Downscaling to a site: Bridging the gap between model and reality

Numerical Weather Prediction (NWP) models, such as the ECMWF IFS model that underpins many forecasts, represent the atmosphere on a grid. For Ireland, a common resolution for global models is approximately 9 km. This means each grid cell covers an area of 81 square kilometres. Within such a large area, there can be significant variations in topography, surface roughness, and local sheltering or exposure.

The model's output for a grid cell is an average representation of the conditions across that entire area. It cannot resolve individual hills, valleys, buildings, or even large forests that might be present within the cell. For example, a site located at the top of a coastal cliff within a 9 km cell will experience significantly different wind speeds and directions than a site in a sheltered valley just a few kilometres inland, even if both fall within the same model grid cell. The model's 'average' wind for that cell will not accurately reflect either extreme.

This discrepancy is particularly pronounced in areas with complex terrain or heterogeneous land use. Without downscaling, a direct reading from the model for a specific point within a grid cell can lead to systematic biases, consistently under- or over-predicting wind conditions depending on the site's unique characteristics relative to the cell average.

Statistical downscaling involves using historical observations from a specific site to develop a relationship between the site's measured wind and the corresponding NWP model output for that grid cell. This method relies on a sufficiently long and high-quality record of local measurements.

The process typically involves:

1. Data Collection: Gathering concurrent historical wind observations (speed and direction) from the target site and the corresponding model reanalysis or forecast data for the same period and grid cell.
2. Relationship Modelling: Developing a statistical model (e.g., linear regression, quantile mapping, or machine learning algorithms) that maps the model's predicted values to the observed values. This model captures systematic biases and local effects.
3. Application: Applying the derived statistical relationship to future model forecasts to adjust them for the specific site.

For example, if a site consistently measures wind speeds 1.2 times higher than the model's prediction for a given wind direction, the statistical model would apply a factor of 1.2 to future forecasts from that direction. This method is effective for capturing consistent local effects but assumes that the relationship between the model and the site remains stable over time and that the historical data is representative. It cannot account for changes in the site's environment (e.g., new buildings, tree growth) unless the statistical model is retrained.

Dynamical downscaling involves running a higher-resolution NWP model over a smaller domain, nested within a coarser global or regional model. This finer-scale model can explicitly resolve smaller topographical features and land-use variations, providing a more detailed representation of local wind flows.

For instance, a global model might operate at a 9 km resolution, while a nested regional model could run at 1 km or even a few hundred metres. This higher resolution allows the model to better simulate:

• Orographic effects: How wind flows over and around hills and mountains.
• Coastal effects: Sea breezes, land breezes, and the interaction of wind with coastlines.
• Urban effects: Channelling and sheltering by buildings.

While dynamically downscaled models offer a more physically consistent approach to resolving local wind patterns, they are computationally intensive and require significant resources. The output from such models can still benefit from further local adjustments, as even a 1 km resolution model may not capture all micro-scale effects at a specific point. For example, a 1 km model would still average wind conditions over a square kilometre, potentially missing the exact wind conditions at a specific turbine location or crane site within that area.

Beyond model resolution, specific adjustments can be applied to account for the immediate local environment. These are often based on terrain and surface roughness characteristics derived from high-resolution digital elevation models (DEMs) and land-use databases.

1. Terrain Effects: Wind flow is significantly altered by topography. Wind speeds tend to accelerate over hilltops and ridges (speed-up effects) and decelerate in valleys and on leeward slopes (sheltering effects). Direction can also be channelled by valleys or deflected by obstacles. These effects can be modelled using computational fluid dynamics (CFD) or simpler empirical models based on terrain slopes and exposure.
2. Surface Roughness: The type of surface (e.g., water, short grass, forest, urban area) determines its aerodynamic roughness, which dictates how much friction the wind experiences. Rougher surfaces slow the wind more effectively and create greater turbulence. Adjustments can be made using roughness length (z₀) values, which are typically incorporated into wind profile laws (e.g., logarithmic or power law profiles) to adjust wind speeds from a reference height to the specific height of interest, considering the local surface type.

