Hidden Winds Aloft: Higher ERA5 Levels Unlock Mountain Wind Power Secrets

High in the tropical Andes of southern Ecuador, where ridgelines rise above 3,000 metres and the wind howls across páramo grasslands, one of the most basic questions in wind energy has just received a surprising answer. A research team led by Juan Contreras of the Universidad del Azuay, working with colleagues at KU Leuven and the Universidad de Cuenca, has shown that the standard way of estimating wind speeds for wind farm planning in mountains is looking in the wrong place in the atmosphere. Instead of relying on the near-ground wind data that the wind industry has treated as gospel for years, the researchers found that wind speeds from hundreds of metres higher up in a widely used global reanalysis dataset predict conditions at turbine height far more accurately. The finding, published in the journal Wind Energy Science, could reshape how wind resources are assessed in complex terrain worldwide.

The dataset at the heart of the study is ERA5, the fifth-generation atmospheric reanalysis produced by the European Centre for Medium-Range Weather Forecasts. Reanalysis products combine numerical weather prediction models with decades of historical observations to create a continuous, globally consistent record of the atmosphere, and ERA5 has become a standard reference for the wind energy industry because it offers hourly data stretching back to 1940. Most practitioners, however, use only the so-called single-level dataset, which provides wind components at 10 and 100 metres above the ground. That convention works reasonably well offshore and over flat land, where the model’s roughly 31-kilometre grid cells capture the flow without much distortion. In mountains, it breaks down badly.

The problem is one of geometry. ERA5’s coarse grid smooths away the jagged detail of real topography, so the elevation the model assigns to a grid cell can differ from the true elevation of a prospective wind farm site by hundreds of metres. In the Ecuadorian study area, those discrepancies ranged from about 130 metres to more than 1,100 metres. Because numerical weather models are highly sensitive to their lower boundary conditions, this smoothing means terrain-driven phenomena such as anabatic and katabatic slope winds, mountain waves and valley channelling are simply lost. Previous research has shown that reanalysis products tend to underestimate wind speeds at sites above the grid-cell mean elevation and overestimate them at sites below it, and that ERA5 systematically underestimates both wind speeds and their variability in mountainous regions.

Contreras and his colleagues hypothesised that if the model’s surface is effectively buried beneath smoothed terrain, then the winds that best represent a real ridge-top site might be found not at 10 or 100 metres above the model’s ground, but much higher in the model’s vertical column. ERA5’s model-level dataset offers exactly that opportunity: 137 vertical levels spanning the atmosphere from the surface up to roughly 80 kilometres, of which the team examined 33 levels between about 10 and 3,000 metres above ground. To test the idea, they turned to four meteorological masts operated by the Electric Corporation of Ecuador, perched at elevations between 2,829 and 3,796 metres across the provinces of Cañar, Azuay and El Oro, each instrumented with first-class cup anemometers measuring wind at 80 metres, a typical turbine hub height.

The correlation analysis produced a striking pattern. When the researchers compared hourly mast observations with ERA5 wind speeds at every model level, they found an inverted parabolic relationship: correlations rose steadily with height, peaked at levels roughly 600 to 1,600 metres above ground, and then declined. At one ridge-top site, the correlation climbed from a dismal 0.27 at the model level closest to the measurement height to 0.855 at the optimal level. Across the exposed sites, peak correlations reached between 0.79 and 0.88, far above anything achievable with the conventional single-level data. Crucially, control analyses at three flat coastal sites in Ecuador showed the opposite behaviour: there, the best correlations occurred near the actual hub height, confirming that the mountain result reflects a genuine topographic effect rather than a universal quirk of the dataset.

Identifying the optimal level was only the first step. The team then built site-specific machine learning models, using random forests to translate ERA5 wind speeds into calibrated estimates of hub-height wind, training on three years of hourly observations from 2021 to 2023 and validating against the independent year 2024. They compared two input configurations: the traditional single-level winds at 10 and 100 metres, and the wind speed at the optimal model level. A linear regression method with residuals, a mainstay of the industry’s measure-correlate-predict workflow, served as a benchmark. The random forest approach was chosen for its demonstrated robustness in complex terrain and its ability to mitigate the influence of interannual wind variability.

