| As a kind of secondary energy in human society,electric energy has become the important basic livelihood of economic development and the people all over the world.Due to the fact that considerable changes in industrial production and consumption,the electrical power system are expected to supply the power continuously at high quality to the consumers.In the power system,any deviation of voltage,current and frequency from the standard rating value is treated as a power quality(PQ)problem that results in malfunctioning of electrical/electronic equipment,and increase the risk of blackout.On the other hand,with the rapid development of renewable energy and the rise of high voltage direct current(HVDC)transmission,power quality problems are spreading from the consumer side to the power generation,transmission and distribution.In order to find out who is responsible for each power quality event,it is very important to monitor and classify the power quality disturbances in smart grid.In this dissertation,the power quality problems in smart grid have been studied deeply.The main research contents are as follows:(1)A multi-label classification method based on multi-task convolutional neural network is proposed for multiple power quality disturbances(MPQDs).Due to the overlapping and cross of the features between MPQDs samples,it is difficult to extract the distinguishing features for 41 types of PQ disturbances including single,double and triple combined disturbances.In this dissertation,the basic PQ disturbances are assigned to three groups corresponding three learning tasks and the correlations among various PQ disturbances are utilized in the joint learning of interrelated tasks.Furthermore,to improve the recognition rate of transient disturbances,a deconvolution network is employed.The experimental results show that the proposed method can greatly improve the accuracy rate for MPQDs compared with the existing classification methods.It provides a new way to obtain the correlation information between PQ disturbances.(2)A classification method for MPQDs based on time domain feature enhancement is proposed.In this dissertation,the frequency variations and inter-harmonic are added in MPQDs classification for the first time,so as to meet the urgent needs in modern power grid.Due to the fact that the characteristics of frequency variations and normal signal,inter-harmonic and harmonic are very similar,the traditional methods in face of these signals are powerless.To solve this problem,the rotational coordinate component information(RCCI)of the PQ signals is introduced,which not only can eliminate the AC background but also obtain more distinctive features.By constructing a fusion network,the feature of the original PQ signals and their RCCI are combined.In additional,the time-domain statistical characteristics of the original PQ signals and their RCCI are calculated separately,which are added with the semantic features extracted by convolution neural network.The experimental results show that the proposed method can effectively identify 54 kinds of MPQDs including frequency variations and inter-harmonics,and the accuracy rate can reach 97.75%.(3)A classification method for variable length PQ disturbances based on periodic difference and spatial pyramid pooling is proposed.Due to the different fault duration,the samples length stored in fault recorder is different for each power event.Existing methods usually uses 10 T or 12 T fixed window to segment a complete sample.However,such processing will make the disturbance signal in a sample loses the global characteristics.After analyzing the characteristics of disturbance process,the periodic difference feature(PDF)is introduced,which makes the disturbance features in samples more concentrated.Then,a multifusion convolution neural network in which the PDF and the original PQ disturbances are used as inputs,is designed to classification for 10 T to 50 T samples.Finally,by applying the spatial pyramid pool realize the classification of variable length PQ disturbances.Compared with only using spatial pyramid pooling,the PDF information greatly improved the size range of input signals and the classification accuracy.The experimental results show that the accuracy rate can reach 95.15% for 10 T to 50 T sample including single and combined disturbances.The method also provides a new idea for the classification of variable length PQ disturbances.(4)A voltage sag classification method based on double convolution neural network is proposed.The existing methods mainly classify voltage sags from the amplitude information of each phase.However,in view of the voltage sag,not only the amplitude of each phase is changed,but also the phase angle between the three-phase.The phase information is also an important feature for sag classification.Based on this view,two symmetrical convolution subnetworks are designed.One of them is used to capture the amplitude feature of each phase voltage,the other is used to obtain the correlation information of three phase voltages.Then,multi-task learning is adopted to classify voltage sags in two different category schemes.The experimental results show that this method can effectively obtain the changes of three-phase voltage in amplitude and phase and effectively improve the accuracy of the existing classification methods. |
