Power Quality Disturbance Recognition Method Based On GAF-MTF And Multiflow Convolutional Neural Network

With the wide application of non-linear power electronic devices and the wide access of new energy sources in the power grid,the disturbance of the power network becomes more and more complex.Therefore,the intelligent diagnosis of power quality disturbance has become the focus of many experts and scholars.The key point of power quality disturbance identification is to extract the features of disturbance signals and classify them accurately.In this thesis,a method for identifying power quality disturbances based on Gramain angular fields and Markov transition field(GAF-MTF)and multistream convolution neural network(CNN)is presented,which is mainly studied from two aspects:data preprocessing of input signals and the selected disturbance identification network structure.Considering the strong ability of deep learning to extract features automatically,deep learning technology is used to classify and identify power quality disturbance signals,which can directly combine feature extraction and disturbance recognition into one step.Compared with traditional classification algorithm,deep learning technology saves time and has a high recognition accuracy.Therefore,this thesis uses convolution neural network to classify and identify power quality disturbance signals.Convolutional neural networks are mainly used for in-depth learning models in the field of image recognition.Therefore,this thesis needs to convert one-dimensional power quality disturbance signal into two-dimensional image,so the classification of one-dimensional power quality disturbance signal becomes a problem of two-dimensional image classification.First,a power quality disturbance identification model based on GAF-MTF and twostream convolution neural network(TSCNN)is presented.GAF and MTF are common methods for converting time series into two-dimensional images,but GAF technology retains most of the static information of the original signal in the encoding process,and MTF technology retains a large amount of dynamic characteristics.Therefore,this thesis uses a compromise method to unify the two encoding images: using TSCNN network to extract features from the two images,and then using full-connection layer linear fusion structure to fuse the features.Finally,the classification results of the disturbances are output through the softmax layer.The network structure of the two channels of TSCNN is the same and the weights are shared,which makes the model have good convergence.Compared with other methods for identifying and classifying disturbance signals,the recognition accuracy of this model is the best.The anti-noise experiments show that the model has good anti-interference ability and robustness.Secondly,considering the loss of time characteristics in the process of converting one-dimensional disturbance signals to images,a power quality disturbance identification model based on time and GAF-MTF THREE-STREAM convolution neural network(TCNN)is presented.The features extracted from the power quality disturbance signal through a one-dimensional convolution neural network(1DCNN)and from the GAFMTF-TSCNN model are fused with the weights of the full connection layer,and then classified and identified.Through comparative experiments,it is proved that the feature fusion method of full join layer weight fusion is the best.The traditional Machine Learning algorithm Extreme Learning Machine(ELM)and Random Forest(RF)are introduced for comparison experiments.The results show that the TCNN model has a significantly higher classification accuracy and good noise resistance than other methods.Finally,the power quality disturbance signal measured in a steel plant is used as the research object for engineering analysis.By comparing the two models presented in this thesis,it is found that the TCNN model has better disturbance recognition ability than the GAF-MTF-TSCNN model but takes longer time.The GAF-MTF-TSCNN model performs better than the TCNN model in identifying some disturbance signals.