The improvements were substantial across every metric the team tested. Random forest models driven by the optimal model level achieved, on average, a 59 percent improvement in the Perkins skill score, which measures how well the predicted wind speed frequency distribution overlaps the observed one; a 40 percent improvement in the coefficient of determination; and 23 percent reductions in both mean absolute error and root mean square error compared with models using single-level inputs. The gains were largest at well-exposed ridge-top sites, precisely the kind of locations developers favour for wind farms, and smallest at a sheltered foothill site where local flow effects dominate. A direction-dependent analysis added a physical explanation: the strongest correlations occurred when the large-scale flow was aligned with each site’s prevailing wind sectors, suggesting the upper model levels capture the free atmospheric flow that sweeps over exposed peaks, unencumbered by the model’s blurred representation of the terrain below.

Perhaps most importantly for the industry, the team translated their wind speed estimates into money and megawatt-hours. Using the power curves of two turbines actually operating in Ecuadorian wind farms, the Goldwind GW70/1500 and the Vestas V112/3450, they calculated annual energy production from the calibrated wind speeds and compared it with production computed from the mast observations. The percentage error in annual energy production fell to between roughly 2 and 7 percent, compared with 8 to 22 percent when using conventional single-level data, a threefold reduction in error. The best result came at the highest-elevation site, where the error was around 2 percent, while the largest deviation, about 7 percent, occurred at a site sharing its ERA5 pixel with a better-represented neighbour, underscoring how much local terrain context still matters within a single 31-kilometre grid cell.

The study is not without limitations, and the authors are candid about them. Only four masts were available, most at well-exposed high-wind sites, so the results may not generalise to sheltered valleys or other mountain ranges. The four-year observation period is too short to capture longer-term climate variability, and the optimal heights were identified through a detailed search that demands considerable data processing. The random forest models also showed familiar weaknesses at the extremes of the distribution, underestimating the frequency of winds above 15 metres per second and struggling with the very calmest conditions, although the latter mattered little for energy calculations because the turbines’ cut-in speeds lie above those values. Future work, the researchers suggest, could feed multiple neighbouring model levels into the machine learning models to emulate wind shear effects, or incorporate sub-grid variables such as gravity wave activity, which other studies have flagged as important in mountainous flow.

Even so, the practical implications are hard to overstate. Wind power currently supplies just 0.6 percent of Ecuador’s electricity despite considerable Andean potential, and severe droughts in 2023 and 2024 exposed the fragility of a grid heavily dependent on hydropower. Because the method relies entirely on freely available reanalysis data and modest computational resources, it offers a cost-effective alternative to running bespoke mesoscale simulations, which are computationally expensive, cover limited regions and typically span fewer than two decades. For developers eyeing wind projects in the Andes, the Himalayas, the Alps or any other range where the wind blows strongest above the clouds, the message is clear: the most faithful picture of what a turbine will experience at hub height may lie far above the model’s notion of the ground, and the data needed to find it have been sitting in the archive all along.

Subject of Research: Improving hub-height wind speed estimation for wind resource assessment in mountainous terrain using ERA5 model-level data and machine learning

Article Title: Identification of optimal ERA5 model level for wind resource assessments in mountainous terrain

Article References: Identification of optimal ERA5 model level for wind resource assessments in mountainous terrain. (n.d.). https://doi.org/10.5194/wes-11-3823-2026

Image Credits: AI Generated

DOI: 10.5194/wes-11-3823-2026

Keywords: ERA5, wind resource assessment, reanalysis data, Andes, random forest, machine learning, hub-height wind speed, complex terrain, annual energy production, measure-correlate-predict, renewable energy, Ecuador