Blade Icing Detection Algorithm Based On SCAD A Data Of Wind Turbines

Wind energy is an inexhaustible clean energy.In order to achieve the goal of"emission peaking and carbon neutrality" around the middle of this century,in recent years,the construction planning of wind energy utilization are being accelerating in the world.In order to make full use of wind energy,the location of wind farms is mostly concentrated in the regions with the most abundant wind resources,and these regions often have the geographical characteristics of high latitude and high altitude.Therefore,wind turbines often face cold and harsh climatic conditions,which makes the phenomenon of wind turbine icing often occur.Once the blades of a wind turbine are covered with ice,their shape will change and weight will inscrese,and the uneven distribution of blade mass will lead to increased vibration of wind turbine,ultimately affecting the operation safety and greatly affecting the output power of the wind turbine.Therefore,timely and accurate detection of blade icing status is of great significance to ensuring the safety of wind turbines and improve the utilization of wind energy.In this paper,the problem of wind turbine blade icing is taken as the research object,and multiple machine learning ice detection models are constructed using wind turbine SCADA data to fully mine the hidden ice related features.The research work and results of this article are as follows:1.A data preprocessing and optimal value feature extraction algorithm has been proposed.A whole set of pretreatment methods are proposed for fan monitoring data,including eliminating singular values,removing redundant data,standardizing data,etc.,and oversampling and undersampling methods are proposed for extreme imbalance of data;A feature extraction algorithm based on the combination of evaluation function and logistic regression is proposed,which extracts the optimal icing related feature set with better accuracy and efficiency than using full feature sets for icing detection.2.An ice detection algorithm based on support vector machine and particle swarm optimization is proposed.The key and difficulty in building an excellent support vector machine model lies in selecting appropriate kernel functions and related parameters.Using particle swarm optimization algorithm,the optimal model parameters are obtained by iteratively updating the parameter combination through the optimal value proximity random search strategy,constructing an evaluation function,and undergoing multiple rounds of cross validation.Through comparative verification,this algorithm is significantly superior to the logistic regression algorithm in terms of accuracy and precision.However,when faced with big data samples,the efficiency drops sharply,which limits the further improvement of the algorithm’s accuracy.3.To improve the efficiency of big data icing detection and enhance the adaptability of the model,an ice detection algorithm based on deep neural network and transfer learning was proposed.The algorithm first learns and trains a primary icing detection model based on a deep neural network model on the icing data of a single wind turbine,and then tranfer the model to other wind turbines for verification through transfer learning.By optimizing some model parameters,an icing detection model with good generalization performance is obtained.4.In order to further optimize the effect of icing detection,time series dependence is introduced on the basis of deep neural network model,and an improved ice detection algorithm based on long short-term memory network is proposed.This algorithm first creates a cross prediction model to obtain the residual between the predicted value and the actual value,then extracts the feature amount of the residual,and finally creates a sequence classification model based on the residual feature for ice detection.This algorithm has better ice detection accuracy and generalization performance compared to traditional methods.