Wind Speed Dataset with Observation- And Copula-based Bias Correction and Multi-Model Fusion at 23 U.S. Locations

This dataset provides two related products of bias-corrected hourly wind speed at 23 U.S. sites spanning coastal, inland, mountainous, and plains regions, designed for energy applications and wind-related research. Physics-based models carry systematic biases relative to observations from spatial resolution, parameterization uncertainty, and numerical schemes; they tend to underestimate extremes and misrepresent seasonal patterns. Accurate correction therefore has to span the full distribution, not just the mean. Both products use copula-based bias correction trained and evaluated against hourly in situ measurements at variable hub heights.

The first product corrects Sup3rCC downscaled climate model output for 2018-2020. It models the marginals with the parametric Bulk-and-Tails (BATs) distribution, which captures both central behavior and extremes, and links modeled and observed wind speed with a Gumbel mixture copula. Unlike data-driven approaches such as quantile mapping, the parametric marginals allow reliable correction in the tails, where observations are sparse. The second product covers 2015-2020 and combines ERA5 (ECMWF) and MERRA-2 (NASA GMAO). Each is copula bias-corrected and then fused with Bayesian Model Averaging (BMA), which assigns performance-based weights to each source. The result is a probabilistic integrated product that quantifies uncertainty and outperforms either reanalysis alone, since neither is uniformly superior across sites and seasons. To preserve site anonymity, locations are reported at the state level, exact coordinates and measurement heights are available on request. For more information see the GitHub Repo and Powerpoint resources below.