Wind Power Curve Data Cleaning Algorithm via Image Thresholding

Abstract

Wind turbine data from the Supervisory Control And Data Acquisition (SCADA) system is very important for wind turbine conditional monitoring, wind power prediction and wind turbine performance evaluation. However, the SCADA data usually contains lots of abnormal data. This paper presents an image-based algorithm for abnormal data cleaning of wind power curve (WPC) data via image thresholding. The basic idea is to build a gray-level representation of the original binary image of WPC which is able to preserve the normal part as much as possible. Therefore, the cleaning operation is then turned into a problem of image segmentation. The proposed algorithm includes the following steps: First, the scatter data is converted into a binary image. Then the median of four distances are computed from each pixel in the image to the nearest connected domain boundary along four directions, and a gray level image is generated to strengthen the normal part, in the meantime, weaken the abnormal part. For all possible threshold t, the optimal t o which makes the smallest Hu moment based dissimilarity of the segmented normal part with a reference WPC template, is finally determined. The proposed algorithm is compared with some data-based algorithms as well as an image-based mathematical morphology operation (MMO) algorithm. Experiments carried out on WPC data of 17 wind turbines from a wind farm verified the effectiveness and accuracy of the proposed method.