應用自適應增強演算法於風機控制系統異常之預警

EARLY ANOMALY WARNING OF THE WIND TURBINE CONTROL SYSTEM BASED ON ADABOOST ALGORITHM

本研究主要針對台電某風場15部風機的監控和數據採集系統SCADA（supervisory control and data acquisition）資料進行分析，首先從15部風機中找出當月份的模範風機，再以自適應增強演算法（adaptive boosting, Adaboost）學習並回歸當月份模範風機之發電機轉速對風速之資料，計算其他幾部風機的資料的實際值和預測值的殘差，藉由指數加權移動平均管制圖（exponentially weighted moving average control chart, EWMA control chart）得到超出管制圖殘差資料點的比例，對比例超過一倍標準差閾值之風機進行第二階段分析。在第二階段的分析中，把風速按Vestas風機控制邏輯分類為4個類別，以自適應增強分類學習模範風機之發電機轉速與葉片旋角，並利用混淆矩陣比較實際值與預測值的落差，進而得知該部風機出現問題的風速區間，找出其控制系統的異常，作為風機控制系統異常之預警。

The 15 wind turbines' SCADA data from a Taipower wind farm were analyzed in this study. First, an exemplary wind turbine was chosen by the least square error with the guaranteed power curve (GPC) among the 15 wind turbines. The generator speed versus wind speed data of the exemplary wind turbine was learned based on the Adaboost algorithm to predict the generator speed of the other wind turbines. The residues between the generator speed of the exemplary turbine and the other turbines were calculated. By figuring out the percent of the residue that is outside the exponentially weighted moving average control chart and then using one standard deviation of the proportional data as the threshold, the wind turbine's proportion of residue beyond the threshold would be further analyzed. In the second stage, the wind speed was classified into 4 categories according to the control logic. The relation-ship between the generator speed and the blade pitch angle of the exemplary wind turbine was learned to predict the wind speed category. By comparing the discrepancy between the actual categories and that of the predicted categories with the confusion matrix, the fault of the wind speed category could be found, and the control system anomaly could be identified. Therefore, it could be used as the early anomaly warning of the wind turbine control system.