gaps and uncertainties. Each modelling centre (e.g., ECMWF, NOAA, Met Éireann) uses a slightly different set of observations and a different method to combine them.

Even tiny differences in these initial conditions can grow significantly over time, a phenomenon known as the 'butterfly effect' or sensitive dependence on initial conditions. For example, a small, unobserved perturbation in the Atlantic could lead to a different track for a low-pressure system days later. This is a fundamental limit to predictability.

To account for this, ensemble forecasting runs the same model multiple times with slightly perturbed initial conditions. The spread of these ensemble members provides an estimate of the forecast uncertainty arising from initial condition errors. A wide spread indicates high uncertainty, often due to an unstable atmospheric state where small initial errors amplify rapidly.

The atmosphere is a continuous fluid, but NWP models represent it on a discrete grid of points, both horizontally and vertically. The spacing of these grid points, known as resolution, varies between models. Global models like ECMWF's IFS or NOAA's GFS might have horizontal resolutions of 9–13 km, while regional models (e.g., Met Éireann's HARMONIE-AROME, UK Met Office's UKV) can be 2.5 km or finer.

Processes smaller than the grid resolution, such as individual clouds, turbulence, or convection, cannot be explicitly resolved. Instead, they are represented by parameterisation schemes – simplified mathematical approximations of complex physical processes. Each modelling centre develops and refines its own set of parameterisations for:

• Cloud microphysics (how clouds form, grow, and precipitate)
• Radiation (how solar and terrestrial radiation interact with the atmosphere)
• Boundary layer processes (friction, heat exchange near the surface)
• Convection (updrafts and downdrafts in thunderstorms)

For example, one model's convection scheme might initiate precipitation earlier or more intensely than another's, leading to differences in forecasted rainfall and associated wind shifts. These differences in grid resolution and parameterisation are a primary reason why forecasts diverge, especially for local phenomena.

When forecasts disagree, it is often useful to distinguish between errors in timing (phase errors) and errors in intensity (amplitude errors). A model might accurately predict the strength of a weather system but misjudge its arrival time by several hours, or vice versa.

Consider a forecast for a strong southerly wind associated with an approaching front. Model A predicts 15 m/s at 14:00, while Model B predicts 15 m/s at 17:00. This is primarily a timing error. If Model A predicts 15 m/s and Model B predicts 10 m/s for the same time, this is an intensity error.

Worked Example:

Suppose the observed wind speed is 12 m/s at 15:00. Model A forecast 15 m/s at 14:00, and Model B forecast 10 m/s at 15:00.

• Model A: Intensity error = |15 m/s - 12 m/s| = 3 m/s. Timing error = 1 hour early.
• Model B: Intensity error = |10 m/s - 12 m/s| = 2 m/s. Timing error = 0 hours (correct timing).

In this scenario, Model B had better timing but slightly underestimated intensity, while Model A overestimated intensity and was early. Both types of errors are important for operational decisions. The Agreement Spine on The Wind Agent helps visualise these differences across models, allowing users to assess the consistency of timing and intensity predictions.

Fronts, low-pressure systems, and high-pressure systems are dynamic features that move across the forecast domain. A 'phase error' occurs when a model predicts the position or timing of these features incorrectly. For example, one model might place a cold front over County Kerry at 09:00, while another places it over County Cork at the same time. This can lead to significant differences in forecasted wind, temperature, and precipitation for a given location.

These errors are particularly common with rapidly developing or fast-moving systems. The exact speed and track of a low-pressure centre, for instance, can vary between models. A difference of just 50 km in the predicted track of a deep depression can mean the difference between strong winds and gale-force conditions for a coastal site. Similarly, the timing of a frontal passage, which often brings a sharp wind shift and gust increase, can be off by several hours.

Phase errors are a major contributor to forecast divergence beyond 48 hours. When models disagree on the position of a significant weather feature, the resulting wind forecasts for specific locations will naturally diverge. The Agreement Spine can show this as a shift in the peak wind speed between models.

