Using edge-detector to model wake effects on wind turbines

Abstract

A healthy wind turbine is essential for efficient wind energy generation, and fault monitoring is required to ensure that. To detect faults accurately, while reducing false alarms, we need to identify the temporary power generation loss that is not due to fault. One major phenomenon that is not a fault but will cause power reduction is the wake effect. Wake effect is the blocking of the air flow towards one turbine by another turbine or other structures, when the wind is blowing in certain direction. Identifying the existence of wake can effectively reduce the false alarms of faults. Meanwhile, although wake is not a fault, it increases the turbine's structural loading and may eventually lead to a fault. Therefore, it is beneficial to understand how susceptible a turbine is to wake effect, which is affected by the layout of the wind farm and reflected in the data. Currently the wind farms are typically equipped with the capacity to collect SCADA (supervisory control and data acquisition) data in this standard data format; however, the large amount of SCADA data may not be fully utilized to its potential, and we propose to use the SCADA data to learn the wake pattern for each turbine. Our approach is driven by the data, without any pre-determined modeling, and hence it is automatic, adaptive and widely applicable. Specifically, we propose to represent the wind speed difference data against wind direction, similar to an image, and then we use an edge detector to discern the pattern in the data to capture the wake effect. Due to the nature of the images to be processed, where outlier pixels can be viewed as “salt & pepper” noise, we propose to use a linear prediction based entropy thresholding method for edge detection, to illustrate our concept for wake effect detection. To further improve wake detection accuracy, each extracted edge map was converted into a linear data series so that an envelope can be built based on the pixels forming the “valleys” and the “peaks”. The envelope enables simple and accurate measurements of the width and depth of the “valleys”. Finally, we generate a single wake pattern for each turbine of interest, by fusing the wake patterns caused by all the neighboring turbines, no matter how many neighboring turbines there are. The more the wind directions in the wake pattern, the more susceptible the turbine of interest is to wake effect. The accurate wake pattern generated by our approach is helpful to separate wakes from true faults, and to understand the vulnerability of the turbines.