Transformation of reanalysis data for improved long-term estimation of wind speed and direction at a target site

Abstract

This paper proposes the use of measure-correlate-predict (MCP) methods based on supervised machine learning (ML) techniques to transform reanalysis data from ERA5 and MERRA2, aiming to improve the long-term estimation of wind speed and direction at locations with limited on-site measurements. The study analyzes modelsthat directly estimate the target variables—wind speed and direction—as well as two-stage models that first estimate the Cartesian components of wind velocity and subsequently transform them into polar coordinates.As a case study, hourly mean wind data recorded between 2001 and 2023 at 10 m above ground level are used. The data were collected from an anemometric station located on the island of Gran Canaria (Canary Archipelago, Spain). Key findings include the following: (a) Reanalysis data underestimate actual wind speeds and fail to adequately represent the mean wind direction; (b) although reanalysis data poorly represent the daily wind speed profile, the MCP model significantly corrects this, achieving a Pearson correlation of 0.994; (c) the MCP method minimizes the differences between observed and estimated values (7.2 m/s vs. 7.13 m/s, and 4.49◦ vs. 4.50◦, respectively); (d) the combination of ERA5 and MERRA2 consistently yields the lowest estimation errors, regardless of model type; (e) artificial neural networks outperform other ML techniques in all scenarios; and (f) the proposed method reduces the mean relative error in wind power density estimation to 13.89 %, compared to 43 % and 63.1 % using MERRA2 and ERA5 alone, respectively.