Posters

Abstract

K2M has leveraged machine learning to assess the success of a WTG or plant optimisation strategy more accurately. Operational data recorded from 4 real-world wind farms is used to demonstrate that the machine learning method significantly reduces prediction uncertainty when compared to the traditional power prediction models most commonly used in the industry. Determining how successful an optimisation strategy is, relies on the accurate quantification of the resulting power production increase. Put simply, this quantification is achieved by comparing power production with and without the optimisation strategy applied. In practice this is done by producing power curves for a range of conditions both with and without the optimisation. These power curves are then used to build up a picture of production for a long-term period at the considered turbine location. The industry standard method of power prediction relies on nacelle anemometer bin-averaged power curves derived from operational SCADA data recorded at the WTGs. This assumes average wind speed to be the sole driver of energy production. Various approaches to improving the traditional method are available, such as binning by direction sector or calendar month as proxies for certain atmospheric conditions, but the accuracy improvement is limited, and highly site dependent. From consideration of operational data, K2M has seen a variation of up to half a megawatt at the peak-cp part of the power curve due to the variation of environmental conditions. The main limitation of the traditional approach is that it does not account for the factors affecting performance variation such as air density, turbulence, wind shear, and rotor inflow angle. To fully capture the influence of all performance-affecting conditions, while keeping uncertainty to a minimum, a balance must be found between the number of conditions, the size and number of the bins, and the amount of data considered; a process which contains subjectivity. The machine learning approach creates an optimal number of power curves, with optimal bin size, for the most performance-affecting environmental conditions. K2M has demonstrated the value of this approach by comparing machine learning power predictions against real measurements produced when WTGs are operating normally at a variety of different wind farm sites. K2M's machine learning approach has been considered against each of the industry standard binned power curve methods and consistently demonstrates a reduction in prediction error when compared to the other methods. Due to this the machine learning power prediction model can be confidently applied to more accurately quantify production uplift from WTG and plant optimisation strategies.