【摘要】 风电场异常运行数据清洗对风电场功率预测、理论功率计算以及发电性能评估至关重要。针对现有数据清洗方法中基于统计特征的清洗算法无法有效清洗风功率曲线中部的堆积型异常数据，基于图像识别的直接清洗算法会忽略不同风速区间的像素点密度分布差异等问题，文章提出了基于分段图像识别的风电场异常运行数据清洗方法。首先，以切入风速和额定风速为数据集分段点，生成分段二值图像集，构建基于Canny边缘检测的分散型像素点辨识模型；然后，基于边缘内图像集，构建基于数学形态学的堆积型异常像素点辨识模型；最后，以中国西北地区某两个风电场的实际运行数据为算例，与四分位法、基于密度的空间聚类（DBSCAN）法、基于图像的直接清洗算法进行对比分析，验证了所提算法的有效性和适用性。所提方法采用数学形态学分割图像的主要部分和突起部分，从而有效辨识风功率曲线中部的堆积型异常数据；通过设置不同分段图像的清洗参数以改进图像直接清洗算法对局部细节的识别性能，清洗后的正常数据散点分布更为平滑。更多还原

【Abstract】 Wind farm abnormal operation data cleaning is of great importance for wind farm power prediction, theoretical power calculation and power generation performance evaluation. The existing cleaning algorithm based on statistics feature cannot effectively clean the stacked abnormal data in the middle of the wind power curve, and the direct cleaning algorithm based on image recognition technology ignore pixel density distribution in different wind speed range, based on above, a wind farm abnormal operation data cleaning method based on piecewise image recognition is proposed. Firstly, with the input speed and the rated speed as the segmented points of the data set, the segmented binary image set is generated, and a distributed abnormal pixel recognition model based on edge detection method is constructed. Secondly, based on the inner edge image set, a stacked abnormal pixel recognition model based on data morphology is constructed. Finally, taking the actual operation data of two wind farms in northeast China as examples, the proposed algorithm is compared with the quartile method, DBSCAN method and the image-based direct cleaning method to verify the effectiveness of it. The results show that, the proposed method uses mathematical morphology to segment the main part of the image and the protrusions, so as to effectively identify the stacked abnormal data in the middle of the wind power curve, and by setting the cleaning parameters of different segmented images to improve the recognition performance of the image direct cleaning algorithm for local details, the smooth scatter plots distribution of the normal data obtained after cleaning verifies the effectiveness of the proposed method.更多还原