RWE put our models to the test
Across 170 sites worldwide, Whiffle Meso and LES delivered 30% less wind speed uncertainty than ERA5.*
*Whiffle LES vs ERA5 on the 71-site comparison subset. Full numbers and methodology below.
170 sites validated, globally 300+ met mast measurement locations 170 years of LES simulation data analysed 60,000+ parallelised 24-hour simulations. RWE built the largest measurement database of its kind, selected the 170 sites for the study, and defined the inclusion criteria – at least 90 days of model-observation overlap, heights with at least 50% data availability, timestamps where all heights reported valid measurements. Whiffle submitted model version 13, our default Whiffle Wind configuration: Mesoscale at 2 km, LES nested at 100 m, with site-specific domains and standard spin-up. No tuning to the sites, no adjusted parameters – the same model any customer runs.
The results of the landmark study
In their comprehensive validation, RWE’s benchmark compared nine modelled datasets, including Whiffle LES and Whiffle Meso. The results below focus on the 71-site subset where Whiffle Meso, Whiffle LES, and ERA5 outputs were all concurrent.
30% less wind speed uncertainty than ERA 5
Whiffle Meso competitive for standard sites
Whiffle LES earns it place in complex terrain
Whiffle LES shows lowest AEP error after bias correction
30% Lower wind speed uncertainty
Both Whiffle models return mean bias close to zero. ERA5 carries a systematic negative bias of −0.53 m/s. On error spread, Whiffle LES returns σ=0.44 m/s and Meso σ=0.48 m/s, against σ=0.66 m/s for ERA5. A narrower error distribution means lower wind speed uncertainty in your P90 estimate — and a smaller correction to apply in the first place.
LES closest to observed mean wind speeds
Each point compares a model’s mean wind speed at a site to the measured mean. Whiffle LES sits tightest to the 1:1 line — within ±5% at most sites — meaning the model is consistently close to observed mean wind speeds across the 71-site subset. ERA5’s systematic underprediction is visible as a downward shift below the line.
Scaled power bias close to zero
A standard long-term correction removes the mean wind speed error. The scaled power bias measures what remains — how well the model captures the shape of the wind speed distribution. Because AEP scales non-linearly with wind speed, a distribution wrong in shape still produces systematic yield error after correction. It is the most direct proxy for residual AEP error available from this dataset.
Both Whiffle models show median scaled power bias close to zero across all site types. LES shows a somewhat tighter spread. An offshore positive bias is present in both models, consistent with overprediction at sub-rated wind speeds. This is identified, understood, and in active development.
What 30% less uncertainty means for you
Wind speed uncertainty flows directly into P90, and P90 flows directly into debt sizing, cost of capital, and how competitively you can bid. A narrower error distribution — across site screening, your long-term reference, and full resource modelling — means a less conservative P90 that sits closer to P50, a lower uncertainty allocation from lenders, and a more competitive LCOE on your bid.
How much of that translates to your project depends on where Whiffle sits in your workflow, your campaign length, your uncertainty budget composition, your lender’s methodology, and how much of the 71-site aggregate gain holds up at your specific site. The σ reduction is the signal — the project-level impact is a conversation.
What comes next?
The RWE benchmark is the baseline. The industry is asking harder questions about model accuracy. We think that’s exactly right.
Joint Ørsted–Whiffle cross-prediction study
Extending the question from “how accurate at the mast” to “how accurate across the site. Presented at WindEurope Annual Event 2026.
Joint RWE–Whiffle scientific paper
Building on results presented at WindEurope Annual 2025 in Bilbao