Study On Recognition Of Complex Power Quality Disturbances Based On Deep Learning

Compared with traditional energy,renewable energy can reduce carbon emissions,and the amount of resources is much larger than conventional energy,and the proportion of applications in the power system is increasing.At the same time power quality disturbances(power quality disturbances,PQDs)have increased,and today,as users have increasingly higher requirements for grid stability and power quality,the classification of PQDs has important research significance.This paper uses the recognition methods of images and time series in deep learning,the mathematical models of various power quality disturbances,deep learning models such as convolutional neural networks and gated loop units,multi-task learning modeling methods,and multi-task learning.Uncertainty has been studied,and the main contents are:(1)PQDs are divided into four types of single disturbances and complex disturbances composed of four types of simple disturbances,and different simple and complex disturbances data sets are generated on MATLAB according to their mathematical models.Introduced the basic principles of deep learning and the advantages of deep learning applied to power quality disturbance classification.Integrating the ideas of multi-label classification and multiclassification,it is proposed to apply multi-task learning classification method to disturbance classification,and the corresponding coding method and loss function are proposed.After experimental verification,the accuracy of multi-task learning classification is higher.(2)Introduced the basic principle and main structure of convolutional neural network,applied it to PQDs classification,and proposed a power quality compound disturbance classification method based on one-dimensional convolutional neural network and multi-task learning.Compared with the traditional classification method reduces the feature extraction stage,and directly uses the original signal to be recognized as the input of the network and outputs the classification result,which realizes end-to-end recognition.The results of simulation experiments and measured experiments show that this method has higher classification accuracy than traditional methods and other known deep learning methods under different signal-to-noise ratios.(3)In order to solve the problem that one-dimensional convolutional neural network does not fully consider the dependency of time series before and after,combined with onedimensional convolutional neural network,a deep learning network structure of convolutional neural network combined with gated recurrent unit is proposed.Simulation results and actual test results show that compared with a single convolutional neural network structure,this structure can effectively extract the time-dependent features of the signal,and has a higher accuracy rate,especially in improving the classification accuracy of transient disturbances.The rate has increased significantly.(4)The loss function of multi-task learning is studied.In order to solve the problem that the weight of the loss function of each sub-task needs to be manually set,the uncertainty weight of the multi-task learning loss function is proposed,so that different sub-tasks have different dynamic weights.,As a parameter of network learning,adjust the size with the deepening of network training.This optimization method is applied to the deep learning network structure of identifying PQDs,and a higher classification accuracy is obtained under the random signal-tonoise ratio,while the anti-noise ability of the system is improved,showing a good application prospect.