Research On Icing Fault Prediction Of Wind Turbine Based On Hybrid Spatiotemporal Graph Convolutional Network

Wind power is clean,safe,and pollution-free and is an essential renewable energy source.However,due to the influence of geographical environment factors,areas with high wind energy density are prone to icing failure of fan blades in winter,which destroys the original mechanical balance of the fan and accelerates the aging and damage of components.Therefore,it is of great practical significance to establish a model in a data-driven way to predict icing faults,issue early warnings and take measures before icing faults occur,and improve the operational efficiency of wind farms.However,the current research methods often only focus on the time series features in the data,ignoring the spatial relationship between the data,and it is not easy to fully mine the multi-scale features in the data.As a result,the accuracy and timeliness of the current model for icing fault prediction still need to be improved.This paper takes wind turbine icing fault prediction as the research topic.It decomposes it into two sub-tasks: wind turbine operation data trend prediction and wind turbine icing fault prediction.The main contents and contributions are as follows:1)A wind turbine operation data trend prediction model is proposed based on the hybrid spatiotemporal graph Conv LSTM(Convolutional LSTM Network).First,the original data from the actual production environment is cleaned and processed.The original onedimensional data is combined into two-dimensional space-time map data in the time dimension and space dimension to improve the ability of the data to represent the spatial relationship.Secondly,the Conv LSTM is mixed with the spatiotemporal graph in the model training.The mean square error(MSE)is used as the loss function for fitting to eliminate the influence of noise and improve the model’s accuracy.Finally,in the experiment,a more accurate prediction of the wind turbine operation data trend in the next 15 minutes is achieved,and the performance is better than in other models.2)A wind turbine icing fault prediction model is proposed based on a Multi-Scale Convolutional Neural Network(MSCNN).Firstly,the shortcomings of existing automatic feature construction methods are analyzed,and a multi-scale stacked mechanical feature construction structure is designed to explore wind turbine data in different receptive fields.Secondly,the multi-scale features are fused with convolutional neural networks.The loworder features and high-order features in the wind turbine data are simultaneously mined to explore the profound relationship between wind turbine data and icing faults.Finally,compared with the existing methods,the advantages of this model in the prediction accuracy of icing faults are verified in the comparative experiments.Compared with the existing methods,a large amount of actual wind turbine data is used.The experimental results show that the proposed algorithm can predict the icing failure of the wind turbine more accurately and,which further verifies that the spatiotemporal graph data processing method and multi-scale features can be used in this method’s effectiveness in the task.At the same time,it also provides a new solution for wind turbine icing failure prediction.