Icing Detection Of Wind Turbine Blades Based On Machine Learning

With the proposal of peaking carbon dioxide emissions and carbon neutralization,the demand for clean energy in countries all over the world is increasing,contributing to renewable energy has attracted extensive attention.Wind energy is one of the fastest-growing renewable energy fields in the world because of its characteristics of abundant availability,cost-effectiveness,and technological maturity for commercial use.However,wind farms are usually located at sea or plateau mountainous areas with low temperature and high humidity,resulting in the problem of blade icing.Icing on the blades will not only affect the power generation performance of the wind turbine,but also damage the wind turbine itself,and even bring safety problems near the power plant.Therefore,the detection of wind turbine blade icing is of great significance.Traditional wind turbine blade icing detection usually depends on manual observation or installation of sensors and other equipment.However,these technologies are limited by additional labor costs or device installation and maintenance costs.Fortunately,with the wide installation and use of sensors,a large number of wind turbine monitoring data are generated and collected.In addition,the development of big data analysis and mining technology has built an available platform for industrial data analysis and exploration.This paper models and analyzes the icing data of wind turbine blades based on the deep learning method and federated learning method in the field of machine learning.The main research work and results are as follows:1.Since the monitoring data of the sensors in wind turbine is essentially multivariate time series,and the data features are typically complex due to the changeable and complex internal and external environment of wind turbines.According to the characteristics of the data features,the multilevel discrete wavelet transform is used to extract the time-frequency domain features of the data.Then combining with the powerful time series and depth feature extraction capabilities of the deep learning model,the proposed model can automatically analyze the icing data of wind turbine blades.From the data level and algorithm level,the sensor data imbalance problem is investigated by using the data resampling and rebalancing loss function strategies.Finally,a multi-step accumulation strategy is proposed to enhance the robustness of the model in real-time icing detection.2.The combination of data from multiple wind power plants has the perseverance to enrich data characteristics,so a heterogeneous federated learning model that can protect data privacy is proposed.The model can learn the data of different power plants distributed,expand the data characteristics and avoid the leakage of original data.By setting the server-side model and client-side model heterogeneous,this model can not only mitigate the limitation of client-side computing power but also reduce the waste of computing resources on both server-side and client-side.Different client models are also heterogeneous,which can adapt to the data mining of the situation with different data volumes and data characteristics.The proposed model also introduces a rebalancing loss function to alleviate the imbalance of sensors data.