Wind Turbine Condition Monitoring Based on Variable Importance of Random Forest

Abstract

SCADA data lacks sensory data such as vibration and strain measurement for traditional wind turbine condition monitoring; it is updates in low frequency, one piece of data per 10 minutes in the main, which is also low for failure prediction. Thus it is a tough work to monitoring wind turbines' working condition based on SCADA data. To this end, this paper proposes a wind turbine condition monitoring method based on variable importance of random forest by utilizing the SCADA data. First, to minimize the misjudgment caused by individual outliers, we divide the SCADA time series into segments in unit of time period T. Second, we use decrease accuracy method to calculate the variable importance of random forest, as the feature vector of each segment, which characterizes a turbine's condition. Third, we compare a specific turbine's variable importance with the standard feature of healthy turbines to obtain the proximity of them. Fourth, the monitoring baseline is determined according to 3σ, and the deterioration function is applied to construct the failure probability model. To show the effectiveness, we apply the proposed method to four real cases from wind farms in China.