Electric Power Technology and Environmental Protection

【目的】系泊安全是影响海上漂浮式风力发电机稳定运行的关键。提前识别系泊失效特征、降低系泊监测难度和成本，有利于海上风电健康发展。【方法】本文使用 ITI Energy Barge 平台搭载 NREL 5 MW 风力机，建立了漂浮式风力机模型，并以该模型的系泊系统为研究对象，基于卷积神经网络（convolutional neural network，CNN）模型对系泊系统不同位置系泊失效后的响应数据进行特征自学习和分类，识别系泊失效特征，并对结果进行可视化分析。【结果】研究表明：CNN模型可从复杂载荷作用下漂浮式风力机的位移、速度和加速度 3 个响应变量中准确识别出各系泊失效特征。当 CNN 模型迭代 200次后，位移、速度及加速度的训练集与验证集的准确率均达到了98.0%以上，其中训练集中位移、速度及加速度这3个变量的损失分别为0.13、0.02和0.05，验证集中这3个变量的损失分别为0.20、0.05和0.05。在混入6种不同高斯白噪声后，CNN模型对风力机的 3 个响应变量的识别准确率仍保持 99.0% 左右，模型具有良好的鲁棒性；通过 t-SNE 可视化方法可逐步降 维观察3个响应变量的数据特征，其区分度随卷积的逐步深入变得明显，有利于失效特征识别。【结论】本文所述方法有利于提前识别系泊失效特征，保障海上漂浮式风力发电机稳定运行。

[Objective] Mooring safety is the key to the stable operation of offshore floating wind turbines. Identifying mooring failure characteristics in advance and reducing the difficulty and cost of mooring monitoring are conducive to the healthy development of offshore wind power. [Methods] this paper establishes a floating wind turbine model of ITI Energy Barge platform equipped with NREL 5 MW wind turbine. Taking the model mooring system as the research object, based on the convolutional neural network (CNN) model, the platform response data after mooring failure at different positions are self-learned and classified, the mooring failure characteristics are identified, and the results are visualized and analyzed. recognized under different mooring states. [Results] The results show that the CNN model can accurately identify the mooring failure characteristics from the 3 response variables of displacement, velocity and acceleration of the floating wind turbine under complex loads. When the CNN model is iterated 200 times, the accuracy of the training set and the verification set of displacement, velocity and acceleration reaches more than 98.0%. The loss of the training set of displacement,velocity and acceleration is 0.13, 0.02 and 0.05, and the loss of the verification set is 0.20, 0.05 and 0.05. After mixing 6 different Gaussian white noises, the CNN model still maintains about 99.0% recognition accuracy for the 3 response variables of the wind turbine, and the model has good robustness. The t-SNE visualization method can gradually reduce the dimension to observe the data characteristics of the three response variables, and its discrimination becomes obvious with the gradual deepening of convolution. [Conclusion] The method described in this paper is helpful to identify the mooring failure characteristics in advance and ensure the stable operation of offshore floating wind turbines.