Scientists cracked the Andes wind puzzle by reading the sky “upstairs,” far above the turbines they were trying to predict
When engineers plan a wind farm, they typically pull atmospheric data from right around turbine height — the zone where the blades will actually spin. In the high Andes, that standard approach quietly breaks down. Across rugged ridgelines above 2,800 meters in southern Ecuador, one of South America’s least-tapped wind corridors, the most widely used reference dataset consistently underestimates wind speeds and skews energy projections by as much as 22%.
A research team found the fix not by improving instruments on the ground, but by looking somewhere unexpected — hundreds of meters higher in the sky.
Why standard ERA5 wind data fails in the mountains
ERA5 is the go-to global reanalysis dataset for wind energy planning. It holds up well offshore and over flat terrain, but mountains expose its limits fast. The dataset’s spatial resolution of roughly 31 km smooths ridges, peaks, and valleys into averaged grid cells — systematically underestimating wind speeds at sites sitting above the model’s flattened terrain.
At those elevations, the air flowing over the turbines isn’t surface air shaped by local friction — it behaves more like free atmospheric flow.
The standard outputs most developers use — wind components at 10 m and 100 m above ground — were designed for relatively uniform surfaces. Complex terrain tells a different story. Mountain waves, katabatic flows, and valley channeling can dominate local wind behavior yet stay largely invisible to a coarse-resolution model.
Ecuador’s stakes are high. Wind power supplies just 0.6% of national electricity, while the grid leans heavily on hydropower — a vulnerability that severe droughts in 2023 and 2024 exposed painfully. Accurate wind resource assessments aren’t an academic exercise there; they’re a resilience question.
Looking up: the case for higher atmospheric levels
The research team started with a physical intuition. Wind farm sites in the Andes sit on exposed ridgelines, often thousands of meters above surrounding valleys. At those elevations, the air flowing over the turbines isn’t surface air shaped by local friction — it behaves more like free atmospheric flow. Data from higher up in ERA5’s vertical structure should therefore correlate better with what instruments on the ground actually measure.
ERA5’s model-level dataset offers 137 vertical layers. The team tested 33 of those levels, spanning roughly 10 m to 3,000 m above ground, computing how well ERA5 wind speeds at each level matched observed wind speeds at 80 m hub height across four mast sites.
The result was an inverted parabola. Correlation started weak near the surface, climbed through the lower levels, peaked between 600 and 1,600 m above ground, then fell off. At site M3 — an exposed peak in the southwest — correlation rose from just 0.27 at hub height to 0.855 at the optimal upper level.
Four masts, four ridgelines: testing the idea in the field
To move from hypothesis to evidence, the team needed real observations. Ecuador’s national electricity corporation, CELEC EP, operates four meteorological masts across the southern Andes, spanning elevations from 2,829 m to 3,796 m above sea level. Each mast recorded hourly wind speeds at 80 m height over four years, January 2021 through December 2024.
The team trained site-specific random forest models on three years of data, then validated predictions against the untouched fourth year. Two model versions were built: one using standard ERA5 single-level data at 10 m and 100 m, and one drawing wind speed from the optimal upper ERA5 model level for each site. Linear regression ran alongside as a conventional benchmark.
The numbers: sharper wind predictions, more reliable energy estimates
The upper-level approach won across every metric. Using the optimal ERA5 model level improved the Perkins skill score by an average of 59% compared to single-level inputs, while R² improved by 40% and mean absolute error dropped by 23%.
The energy numbers were telling. Standard ERA5 data produced bias as high as 22% in annual energy production calculations; the optimal model-level approach cut that figure to less than 7% — roughly a threefold improvement. Gains were largest at well-exposed ridge-top sites M1 and M3, precisely the locations most attractive for wind farm development. The random forest algorithm outperformed linear regression overall, though linear regression handled extreme wind speed values slightly better.
A low-cost fix with broader implications
One of the most useful things about this finding is what it doesn’t require. The ERA5 model-level dataset is freely available through ECMWF’s Climate Data Store — no new field campaigns, no expensive mesoscale simulations, no proprietary data. For wind developers working in data-sparse regions or under budget constraints, that accessibility matters considerably.
The physical logic isn’t unique to Ecuador. High-altitude wind corridors exist across the Andes, the Himalayas, the Alps, and other major ranges. The authors note that more mast data from diverse terrain types are needed before claiming broad applicability.
For Ecuador, the findings add technical weight to a strategic argument. Andean wind is an underused resource in a country whose hydropower dependence has already proven fragile. Accurate wind resource assessment is a prerequisite for grid resilience — and better data was available all along, at a higher altitude.
Full study: Contreras, J., van Lipzig, N., Samaniego, E., and Ballari, D.: Identification of optimal ERA5 model level for wind resource assessments in mountainous terrain, Wind Energ. Sci., 11, 3823–3840, https://doi.org/10.5194/wes-11-3823-2026, 2026.
Daniel García is an Editor-in-Chief with strong expertise in structural work and engineering principles. He combines this technical foundation with deep knowledge of energy, spatial design, and emerging technologies, bringing a forward-thinking and analytical approach to editorial leadership.
