Research On Fault Diagnosis Methods Of Wind Turbine Driven By Wind Farm Big Data

With the depletion of traditional fossil energy sources,environmental pollution,and wind power technology becoming more mature,wind power has become an important part of today’s global energy structure.Wind turbines have co mplex operating conditions and high failure rates,traditional fault diagnosis methods are difficult to meet current requirements.Fault diagnosis technology based on wind field data-driven and artificial intelligence methods can deeply dig into the intern al characteristics of unit data and diagnose and analyze unit faults.Deep learning algorithm is an end-to-end learning method.Compared with traditional algorithms,it can automatically construct features based on input non-linear transformation layer by layer without the need for artificial feature engineering.Based on big data of wind farms,artificial neural networks and deep learning theory,this paper has conducted in-depth research on fault diagnosis methods for wind turbines.The main contents are:Based on the analysis of wind turbine operation principle and SCADA system in wind field,a fault diagnosis method of wind turbine based on deep learning theory is proposed,and the icing of wind turbine blades is taken as an example to carry out relevant experimental investigations.The data of a SCADA system in a wind farm is collected and filtered for data pre-processing,that is,after completing data cleaning and normalization,analyze data correlation and remove redundant dimensions;A combination of Border-SMOTE and under-sampling samples to deal with the imbalance of data samples is used.Based on principles of neural network and deep learning optimization algorithm,a diagnosis model of fan blade icing fault based on deep fully connected neural network(FC-DNN)is proposed.The model was verified using the 10-fold crossover method,and influence of model structure parameters on accuracy of the model was discussed.The model was compared with a neural network model that did not use a deep learning-related optimization algorithm and traditional machine learning algorithms.FC-DNN model’s F1 score and accuracy are 0.938 and 0.921,respectively,which are significantly better than the neural network models that do not use deep learning-related optimization algorithm and traditional machine learning algorithms.Our proposed model can quickly and effectively diagnose the icing problem of wind turbine blades.In order to further improve experimental results and reduce model parameters,a one-dimensional convolutional neural network(1-D CNN)based fan blade icing fault diagnosis model was proposed.After sample construction and data enhancement,a simulation experiment was performed.In the experiment,particle swarm optimization(PSO)was used to optimize hyperparameters of the model,influence of data overlap on the model was discussed,and the model was compared with fault diagnosis methods based on other deep neural network and traditional machine learning algorithms.The F1 score and accuracy of 1-D CNN model are 0.962 and0.972,respectively.Experimental results show that the method of increasing data overlap can enhance data set and improves model training stability and final model result;Our proposed 1-D CNN model performs significantly better than ot her fault diagnosis models on fan blade icing,which can further improve fault diagnosis results of FC-DNN model previously proposed.