Reducing Wind Power Prediction Uncertainty with Atmospheric Characterization and Machine Learning Models

Accurate and precise wind power estimates are necessary during preconstruction phases of development to inform stakeholders’ decisions regarding turbine selection and to deliver an accurate assessment of the project’s overall economic viability. Unfortunately, power prediction uncertainty remains an industry challenge, leading to wind farm underperformance bias, in which a project’s operational energy yield is lower than expected prior to construction. Manufacturer’s Power Curves (MPCs) are often created to predict power as a function of hub-height wind speed alone. However, some research shows that power prediction may be improved by replacing the hub-height wind speed term with Rotor Equivalent Wind Speed (REWS) which accounts for variations in wind speed throughout the rotor-layer and the corresponding changes in turbine available power. It is also well known that atmospheric conditions other than wind speed impact turbine efficiency. The aim of this work is to understand how a variety of key atmospheric variables may be used to help better predict a turbine’s power output and, thus, reduce power prediction uncertainty. The dataset used in this study consists of measurements from a scanning Doppler wind lidar, a meteorological tower, a microwave radiometer, and several flux towers, obtained in front of a 2MW coastal turbine during the VERTical Enhanced miXing campaign (VERTEX).

A new method is introduced to reconstruct 10-minute average vertical wind profiles upwind of the turbine using Scanning Doppler wind lidar and Optimal Interpolation. Preconstruction power estimates are calculated for both hub-height wind speed and REWS predictions, and results are compared to the actual power produced by the turbine. Further, machine learning algorithms are introduced to group atmospheric conditions and relate them to 10-minute average power production over a two-month period. Specifically, clustering algorithms are used to create an AtmosMath classification system which groups vertical wind profiles based on profile shear conditions, concurrent turbine efficiency, and coexisting mesoscale patterns. These classifications elucidate relationships between atmospheric conditions, weather events and turbine performance. Decision trees and random forests are used to predict turbine performance based on wind direction, veer, shear, wind profile deviation from logarithmic shape, and atmospheric stability in addition to hub-height wind speed or REWS. The reduction in power prediction uncertainty achieved by the machine learning algorithms is quantified. In addition, the results of the decision tree and feature reduction algorithms are also used to determine the relative importance of measuring and characterizing each of these variables during different synoptic conditions.

Results demonstrate that hub-height speed alone overpredicts the turbine’s power by an average of 169.6 kW. REWS improves power estimates by reducing overprediction by 1% compared to the hub-height wind speed approach. Further, 10-minute average wind profiles that fit well to the expected logarithmic shape with low shear are associated with the least overprediction compared to other classified shear profiles. Results of the machine learning models demonstrate that incorporating additional variables reduces overprediction of wind power by up to 15%. Furthermore, the relative importance of each measured variable varies during different weather patterns. Therefore, the results of this work demonstrate that using a machine learning approach to incorporate a variety of atmospheric variables in power prediction significantly reduces prediction uncertainty and may be used to develop cost-effective, accurate, and precise power prediction models.