基于数据扩充及风力发电机组功率曲线分段回归的自适应监测方法

A data expansion based piecewise regression strategy for incrementally monitoring the wind turbine with power curve

准确的功率曲线可以反映风机的运行状况,在风力发电机组的监测中至关重要。然而,一些新安装的风力发电机组由于没有足够的训练数据来拟合准确的功率曲线,从而导致监测结果不佳。本文提出了一种基于数据扩充的分段回归方法实现对风电机组的自适应监测,该方法可分为离线建模阶段和在线监测阶段。在离线建模时,首先,设计了一种新的映射函数,通过将其他数据集的数据映射到目标数据集上,从而实现对目标数据集的扩充,进而使得目标风力发电机组有足够的数据用于模型训练。然后,设计了一种分段建模策略,将扩充后数据的信息精炼到少量样本中,再进行功率曲线拟合,可以以较低的计算复杂度准确拟合出功率曲线。在线监测时,可依据功率曲线模型对功率进行预测,最后将预测结果与实际功率进行比较,从而实现对运行状态的监测。此外,我们提出一种增量学习策略,以利用新的数据实时更新模型提高预测和监测的准确度。实验采用真实风电数据,结果表明,所提监测方法能够在数据不足的情况下能准确地发现异常行为,监测准确率可达92.77%。

An accurate power curve is essential for monitoring the wind turbine because this curve can reflect the operating condition of the equipment.However,some newly installed wind turbines may not have enough training data to fit an accurate power curve and lead to poor monitoring results.In this work,considering the insufficient data,we proposed a data expansion based on piecewise regression strategy to monitor the wind turbine with an incremental monitoring strategy.The proposed method can be divided into the offline modeling stage and the online monitoring stage.During offline modeling,a novel mapping function was first designed to expand the insufficient data by mapping the data of other data sets onto the insufficient target data set.In this way,there will be enough training data for the target wind turbine.Then,a piecewise modeling strategy was designed to condense the information of the expanded data into a small number of samples and then fit the power curve.Based on this strategy,the power curve can be accurately fitted with low computational complexity.During online monitoring,the power was predicted by the power curve,and finally,the operating condition can be monitored by comparing the prediction with the observed power.Meanwhile,an incremental learning strategy was proposed to improve both the prediction and monitoring accuracy by updating the power curve model using the newly arrived data.A real case in the experiment illustrated that the proposed monitoring method can accurately detect abnormal behavior with 92.77%detection accuracy while facing insufficient data.