Posters

Abstract

With the continuous growth of wind energy installed capacity, the number of wind farms currently under-development and already in operation has increased and will continue to do so in the upcoming years. In order to manage this increasing number of assets and assure their proper operation, the wind industry needs to be equipped with tools and methods that allow for a fast and automated wind turbine performance monitoring. This objective has proven sometimes difficult to be accomplished as the available turbine operational data often lacks the quality and precision required to be used on a large-scale. With that in mind, this abstract presents an innovative approach on how operational data (wind speed, power, temperature, turbulence intensity, …) can be analysed in order to obtain a more reliable wind farm performance analysis usable at a large-scale and that could handle the known data quality issue. The method developed is composed of two main parts: firstly, a data quality analysis is applied on the historical wind speed data (main source of performance uncertainty) and secondly, based on those results, the overall turbine's performance is quantified by comparing a normalised power curve to a learned reference behaviour. The data quality analysis is obtained by building a turbine specific wind speed predictive model during a training period. The predictive model is based on Gaussian Process Regressions and kernel functions consisting of a combination of linear contributions (higher wind speed implies higher production and rotor speed for example), non-linear contributions (turbine dynamics) and white noise to fully model the impact of the multiple parameters on the wind speed. The data used as input to the model is composed of a combination of external condition (temperature, wind sector, turbulence intensity), turbine parameters (power, rotor speed, pitch angles) and neighbouring turbines parameters. The comparison between the predicted values and the wind speed SCADA measurements allows for a detection of any significant change in the wind speed during the analysis period. If a change is detected, a statistical analysis of the differences provides an estimation of the impact on the wind speed values that can then be used to "correct" the SCADA measurements. The developed method was initially tested on 2 different wind farms that had a met mast available on-site. The results demonstrated the model's ability to detect wind sensor changes (using the met mast wind speed as a reference) with a sensitivity of detection of around 0.3 m/s. After the initial validation, the method was applied to multiple operating wind farms where historical SCADA data of at least 24 months was available. The results allowed the detection of multiple wind sensor issues, provided a more stable performance monitoring and a more precise performance change estimation when compared to a non-corrected SCADA data analysis. In conclusion, this abstract presents an automated method able to use the largely available turbine data for a reliable performance monitoring, reducing the overall uncertainty of the existing methods by assuring a stable behaviour of the wind speed measurements throughout the analysis.