Many meteorological phenomena are highly localised and depend on fine-scale interactions between the atmosphere and terrain. These include:

• Sea breezes: Thermally driven winds that develop on sunny days near coastlines.
• Topographic channelling: Wind acceleration or deflection through valleys or around headlands.
• Lee effects: Sheltering or enhanced turbulence downwind of hills and mountains.
• Convective showers: Localised downpours and associated gust fronts.

Lower-resolution global models often struggle to resolve these features. For example, a global model with a 13 km grid might smooth out the topography of the Wicklow Mountains, failing to capture the channelling of wind through a specific pass. A higher-resolution regional model, such as Met Éireann's HARMONIE-AROME (2.5 km), is more likely to represent these local effects.

However, even high-resolution models have limitations. They still rely on parameterisations for sub-grid scale processes, and their accuracy is highly dependent on the quality of their initial conditions and boundary conditions (data fed from a larger-scale model). Therefore, even for local effects, different high-resolution models can still show discrepancies based on their specific configurations and how they handle these complex interactions.

Even if models perfectly agreed on the atmospheric state, their output can still appear to differ due to how wind speed is defined and presented. Key differences include:

• Averaging Period: The World Meteorological Organisation (WMO) standard for mean wind speed is a 10-minute average. However, some models or data providers might use 2-minute, 1-hour, or instantaneous values. Gusts are typically defined as the highest 3-second average within a 10-minute period. If one model reports a 1-hour average and another a 10-minute average, their 'mean wind speed' values will naturally differ, especially in turbulent conditions.
• Reference Height: Wind speed varies significantly with height due to friction. The standard reference height is 10 metres above ground level (AGL). However, some models might output wind at their lowest atmospheric level, which could be 20 m or 30 m AGL. If not consistently adjusted to 10 m, this will lead to apparent discrepancies. The Wind Agent's Shear Glass explicitly shows wind at 10, 80, 120, and 180 m to address this.

These seemingly minor differences can lead to confusion. For instance, a 10-minute mean wind of 10 m/s might correspond to a 2-minute mean of 11 m/s in gusty conditions. Always check the definitions when comparing raw model outputs. The Wind Agent standardises all displayed wind speeds to 10-minute means at 10 m AGL for consistency, unless explicitly stated otherwise for height-matched data.

While model disagreement can be frustrating, it is crucial to view it as valuable information rather than mere noise. The spread among different model forecasts, particularly from independent centres, provides an indication of the inherent uncertainty in the prediction. When models largely agree, confidence in the forecast is higher. When they diverge significantly, it signals increased uncertainty and the need for a more cautious approach.

This concept is central to ensemble forecasting, where multiple runs of the same model with perturbed initial conditions or physics provide a range of possible outcomes. The spread of the ensemble members, or the spread between different deterministic models, directly quantifies the forecast uncertainty. For example, if one model predicts 10 m/s and another 20 m/s, it suggests a high degree of uncertainty, and a decision based on a single model's output could be risky.

Instead of seeking a single 'correct' forecast, the objective is to understand the range of plausible scenarios. This allows for better risk assessment and more robust decision-making. The Wind Agent's exceedance fan and Agreement Spine are designed to present this disagreement as actionable information, helping users understand the probability of exceeding their operational limits.

The Wind Agent's Agreement Spine is specifically designed to distil the complex information of model disagreement into an easily interpretable format. It presents the hourly wind speed and gust forecasts from multiple independent models (e.g., ECMWF, GFS, ICON) side-by-side.

By visualising these forecasts together, the Agreement Spine allows you to quickly identify:

• Consensus: When all models show similar speeds and timings, indicating high confidence.
• Divergence: When models show different magnitudes or timings, highlighting periods of higher uncertainty.
• Outliers: If one model consistently predicts significantly higher or lower values than the others, which might indicate a unique scenario or a model struggling with a particular atmospheric setup.

For example, if the Agreement Spine shows ECMWF predicting 12 m/s, GFS 10 m/s, and ICON 14 m/s for a specific hour, it immediately communicates a range of 10–14 m/s. If, for the same hour, one model predicts a peak at 14:00 and another at 17:00, it highlights a timing uncertainty. This direct comparison helps you to form a more complete picture of the forecast confidence, rather than relying on a single source. It is an essential tool for assessing the robustness of any wind-sensitive decision.