Methods for high-accuracy wind resource assessment to support distributed wind turbine siting

Abstract. Public wind resource datasets are central to wind energy planning, particularly for distributed wind installations where it may be infeasible to collect on-site measurements or run bespoke simulations. Yet despite their broad use, the site-level accuracy at hub height remains only partially quantified. This work addresses this gap in two steps: (i) we develop a unified, observation-based benchmark to evaluate the performance of the most common models used in industry, and (ii) we propose a new machine-learned ensemble approach that leverages multiple models to synthesize improved estimates that address the shortcomings of individual models. For each dataset and observation series we form long-term empirical wind speed quantiles. This quantile representation allows us to compare products with different periods of record without requiring temporal overlap and evaluate both wind speed distribution errors and site-level mean biases. Results show that the ensemble method reduces quantile-dependent mean bias to near zero across the distribution and lowers mean absolute bias in long-term mean wind speed by roughly one-third relative to the best-performing individual dataset. Finally, we use the trained model to produce a national, gridded set of wind speed quantiles for the publicly accessible WindWatts platform. Together, the benchmark, ensemble model, and deployment dataset demonstrate that machine learning can meaningfully correct and combine existing public datasets, providing more reliable, distributional wind resource information for early-stage assessment and planning.