Research On Power Signal Classification Algorithms Based On Deep Learning

With popularization of the concept of low-carbon and environmental protection,the demand from consumer side of the distribution grid on the green power supply with a higher and higher quality is increasing.In the operation of the large and complex industrial power systems,due to severely disturbed voltage or current signals,it is difficult for the monitoring utilities to analyze the power quality signals.In addition,the non-intrusive power load recognition technique can capture the information of the operating conditions of the power equipment in consumer side(such as which kinds of utilities are running and what status they are in)which is benefit for the power demand prediction and estimation,while the current non-intrusive power load information recognition techniques still have room for improvement.Therefore,this paper mainly studies the power quality distribution recognition and the power load classification.The main contents are as follows:(1)The recognition of the voltage dip types.This paper has chosen the voltage dip as one of the study contents due to the wide range of the power quality distribution types.Feature extraction is the critical step in power quality disturbances recognition,while the traditional mathematical manipulations methods combining with the shallow neural networks cannot well extract the features.Therefore,the paper has proposed a hybrid model based on the convolutional neural network(CNN)and random forest(RF)to perform the automatic feature extraction and classification of the three phase voltage dip data.Firstly,the three phase voltage dip data is transformed to the space phasor model(SPM).Secondly,CNN is used for extracting the features of the SPM.Finally,RF is applied for classification.For acceleration of the training of CNN and relief of over-fitting,dropout,exponential decay of learning rate and update of weights by adaptive moment estimation are introduced.Experimental results demonstrate that the proposed method has a better generalization performance and a higher classification accuracy compared to other classification methods,which provides an objective and efficient auxiliary method for voltage dip recognition.(2)The recognition of the power load types.In the light of the problems such as low generalization ability and easy to fall into local optimal solution existing in the current algorithms,this paper proposes an improved denoising deconvolutional autoencoder(IDDAE)which can classify various categories of real-measured power load data through feature extraction and reconstruction.In the mirror symmetric structure of the network,the convolutional module could extract the distinguishing features and the deconvolutional module could reduce the data redundant and keep the region with higher activation values of the data.The preprocessing method reduces the dimension and the spatial complexity of the data.In order to accelerate the convergence and heighten the classification accuracy of the model,we have used unsupervised pretraining and regularization.The results of the experiments demonstrate that the proposed algorithm has a better generalization performance and a higher recognition rate comparing to the other algorithms.