Research On Power Quality Disturbance Detection And Classification Based On Improved VMD And Deep Learning

With the increasing scale of modern power systems day by day,grid-connection of new energy power generation and the number of using impact and non-linear power sensitive devices has greatly increased,resulting in distortion of current and voltage,and the decline in power quality,resulting in serious economic losses for users.To achieve high-precision and intelligent effective analysis of power quality disturbances,which is helpful to formulate reasonable solutions to power quality disturbances,and to maximize the safe and economic operation of the power grid.This paper is based on previous research and analysis,To carry out analysis and research around power quality disturbance denoising,detection,feature extraction and classification.In the denoising of power quality disturbances,this paper proposes an improved VMD-SVD power quality disturbance denoising method.Through in-depth study of the basic principles of VMD,and introduce the energy convergence factor and PSO algorithm to adaptively determine the optimal value of the VMD parameter decomposition number k and the penalty factor α.The improved VMD-SVD is applied to the denoising processing of power quality disturbances containing complex components and compared with the denoising effects of SVD and EMD-SVD,simulation experiments show that the improved VMD-SVD algorithm has better denoising effects.For the disturbance signal with a single component,SVD and EMD-SVD are used for disturbance denoising processing and the denoising effect is compared,simulation experiments show that EMD-SVD has a better denoising effect.In the detection of power quality disturbances,in order to make up for the shortcomings of traditional disturbance detection algorithms and improve the detection accuracy,anti-interference and fault tolerance of power quality disturbances,the improved VMD algorithm and HT transformation are combined to realize the detection of power quality disturbances.In order to test the effectiveness of the method proposed in this paper,the detection effect of the algorithm proposed in this paper is compared with the results of HHT detection of power quality disturbances.The simulation experiment results show that the improved VMD-HT algorithm has stronger detection accuracy,anti-interference and fault tolerance.In terms of feature extraction and classification of power quality disturbances,this paper uses improved VMD-HT and HHT algorithms to extract disturbance features and make their time-frequency diagrams,and use the constructed time-frequency diagrams as the input data samples of the neural network model.First,use Res Net,a typical convolutional network in deep learning,to realize the recognition of disturbance signals;In order to realize the classification of disturbances more quickly and accurately,the method of multi-scale feature extraction in the Inception structure is introduced,and the traditional convolutional layer in the Inception structure is replaced by depth separable convolution,and an improved Inception model is built to realize the recognition of disturbance signals;at the same time,the distraction mechanism and residual structure are used to improve the Caps Net network model to realize the recognition of disturbance signals,partitioning thought and the characteristics of the attention calculation in the group of the distraction mechanism make the network model run more efficiently and obtain more key detailed feature information.The above simulation experiment results of the improved deep learning model show that the improved Caps Net network model requires fewer experimental samples and smaller image size to achieve classification accuracy of 99.78%,and its recognition effect is better than other network models.And through the comparison and analysis of the power quality disturbance classification effect with traditional machine learning BP neural network,SVM and random forest,it can be seen that the improved deep learning model proposed in this paper has higher recognition accuracy than the traditional machine learning model.