Research On Wind Turbine Fault Early Warning Based On Generative Adversarial Network

In order to improve the natural environment and meet the growing human demand for energy,actively developing clean and renewable energy power generation is an effective method.In recent years,wind power generation has gradually become an important part of renewable energy power generation worldwide,and the cumulative grid connected installed capacit y of wind turbines has increased year by year.However,the failure rate of most wind turbines remains high,and the maintenance is difficult and the maintenance cost is high,which has an adverse impact on the economic benefits of wind farms.Therefore,i t is necessary to carry out the research of wind turbine fault early warning.Based on the SCADA data of wind turbine,this paper studies the fault early warning based on GAN.The main research methods are as follows: according to the operation mode of variable speed and constant frequency of wind turbine,the abnormal data in the original SCADA data are eliminated by statistical analysis method and K-means algorithm;Random forest or Pearson correlation coefficient method was used to screen the relevant characteristic parameters;The technical characteristics of traditional GAN are studied,its distance measurement is improved,a model based on WGAN is built,and the threshold calculation method is designed;Aiming at the potential instability of model trai ning,the loss function is improved on the basis of WGAN,and the model based on LSGAN is built to avoid the problems of unstable training results and mode collapse;Finally,the health index method based on sliding window and threshold are used as early w arning strategy to realize fault early warning.Main conclusions of the experiment:(1)By preprocessing SCADA data,the main band of wind speed generator active power curve is obtained,which improves the modeling accuracy of subsequent early warning model;(2)The selected algorithm is used to screen the relevant characteristic parameters,which improves the efficiency and generalization ability of the early warning model;(3)WGAN algorithm gives early warning of gearbox oil pool temperature fault3 days and 2 hours ahead of SCADA system,which achieves the purpose of fault early warning;(4)LSGAN algorithm gives early warning of gearbox oil pool temperature fault3 days and 12 hours earlier than SCADA system,and gives early warning of generator drive end bearing temperature fault 5 days and 1 hour earlier than wgan,and the model is more stable and sensitive;(5)As an early warning strategy,the health index method of sliding window reduces false alarm and improves the reliability of the model.