提出一种基于网格搜索优化（GS）极端随机森林（ERF）模型的风电机组性能预测及异常状态预警方法。首先，采用离散度分析法清洗噪声和异常工况数据，以获取建模用正常运行状态数据。其次，通过分析风机运行与控制原理，选取与转速和功率具有较高相关度的特征参数作为模型输入，完成预测模型训练和验证，并对比ERF模型与其它几种模型的建模效果。最后，基于滑动窗口算法确定窗宽为10 min，增量为1 min，计算窗内数据的平均绝对误差作为状态指标，采用非参数估计确定发电机转速阈值为33.78，有功功率阈值为55.07。借助某风电机组真实历史运行数据和故障样本，验证异常状态预警方法，结果表明，该方法能够对即将发生的故障进行检测，预警时间比实际故障时间有效提前。

A wind turbine performance prediction and abnormal state early warning method is proposed based on the combination of grid search(GS) and extreme random forest(ERF) model. Firstly, the dispersion analysis method is applied to clear the noise and abnormal working condition data, so as to acquire the normal operating data for modeling. Secondly, by analyzing the operating and control principles of the wind turbine, the characteristic parameters with high correlation degree to fan speed and output power are selected as the model inputs, the prediction model is trained and tested with above data, and the modeling effect of ERF model is compared with those of several other models. Finally, it is determined that the window width is 10 min and the increment is 1 min based on the sliding window algorithm, the mean absolute errors of the parameters inside the window are calculated as the state indicators, and the nonparametric estimation method is adopted to determine the generator speed threshold of 33.78 and the active power threshold of 55.07.The abnormal state early warning method is verified by using the actual historical operating data and fault samples of a certain wind turbine.The research results show that this method can detecte the impending failure, which warning time is earlier than the actual failure time effectively.

TM614