Research On Icing Diagnosis Of Wind Turbine Blades Based On SCADA Data

The global energy crisis and environmental pollution are becoming more and more serious.The development of sustainable and environmentally friendly wind power generation is of great significance.Wind turbines are mostly located in mountainous areas and coastal cities with abundant wind energy resources.Blades are an important medium for the conversion of wind energy to electrical energy.Whether the blades are in a healthy state is related to the operation of the wind turbine.Affected by problems such as cloud distribution and cold weather,the wind turbine blades frequently freeze,which causing a drop problem in power generation and making the grid connection unstable,the ice which falling from the rotating blades will damage the surrounding equipment,causing serious injury to the staff,and if the ice covered blades in a long-term would facing the risk of breaking,which brings huge economic loss to the wind farm.SCADA(Supervisory Control And Data Acquisition)system is widely used in wind farms.Based on the historical data derived from it,this paper analyzes the internal laws of various variables and blade icing during the operation of the wind turbine,first conducts the data preprocessing and analysis job,secondly uses the chi-square test method to screen out the features which has strong relevance to blade icing,and then based on the wind turbine operating characteristics and empirical formula analysis,the original variables are used to construct new features with practical significance.Finally,considering the timing characteristics of the fan blade icing problem,in response to the three issues raised,a hybrid model was constructed around Long Short Term Memory Autoencoders,Bidirectional Long Short Term Memory,and Temporal Convolutional Networks,the research content is as follows:(1)In view of the limitation of shallow model facing a large amount of unbalanced wind turbine operating data,and traditional feature extraction methods ignoring the characteristics of blade icing timing correlation.Long Short Term Memory Autoencoder is using for feature adaptive learning,and a hybrid diagnostic model is constructed in combination with twin support vector machines.The flower pollination algorithm optimizes the parameters that affect the classification performance of the twin support vector machine.A comparative test was conducted on the No.15 wind turbine test set and the performance of the model was evaluated to verify the effectiveness of the model.(2)In order to extract more comprehensive blade icing process information,the Biderectional Long Short Term Memory network is used to construct a blade icing diagnosis model to capture the changes in blade icing.In the training,the twin structure is combined to improve the resolution performance of the model for icing samples,and the depth cost sensitivity is used to reduce the influence of the model by the imbalance,a comparative test was conducted on the No.15 wind turbine test set and the performance of the model was evaluated to verify the effectiveness of the model.(3)In view of the poor generalization ability of the single-fan training model on different fans and the unknown labeling of the fan blades in the actual diagnosis task,the blade icing diagnosis model is constructed based on the domain adaptation idea,and the Temperal Convolutional Network with better long-term memory is used to extract features from the raw data of wind turbines with different numbers,and the domain adaptation method based on adversarial learning is used to improve the generalization and transferability of features.Taking the operating data of No.15 and No.21 wind turbines as the source domain and target domain respectively,a comparative test was carried out on the No.21 wind turbine test set and the model performance was evaluated to verify the validity of the model.