Rich Data for Wind Turbine Power Performance Analysis

Abstract

The post-installation verification of wind turbine performance is an essential part of a wind energy project. Data collected from meteorological instruments and from the turbine is analysed to produce an estimate of the annual energy production (AEP) which is compared against expectations. However, turbine warranties can impose very strict data filtering criteria which can lead to high rates of data loss. As a consequence, measurement campaigns may last longer than expected and incur additional costs for the development. This project aims to investigate the extent of the problem and the potential of alternative data filtering strategies with respect to data loss, AEP estimates and the dispersion of points in the power curve scatter plot. In doing so, it targets a wide range of meteorological parameters with theoretical relationships to wind turbine power production with particular interest in those not accounted for in the current standard. The identification of viable filtering strategies with lower data loss would provide significant benefits to wind energy development projects in terms of greater control over timescales and reduced costs. Data from a sample of power performance tests is analysed to explore the range and severity of the problem of data loss. It confirms the wide variation in warranty conditions, demonstrates the extent and likelihood of data losses and quantifies the financial implications within the limits of commercial sensitivity. When indirect costs are taken into consideration, the impact of extended measurement campaigns can theoretically reach tens of millions of pounds. A new, high-fidelity dataset is then compiled so that the effects of alternative filtering strategies can be examined. The dataset covers the whole of 2017 and consists of over 700 parameters of which 74 are selected for investigation here. The eFAST method of global sensitivity analysis is used in combination with correlation analysis to reduce this number to 11 parameters which are then used to define alternative filtering criteria. Similar AEP estimates are obtained by application of conventional and experimental criteria to the research dataset. In the case of the experimental filters, however, the data loss was 11% compared to 63% data loss with conventional filters. Conventional filters were also shown to increase the dispersion in the power curve scatter plot by over 10%, while dispersion did not increase significantly with the experimental filters.