In winter,wind turbines are susceptible to blade icing,which results in a series of energy losses and safe operation problems.Therefore,blade icing detection has become a top priority.Conventional methods primarily re...In winter,wind turbines are susceptible to blade icing,which results in a series of energy losses and safe operation problems.Therefore,blade icing detection has become a top priority.Conventional methods primarily rely on sensor monitoring,which is expensive and has limited applications.Data-driven blade icing detection methods have become feasible with the development of artificial intelligence.However,the data-driven method is plagued by limited training samples and icing samples;therefore,this paper proposes an icing warning strategy based on the combination of feature selection(FS),eXtreme Gradient Boosting(XGBoost)algorithm,and exponentially weighted moving average(EWMA)analysis.In the training phase,FS is performed using correlation analysis to eliminate redundant features,and the XGBoost algorithm is applied to learn the hidden effective information in supervisory control and data acquisition analysis(SCADA)data to build a normal behavior model.In the online monitoring phase,an EWMA analysis is introduced to monitor the abnormal changes in features.A blade icing warning is issued when themonitored features continuously exceed the control limit,and the ambient temperature is below 0℃.This study uses data fromthree icing-affected wind turbines and one normally operating wind turbine for validation.The experimental results reveal that the strategy can promptly predict the icing trend among wind turbines and stably monitor the normally operating wind turbines.展开更多
基金This research was funded by the Basic Research Funds for Universities in Inner Mongolia Autonomous Region(No.JY20220272)the Scientific Research Program of Higher Education in InnerMongolia Autonomous Region(No.NJZZ23080)+3 种基金the Natural Science Foundation of InnerMongolia(No.2023LHMS05054)the NationalNatural Science Foundation of China(No.52176212)We are also very grateful to the Program for Innovative Research Team in Universities of InnerMongolia Autonomous Region(No.NMGIRT2213)The Central Guidance for Local Scientific and Technological Development Funding Projects(No.2022ZY0113).
文摘In winter,wind turbines are susceptible to blade icing,which results in a series of energy losses and safe operation problems.Therefore,blade icing detection has become a top priority.Conventional methods primarily rely on sensor monitoring,which is expensive and has limited applications.Data-driven blade icing detection methods have become feasible with the development of artificial intelligence.However,the data-driven method is plagued by limited training samples and icing samples;therefore,this paper proposes an icing warning strategy based on the combination of feature selection(FS),eXtreme Gradient Boosting(XGBoost)algorithm,and exponentially weighted moving average(EWMA)analysis.In the training phase,FS is performed using correlation analysis to eliminate redundant features,and the XGBoost algorithm is applied to learn the hidden effective information in supervisory control and data acquisition analysis(SCADA)data to build a normal behavior model.In the online monitoring phase,an EWMA analysis is introduced to monitor the abnormal changes in features.A blade icing warning is issued when themonitored features continuously exceed the control limit,and the ambient temperature is below 0℃.This study uses data fromthree icing-affected wind turbines and one normally operating wind turbine for validation.The experimental results reveal that the strategy can promptly predict the icing trend among wind turbines and stably monitor the normally operating wind turbines.
