Image-Based Abnormal Data Detection and Cleaning Algorithm via Wind Power Curve

Huan Long,Linwei Sang,Zaijun Wu,Wei Gu

DOI: https://doi.org/10.1109/TSTE.2019.2914089

2023-07-17

Abstract:This paper proposes an image-based algorithm for detecting and cleaning the wind turbine abnormal data based on wind power curve (WPC) images. The abnormal data are categorized into three types, negative points, scattered points, and stacked points. The proposed algorithm includes three steps, data pre-cleaning, normal data extraction, and data marking. The negative abnormal points, whose wind speed is greater than cut-in speed and power is below zero, are first filtered in the data pre-cleaning step. The scatter figure of the rest wind power data forms the WPC image and corresponding binary image. In the normal data extraction step, the principle part of the WPC binary image, representing the normal data, is extracted by the mathematical morphology operation (MMO). The optimal parameter setting of MMO is determined by minimizing the dissimilarity between the extracted principle part and the reference WPC image based on Hu moments. In the data mark step, the pixel points of scattered and stacked abnormal data are successively identified. The mapping relationship between the wind power points and image pixel points is built to mark the wind turbine normal and abnormal data. The proposed image-based algorithm is compared with k-means, local outlier factor, combined algorithm based on change point grouping algorithm and quartile algorithm (CA). Numerous experiments based on 33 wind turbines from two wind farms are conducted to validate the effectiveness, efficiency, and universality of the proposed method.

Systems and Control

What problem does this paper attempt to address?

This paper aims to propose an image-based method to detect and clean abnormal data in wind turbines. Specifically, the study presents an algorithm that identifies and cleans three types of abnormal data—negative value points, scattered points, and stacked points—through Wind Power Curve (WPC) images.

The algorithm includes three steps:
1. **Data Pre-cleaning**: First, filter out data points where the wind speed is greater than the cut-in wind speed but the power is less than zero (i.e., negative abnormal points).
2. **Normal Data Extraction**: Form WPC images and their corresponding binary images from the remaining wind power data; use Mathematical Morphology Operation (MMO) to extract the main part representing normal data from the WPC binary image; determine the optimal parameter settings for MMO by minimizing the dissimilarity (based on Hu moments) between the extracted main part and the reference WPC image.
3. **Data Labeling**: Sequentially identify the pixel points of scattered and stacked abnormal data; establish a mapping relationship between wind power points and image pixel points to label normal and abnormal data of the wind turbine.

The paper validates the effectiveness, efficiency, and generality of the proposed method through comparative experiments. The experiments are based on data from 33 wind turbines from two wind farms. Compared with methods such as K-means, Local Outlier Factor (LOF), and a Combined Algorithm (CA) based on change point grouping and quartile algorithms, the proposed algorithm shows better performance in handling a large amount of abnormal data and can intuitively provide classification information of abnormal data.