Best Match: blending models into one series

The Wind Agent's 'Best Match' is a proprietary algorithm that selects and stitches together data from multiple numerical weather prediction (NWP) models to create a single, continuous forecast timeline. Its primary goal is to provide the most reliable forecast available at any given time and location, drawing on the strengths of different models across various forecast horizons and geographical regions.

What it means:

• Optimised selection: For each forecast hour, Best Match evaluates available models (e.g., ECMWF HRES, GFS, ICON-EU, UKMO) and selects the one deemed most accurate for that specific time, location, and forecast lead time. This selection is based on historical performance metrics and known model characteristics.
• Seamless timeline: The output is a single, uninterrupted series of forecast values (speed, gust, direction, temperature, etc.), designed to be easy to interpret without requiring the user to compare multiple model outputs directly.
• Transparency: Every value presented by Best Match is attributed to its source model, and the selection criteria are based on objective, performance-driven rules, not subjective adjustments.

What it does not mean:

• A new model: Best Match does not run its own NWP model. It is an intelligent aggregation and selection layer built on top of existing, publicly available model data.
• A 'perfect' forecast: No forecast is perfect. Best Match aims to reduce error by leveraging model diversity, but it cannot eliminate inherent uncertainties in atmospheric prediction.
• A 'smoothed' average: Unlike an ensemble mean, Best Match typically selects a single best-performing model for each point, rather than averaging multiple models. This preserves the detail and resolution of the chosen model.

Different NWP models excel in different geographical regions and over varying forecast horizons. Best Match leverages these known strengths to make informed selections. For instance, a high-resolution regional model might perform better for short-range forecasts (e.g., 0-48 hours) over its domain, while a global model might be superior for longer lead times (e.g., 72-168 hours) or for areas outside specific regional model domains.

Typical selection criteria include:

• Geographical domain: Regional models like ICON-EU (for Europe) or AROME (for France) offer higher spatial resolution and often better representation of local weather phenomena within their specific areas compared to global models like GFS or ECMWF HRES. For Ireland, models with high resolution over the North Atlantic and Western Europe are prioritised.
• Forecast horizon (lead time): Short-range forecasts (up to ~48 hours) often benefit from very high-resolution models that can capture mesoscale features. Medium-range forecasts (48-168 hours) typically rely on global models with robust data assimilation and ensemble capabilities. For example, ECMWF HRES is commonly cited as a leading global model for medium-range accuracy.
• Resolution and update frequency: Models with finer grid spacing can resolve smaller-scale features, which is crucial for detailed local forecasts. Models that update more frequently (e.g., every 3 or 6 hours) provide fresher data for short-range predictions.

Best Match dynamically assesses these factors. For example, for a forecast in County Kerry for the next 24 hours, it might prioritise a regional model with a 2.5 km resolution. For a forecast 5 days out, it would likely switch to a global model like ECMWF HRES, which has a coarser resolution (e.g., 9 km) but is known for its skill in the medium range. This adaptive selection ensures that the most appropriate model is chosen for each segment of the forecast timeline.

When Best Match transitions from one model to another within the forecast timeline, it performs a 'stitch' to ensure continuity. This process is not a simple cut-and-paste, as abrupt changes in forecast values could be misleading. Instead, a blending or smoothing technique is applied over a short transition window to minimise discontinuities. The goal is to make the transition as imperceptible as possible while retaining the integrity of the underlying model data.

Consider a scenario where a regional model (Model A) is used for the first 48 hours, and a global model (Model B) takes over from 48 to 168 hours. If at hour 48, Model A predicts 10 m/s wind and Model B predicts 12 m/s, an abrupt jump of 2 m/s would occur. Best Match addresses this by:

1. Identifying the crossover point: The hour where the primary model switches.
2. Applying a weighted average: Over a short window (e.g., 6-12 hours) around the crossover, a weighted average of the two models' predictions is calculated. For example, at hour 45, the forecast might be 90% Model A, 10% Model B. At hour 48, it might be 50% Model A, 50% Model B. At hour 51, 10% Model A, 90% Model B.

Worked example: Assume at hour 48, Model A predicts 10 m/s and Model B predicts 12 m/s. We use a 6-hour blending window (3 hours before and 3 hours after the switch at hour 48).

• Hour 45: Model A (10 m/s), Model B (11.8 m/s). Weighting: 100% A, 0% B. Blended: 10 m/s.
• Hour 46: Model A (10.2 m/s), Model B (12.0 m/s). Weighting: 75% A, 25% B. Blended: (0.75 * 10.2) + (0.25 * 12.0) = 7.65 + 3.0 = 10.65 m/s.
• Hour 47: Model A (10.5 m/s), Model B (12.2 m/s). Weighting: 50% A, 50% B. Blended: (0.50 * 10.5) + (0.50 * 12.2) = 5.25 + 6.1 = 11.35 m/s.
• Hour 48: Model A (10.0 m/s), Model B (12.0 m/s). Weighting: 25% A, 75% B. Blended: (0.25 * 10.0) + (0.75 * 12.0) = 2.5 + 9.0 = 11.5 m/s.
• Hour 49: Model A (9.8 m/s), Model B (11.9 m/s). Weighting: 0% A, 100% B. Blended: 11.9 m/s.

This method ensures a smooth transition, preventing artificial spikes or drops in the forecast that could arise from simply switching models abruptly. The 'Agreement Strip' (chart id: agreement_strip) on The Wind Agent shows which model is contributing to the Best Match at each hour, and its agreement with other models.

While Best Match primarily selects a single model, the blending process at the seams, as described above, can lead to a smoothing of extreme values. This is an inherent characteristic of any averaging or weighting technique. If one model predicts a very high gust and another a moderate one, a blend between them will result in a value less extreme than the highest prediction.

