Ice Detection Model Of Wind Turbine Blades Based On Self Attention Bi-IndRNN

With the rapid development of modern society,the demand of industrial and domestic electricity is increasing dramatically.The rational use of renewable energy for power generation will help adjust the energy structure and ensure sustainable economic development.As an efficient and clean energy source,wind energy has received wide attention.Installed capacity and new installed capacity of wind power have also increased recent years in China.The major wind farms in China are mainly located in areas with higher latitudes.However,in these areas,winter is extremely cold and last for a long time.Low temperatures with sufficient wind energy can easily freeze wind turbine blades.As an important device for converting wind energy into electrical energy,the iced wind turbine blades will cause serious problems such as reduce the efficiency of wind energy utilization,change of blade structure and shortened service life.Therefore,how to accurately detect the icing conditions of the blades through the wind turbine working conditions and environmental data in time is of great significance to increase the utilization rate of wind energy and equipment security.This thesis studies the problem of ice detection on wind turbine blades.In view of the disadvantages of traditional statistical learning models and time series models,this thesis combines data-driven modeling of independent recurrent neural networks based on self-attention networks and icing mechanism modeling,using hybrid modeling method to build a model for detecting icing on wind turbine blades,the feasibility and effectiveness of the model proposed in this thesis is verified through the data set of wind turbine working conditions collected on the industrial spots.The main contents of this thesis are summarized below:1)Due to the problems of sample imbalance in the experimental data set,feature redundancy,and data skew,this data was down-sampled and cost-sensitive learning was used to reduce the impact of sample imbalance.Using Pearson correlation and tree model for feature selection,based on Makkonen object surface icing mechanism model and Rahimi wind energy comprehensive utilization mechanism model construction priori knowledge features.2)In view of the experimental data is time-series industrial data with a long time-series steps,this thesis first introduces the traditional neural network RNN and LSTM,because of the wind turbine blades ice detection problem is a gradual process,therefore time-series models are more suitable than statistical learning models in this problem.Considering that the traditional time-series deep learning model is likely to gradient disappearance in long time-series sequence tasks,it is impossible to learn the whole icing process at one time.In this thesis,the independent recurrent neural network is adopted.Its unique structure design and the use of residual network and ReLU activation function combined with good initialization and gradient clipping,can avoid the disappearance of the gradient.The effectiveness of the model is also verified in the experimental data set in this thesis.3)Concerning the contribution to the final result is different at each time step.Similarly,during the icing process of the wind turbine blades,the working condition at each moment also has different contributions to the final result.This thesis proposes a self attention network.Because the current time series model cannot effectively distinguish the contribution of each moment,the self-attention network can assign different weights to each time step in the input sequence,so the model can focus on the moment in the input sequence that has a greater impact on the result.4)Considering that the one-way time series model forgets the head information of the input sequence because its length is too long,this thesis proposes a bi-directional independent recurrent neural network model(Bi-IndRNN)to capture the two-way characteristic information of time series data,it avoids losing sequence header information due to model parameter updates in long sequence tasks,which improves the prediction accuracy of the model.Finally,a Bi-IndRNN model based on Self Attention Network is proposed.Compared with the traditional timing model LSTM,the result is improved by 4.7%,which verifies the validity of the model.