Rich Data for Wind Turbine Power Performance Analysis

As the trend towards larger wind turbines continues, there is a growing need to refine and update the methods used for evaluating their performance. Variation in the input conditions over the vertical extent of the rotor leads to considerable uncertainty in the results of performance tests which appears as unwanted dispersion in the traditional power curve scatterplot. Measures to reduce this dispersion would therefore be beneficial. This project examines the effects traditional filtering criteria. It finds that these filters reject data from the centre of the distribution rather than the periphery, and therefore lead to a relative increase in data dispersion. Alternative criteria based on stationarity, intermittency, isotropy and normality show similar effects to those using turbulence intensity and shear, but do not lead to any significant improvement. A high-dimensional dataset representing over 700 factors is prepared and progressively narrowed through correlation analysis and global sensitivity analysis. The eFAST method is applied to an artificial neural network trained with quality-controlled time series data. The eFAST method represents input parameters as synthetic periodic signals so that their relative influence can be determined by applying a Fourier transform to the model output. The sensitivity of the model to a factor is reflected in the amplitude of its corresponding spectral peak. The main contribution of this work is the identification through this process of 16 meteorological factors shown to have an influence on power production. They include previously-known factors as well as others related to turbulence and atmospheric stability. A limitation of the approach is that identifying pairwise interactions it not possible. This is a major limitation of this project, and is recommended as a topic for future research.

随着大型风力涡轮机的发展趋势持续，对优化和更新其性能评估方法的需求日益增长。转子垂直范围内的输入条件变化会给性能测试结果带来显著不确定性，这种不确定性在传统功率曲线散点图中表现为不必要的数据离散度。因此，降低此类离散度的措施具备实际应用价值。本项目探究了传统滤波准则的影响。研究发现，这类滤波器会剔除分布中心区域的数据而非边缘区域，反而相对增大了数据离散度。基于平稳性、间歇性、各向同性和正态性的替代准则，虽与基于湍流强度和风剪切的准则呈现相似效果，但并未带来显著改善。本研究构建了包含700余个因子的高维数据集，并通过相关性分析与全局敏感性分析逐步筛选精简。将扩展傅里叶振幅灵敏度测试（extended Fourier Amplitude Sensitivity Test，eFAST）方法应用于经质量控制的时间序列数据训练得到的人工神经网络。该方法将输入参数表征为合成周期信号，通过对模型输出进行傅里叶变换即可确定各参数的相对影响程度。模型对某一因子的敏感性，可通过其对应频谱峰的振幅加以体现。本研究的主要贡献在于，通过该流程识别出16个对风力发电有影响的气象因子，其中既包含已知的影响因子，也涵盖与湍流和大气稳定性相关的新型因子。该方法的局限性在于无法识别成对交互作用，这是本项目的一大主要局限，建议将其作为未来的研究方向。