Machine learning and statistical approaches for wind speed estimation at partially sampled and unsampled locations; review and open questions

Wind resource assessment (WRA) depends on the availability of accurate, long-term wind speed data. In locations where such data is limited (partially sampled locations, PSL) or completely missing (unsampled locations, USL), various physical, statistical, and machine-learning methods have been developed to address these gaps. This paper presents a comprehensive and up-to-date review of statistical and machine-learning methods for estimating long-term wind speed at PSLs and USLs.<br /> It was found that the “Measure Correlate Predict” (MCP) is still the method of choice for estimation at PSL. However, this approach has evolved with the adoption of machine learning, especially Artificial Neural Networks, and reanalysis wind data as the reference site. In general, reanalysis datasets have seen growing adoption for WRA at both PSLs and USLs due to their global coverage, high temporal resolution, and demonstrated accuracy. At USLs, uncorrected and bias-corrected reanalysis wind speed data are used for WRA, with the Global Wind Atlas predominantly used to correct reanalysis wind speed data. There is also a growing effort to develop machine learning models, including deep learning models for reanalysis bias correction at USLs using explanatory variables derived from high-resolution topographic and land use datasets.<br /> Challenges to estimating long-term wind speed at PSLs and USLs are identified and discussed: data uncertainties, disparity in the accuracy of reanalysis wind data, model transferability, and nonstationary conditions. Finally, recommendations for future research and development directions are presented, including techniques that consider documented non-stationarity in wind speed data.