Wind Turbine Blade Icing Prediction Based On LightGBM And Bayesian Optimization Algorithms

With the global energy crisis and the development of new energy technology,wind energy in renewable energy is receiving more and more attention and application.As wind turbines are mostly installed at higher altitudes and in environments with higher humidity,blade icing is prone to occur,which not only brings large economic losses to wind farms,but also threatens the safety of operation and maintenance personnel.Traditional wind turbine blade icing monitoring is difficult to make accurate prediction of the early stage of icing,triggering the alarm when the blade has already appeared largescale icing phenomenon.In this paper,we use a data-driven approach to firstly preprocess the data,then use the integrated sampling algorithm SMOTE+ENN to solve the data imbalance problem,and filter out the most relevant features and icing according to the icing mechanism and RFE feature selection,and then use the LightGBM algorithm for icing prediction,considering the complex problem of the LightGBM algorithm model tuning reference,and use the adversarial verification The most generalizable hyperparameters are identified with the Bayesian optimization algorithm.The main elements of the study include.Translated with www.Deep L.com/Translator(free version)Firstly,based on the wind turbine blade icing mechanism,the wind speed with the highest degree of icing correlation with the net-side active power and the nacelle temperature with the ambient temperature are selected for scatter plot visualization analysis,and the apparently nonicing data are filtered out by strong rules based on the analysis results to improve the generalization ability of the model.The isolated forest algorithm is used to eliminate the interference of abnormal data due to downtime and turbine failure.For the imbalance of wind turbine blade icing data,the integrated sampling algorithm SMOTE+ENN is used for balanced sampling of data,which solves the problem of sample duplication on the basis of oversampling.The problem of too many dimensions of the original features,combined with the visualization analysis under the three-blade related feature variables,take its mean and standard deviation as the new features and remove its original features,after which the RFE features are selected for the new features and remaining features of the three blades to ensure that the model improves the prediction efficiency without losing the prediction accuracy.Considering the problem that the LightGBM algorithm has too many adjustable parameters and the manual adjustment is inefficient,Bayesian optimization is used to automatically find the optimal tuning parameters.Finally,the completed model was evaluated and compared with other commonly used classification models,and it was found that the built algorithm model outperformed other algorithm models in terms of recall,F1-score and AUC values,which are more important for icing prediction.Figure [42] table [7] reference [80]...