This smoothing effect is generally minor for Best Match because the blending window is short and only applied at transition points. However, it is a key difference when comparing Best Match to an ensemble mean, where all ensemble members are averaged across the entire forecast period. An ensemble mean, by its nature, will almost always be smoother and less extreme than any individual ensemble member, as the averaging process tends to cancel out the highest and lowest predictions.

Implications for users:

• Reduced peak values: If a specific model in the blend predicts an isolated, very sharp peak in wind speed or gust, the blending at a transition point might slightly reduce this peak if the incoming model is less extreme.
• Increased confidence in continuity: The smoothing ensures that the forecast timeline flows logically, reducing the likelihood of sudden, unexplained changes that could undermine user confidence.

Users should be aware that while Best Match aims for accuracy, any blending at the seams can slightly moderate the most extreme values predicted by a single model. For critical operations, consulting the raw ensemble data (e.g., via the 'ensemble_plume' chart) can provide a fuller picture of the range of possible outcomes, including higher extremes that might be present in individual ensemble members.

It is crucial to distinguish Best Match from an ensemble mean. Both techniques combine information from multiple models, but their methodologies and resulting characteristics differ significantly.

Best Match (Selection & Stitching):

• Methodology: Selects the single 'best' performing deterministic model for each forecast point based on predefined criteria (region, horizon, historical skill). Blends only at the transition points between selected models.
• Resolution: Tends to maintain the resolution and detail of the chosen high-resolution deterministic models.
• Extremes: Generally preserves extremes, as it picks a single model. Smoothing only occurs over short blending windows.
• Uncertainty: Does not directly quantify uncertainty itself, but relies on the skill of the selected deterministic model. Uncertainty is shown via the Exceedance Fan or ensemble plume.

Ensemble Mean (Averaging):

• Methodology: Averages the predictions from all members of an ensemble forecast (e.g., 50 members of ECMWF ENS, 21 members of GFS GEFS). Each member represents a slightly different initial condition or model physics.
• Resolution: Inherently smoother and lower resolution than individual deterministic models, as it averages out small-scale features.
• Extremes: Always smooths out extremes. The ensemble mean will almost always predict lower peak winds and higher minimums than the most extreme individual ensemble member.
• Uncertainty: Provides a direct measure of uncertainty through the spread of its members. A wide spread indicates high uncertainty.

For example, if ECMWF HRES (a deterministic model often selected by Best Match) predicts a gust of 30 m/s, and the ECMWF ENS mean predicts 25 m/s, Best Match would likely show a value closer to 30 m/s (if HRES is selected), whereas the ensemble mean would show 25 m/s. The choice between Best Match and an ensemble mean depends on the user's need: for a single, high-resolution timeline, Best Match is preferred; for a comprehensive view of uncertainty and a smoothed outlook, the ensemble mean is more appropriate.

The Wind Agent's Best Match is the default forecast presented on most interfaces, offering a coherent and actionable single timeline. However, there are specific contexts where we either explicitly indicate the underlying model or, in some cases, refuse to show a blended forecast altogether, instead directing users to raw model data or ensemble products.

When Best Match is shown (default behaviour):

• Primary forecast display: The main hourly forecast chart, the 'Shear Glass' (chart id: glass) for height-matched wind, and the 'Exceedance Fan' (chart id: fan) all use the Best Match timeline as their primary input for deterministic values.
• General planning: For most day-to-day operations and planning, the Best Match provides a robust and easy-to-understand forecast.

When we explicitly indicate the source model:

• Evidence Records: Every value stored in an Evidence Record is explicitly tagged with its source model, ensuring full traceability and auditability.
• 'Agreement Strip' chart: This chart (chart id: agreement_strip) visually indicates which model is selected for Best Match at each hour and how other models compare.

When we refuse to show a blended forecast:

• High disagreement: If the underlying models show extreme divergence, indicating very high uncertainty, the system may flag this. While a blend might still be presented, the Agreement Strip will clearly show the model disagreement, prompting the user to consult ensemble products.
• Long-range uncertainty: For very long-range forecasts (e.g., beyond 7-10 days), the skill of any single deterministic model, and thus any blend, diminishes significantly. In such cases, the system will emphasise the ensemble plume (chart id: ensemble_plume) to highlight the wide range of possible outcomes, rather than offering a seemingly precise but potentially misleading single-value blend.
• Specific research or analysis: For advanced users or specific analytical tasks, direct access to individual model outputs (e.g., via the 'model_compare' chart) is always available, allowing for independent assessment beyond the Best Match algorithm.

A core principle of The Wind Agent is transparency and traceability. For every forecast value presented as part of the Best Match timeline, the system meticulously records its provenance. This means that users can always determine which specific underlying NWP model contributed each data point.

This detailed recording of provenance is critical for several reasons:

• Auditability: For operations requiring strict adherence to procedures or for post-event analysis, knowing the source of each forecast value is essential. This allows for auditing against specific model performance or for understanding why a certain forecast was issued.
• Validation and trust: By clearly attributing each value, The Wind Agent builds trust. Users can cross-reference the selected model's performance or compare it with other data sources if they choose. This also helps in understanding the strengths and weaknesses of different models over time.
• Decision support: Knowing the source model can inform decision-making, particularly when users have specific experience or preference for certain models in particular conditions or regions.

This provenance information is integral to the 'Evidence Records' feature of The Wind Agent. When a user creates an Evidence Record, not only are the forecast values at that moment captured, but also the specific model (e.g., 'ECMWF HRES', 'GFS', 'ICON-EU') that supplied each value in the Best Match timeline. This ensures that a complete and verifiable history of the forecast is preserved, supporting operational compliance and retrospective analysis.