Research On The Prediction Method Of Wind Turbine Blade Icing Based On Mechanism Analysis And Model Fusion

With the growth of renewable energy,the normal operation of wind turbines provides an essential basis for ensuring that wind power reaches its capacity.As a serious impact on the work of wind turbines failure,blade icing will lead to poor wind power conversation efficiency.At the same time,blade covered with ice will be vulnerable to injury people and result in other extreme conditions.Therefore,accurate and timely detection of blade icing is of great value based on readily available data such as wind turbine operating and environmental parameters.It is the foundation for improving wind power conversion efficiency,extending equipment life and ensuring personal safety.This paper focuses on the methodology for wind turbine blade icing prediction.The main research elements are shown as follows.(1)Perform feature reconstruction and engineering by fusion mechanism analysis.The original dataset suffers from errors,feature redundancy and high feature dimensionality.Firstly,the dataset is analysed to gain an in-depth understanding for the meaning of the fields and data types.The original data set was then pre-processed,including filling in missing values using Lagrangian interpolation,using the Lajda criterion combined with box plotting to identify outliers,and using appropriate normalisation methods based on the characteristics of the model.Finally,the model features were selected and reconstructed,including: mechanistic analysis based on the icing mechanism model proposed by Makkonen and the wind energy utilisation mechanism model proposed by Rahimi,analysis of the correlation between the icing feature fields with Pearson’s correlation coefficient,and assessment of the feature importance with the random forest algorithm.Knowledge of icing mechanisms is introduced to complete feature reconstruction and feature engineering.It is the data basis for subsequent construction of a blade icing prediction model.(2)Base on the EE-ADASYN algorithm to optimize the imbalanced class distribution of icing datasets.Considering that the extremely unbalanced distribution of data between majority and minority class samples in the dataset leads to the problem of biased prediction in the constructed prediction models,the data set used for model prediction was further improved by simultaneous optimisation at both the data level and the algorithm level.At the data level,a down-sampling method based on strong rule-based filtering and similarity metrics is proposed to reduce the majority sample data through exploratory analysis and visualization of the data.Then an EE-ADASYN algorithm is introduced at the algorithm level.It is an improved oversampling algorithm which based on the idea of Easy Ensemble(EE).The base classifier is Random Forest.The effectiveness of this data set distribution optimization method is verified.It provides the data and algorithm support for the subsequent construction of icing prediction models.(3)Construct a model fusion-based prediction model for wind turbine blade icing.Considering the limitations of the prediction performance of a single model,a model fusion model for the prediction of blade icing is constructed.First,the overall process of model construction is introduced,and the scoring index Score is proposed for model evaluation.Subsequently,the logistic regression(LR)and Xg Boost models based on statistical learning methods were optimized in terms of parameters to obtain their single best model.A deep learning temporal model based on CNN-foc LSTM is proposed,whose main network structures are Convolutional Neural Network(CNN)and Long Short Term Memory(LSTM).At the same time,the Focal Loss function in image processing is migrated as a loss function optimization method for model training.Finally,Stacking was applied for model fusion.The tuned and optimised EE-ADASYN(base classifier is random forest),LR model and Xg Boost model are adopted as three different individual learners.CNN-foc LSTM model is chosen as the meta-learner to obtain the fusion prediction model of wind turbine blade icing.Validation on a Supervisory Control and Data Acquisition(SCADA)dataset demonstrated the validity of the constructed prediction model.