Study On Failure Detection Methods Of Wind Turbine Icing Event Based On Deep Learning

Nowadays the demand for energy is increasing day by day.However,wind energy has been paid great attention from the society due to its clean and renewable characteristics,as well as the abundant resources.Developing the wind energy industry is helpful for achieving the adjustment of energy structure and sustainable development.However,most wind turbines are located at high latitudes such as mountains and costal areas.The turbine blade is likely to encounter the icing event in cold and humid environment for long period.Turbine icing reduces the power production and increases the fatigue load,sometimes even damages the device and reduces the service life of turbine,which greatly influences the stability and benefits of wind farms.Thus,to improve the efficiency of power generation,control the maintenance costs and eliminate the safety hazards,in the condition that SCADA system has been the standard supporting facilities for wind turbines,we collect the real operating data from five turbines by SCADA system.Then we studied the icing failure detection research based on deep learning,despite of the difficulties of various sampling scale,noisy data,weak representation of state variables as well as the complex and dynamic working conditions.According to the characteristics of turbine data,a standardized process of data preprocessing and feature mining was designed.The main algorithm related to the recurrent neural network such as LSTM,GRU was implemented to model and analyze the data,then compare the results.With the combination of actual logic mechanism of turbine,voting system for icing risk estimation is implemented,playing a role in the output of deep learning model.Finally the algorithm framework based both on LSTM network and icing voting mechanism was constructed,capable to learn the progressive trend and hidden messages from historical correlation.In addition,the model performance on the same data distribution and different data distribution was considered,so the generalization and mobility of model in practical application was studied.