For example, if a model predicts 10 m wind speed of 10 m/s over a grid cell classified as 'open land' (z₀ ≈ 0.03 m), but your specific site is within a dense forest (z₀ ≈ 1.0 m), the actual 10 m wind speed will be significantly lower due to increased friction. Adjustments apply a ratio based on the roughness difference. Similarly, a site on an exposed headland will experience higher speeds than the model's average for a coastal cell.

The influence of local terrain and roughness is often highly dependent on the wind direction. A site might be sheltered from northerly winds by a hill but fully exposed to southerly winds from the sea. Therefore, downscaling often incorporates sector-dependent correction factors.

These factors are derived by analysing observed wind data against model output for different wind direction sectors (e.g., 30° or 45° sectors). For each sector, a unique bias or scaling factor is determined. This allows for a more nuanced correction than a single, overall adjustment.

Worked Example: Consider a site where historical data shows the following biases:

• Wind from 0°–90° (N-E): Model consistently underpredicts by 15% (observed = 1.15 × model).
• Wind from 91°–180° (E-S): Model consistently overpredicts by 10% (observed = 0.90 × model).
• Wind from 181°–270° (S-W): Model is generally accurate (observed = 1.00 × model).
• Wind from 271°–360° (W-N): Model consistently underpredicts by 20% due to exposure (observed = 1.20 × model).

If the forecast predicts a 10 m wind speed of 15 m/s from 45° (NE), the corrected speed would be 15 m/s × 1.15 = 17.25 m/s. If the forecast is 12 m/s from 120° (SE), the corrected speed would be 12 m/s × 0.90 = 10.8 m/s.

These factors are typically determined through statistical analysis of long-term observation campaigns at the site, providing a robust, data-driven adjustment for local conditions.

The effectiveness of downscaling techniques, particularly statistical methods and sector-dependent corrections, is heavily reliant on the availability of high-quality, long-term local wind data. When such data is scarce or non-existent, the ability to accurately downscale model output is severely limited.

In data-poor environments, modellers must often rely on:

• Generic adjustments: Applying generalised terrain and roughness adjustments based on broad land-use classifications and simplified topographic models, which may not capture site-specific nuances.
• Analogue methods: Using data from meteorologically similar sites with available observations, though 'similarity' can be difficult to define and verify.
• Expert judgement: Relying on local meteorological expertise to estimate potential biases, which introduces subjectivity.

These approaches carry higher uncertainty. For critical operations, the absence of site-specific data can mean that model output, even with basic downscaling, remains an approximation. This underscores the value of establishing local monitoring stations to build a robust historical dataset for future downscaling and calibration efforts. The Wind Agent's grounded agent feature can help in this regard, by collecting and storing observations for later analysis against model output.

Transparency in how downscaling and corrections are applied is crucial for users to understand the origin and potential limitations of the wind data they receive. The Wind Agent aims to make these adjustments clear.

When a site-specific correction is applied, the instrument indicates that the displayed values are 'adjusted' or 'corrected'. This is distinct from raw model output. The underlying model data remains accessible for comparison, allowing users to see the magnitude of the adjustment.

For example, if a forecast for your site is 10 m/s, and the raw model output was 8 m/s, the instrument would indicate that a +25% adjustment has been applied. This allows the user to understand the source of the difference. The 'Agreement Spine' feature on The Wind Agent provides a visual representation of how well the model output aligns with observed data, which implicitly reflects the quality of any downscaling or calibration applied.

Users should always refer to their own operational documents and risk assessments. The Wind Agent provides the most accurate and transparent wind information possible, but the ultimate decision-making responsibility rests with the user, informed by their site-specific knowledge and procedures. Understanding the downscaling process helps in interpreting the data and assessing its applicability to specific operational thresholds.