CN113780160A - Electric energy quality disturbance signal classification method and system - Google Patents

Electric energy quality disturbance signal classification method and system Download PDF

Info

Publication number
CN113780160A
CN113780160A CN202111051091.8A CN202111051091A CN113780160A CN 113780160 A CN113780160 A CN 113780160A CN 202111051091 A CN202111051091 A CN 202111051091A CN 113780160 A CN113780160 A CN 113780160A
Authority
CN
China
Prior art keywords
model
power quality
disturbance signal
training
quality disturbance
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111051091.8A
Other languages
Chinese (zh)
Other versions
CN113780160B (en
Inventor
王勇
栾乐
许中
莫文雄
彭和平
孔令明
崔屹平
刘俊翔
罗思敏
周凯
徐硕
范旭娟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guangzhou Power Supply Bureau of Guangdong Power Grid Co Ltd
Original Assignee
Guangzhou Power Supply Bureau of Guangdong Power Grid Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guangzhou Power Supply Bureau of Guangdong Power Grid Co Ltd filed Critical Guangzhou Power Supply Bureau of Guangdong Power Grid Co Ltd
Priority to CN202111051091.8A priority Critical patent/CN113780160B/en
Publication of CN113780160A publication Critical patent/CN113780160A/en
Application granted granted Critical
Publication of CN113780160B publication Critical patent/CN113780160B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Signal Processing (AREA)
  • Monitoring And Testing Of Transmission In General (AREA)

Abstract

The invention discloses a method and a system for classifying power quality disturbance signals, which are used for expanding the dimension of one-dimensional disturbance signals, constructing a two-dimensional data sample, reflecting more detailed characteristics of the disturbance signals and improving the noise immunity, respectively training a VGGNet model, a GoogLeNet model and a ResNet model by using the one-dimensional data and the two-dimensional data, extracting the characteristics of the disturbance signals from different dimensions, ensuring the accuracy of the extraction of the characteristics of the disturbance signals, and finally fusing the characteristics of a plurality of disturbance signals by using a characteristic fusion module, so that the robustness of the classification of the power quality disturbance signals is improved, and the technical problems that the accuracy and the interference immunity cannot be considered in the conventional power quality classification method and the reliability is lower are solved.

Description

Electric energy quality disturbance signal classification method and system
Technical Field
The invention relates to the technical field of power quality analysis, in particular to a method and a system for classifying power quality disturbance signals.
Background
The grid connection of distributed energy and the access of a large number of novel power electronic devices and other nonlinear loads in a power system lead to the increasingly serious problem of the quality of electric energy in a power grid, and threaten the operation safety of the power system. The disturbance of the quality of electric energy in an electric power system usually shows as sudden changes of the amplitude, frequency and phase of voltage and current, and according to the type of the changes and the difference of the change amplitude, the disturbance can be generally divided into several different types of voltage sag, voltage interruption, voltage fluctuation and flicker, voltage deviation, frequency deviation, voltage three-phase unbalance, harmonic wave and the like, and the disturbance can cause a series of influences on the electric power system, such as increase of line loss, increase of energy consumption, damage of sensitive equipment, interference of normal communication, even abnormal operation and damage of main electric facilities, serious harm to the operation safety of a power grid, large-scale power failure events, and huge economic loss and consequences for industrial production and daily life. Effective safety assessment and reasonable power resource allocation are effective measures for dealing with the damage caused by the power quality problem, so that detection and accurate classification of various disturbances are of great significance for analyzing and eliminating the power quality problem and making corresponding countermeasures.
The existing power quality classification method mainly comprises two parts of disturbance feature extraction and disturbance identification classification, wherein the disturbance feature extraction is mainly used for transforming a power quality disturbance event sequence or reconstructing the waveform of the power quality disturbance event sequence, and then extracting the feature quantity of a power quality disturbance signal, and the commonly used disturbance feature extraction methods comprise Fourier transform, S transform, wavelet transform, Hilbert-Huang transform and the like. The disturbance signal identification and classification is to send the detected disturbance signals with respective characteristics to a corresponding classifier for identification and classification, and the commonly used disturbance signal classification methods include a support vector machine method, a decision tree method, an artificial neural network method and the like. However, the existing power quality classification method cannot give consideration to both accuracy and anti-interference performance, cannot guarantee classification accuracy and simultaneously has good anti-noise performance, and is low in reliability.
Disclosure of Invention
The invention provides a method and a system for classifying power quality disturbance signals, which are used for solving the technical problems that the existing power quality classification method cannot give consideration to both accuracy and anti-interference performance and is low in reliability.
In view of the above, a first aspect of the present invention provides a method for classifying a power quality disturbance signal, including:
constructing a power quality disturbance signal sample data matrix according to historical power quality disturbance signal data at nodes of the power system;
constructing a two-dimensional data set with time-frequency characteristics according to the power quality disturbance signal sample data matrix;
constructing a training set and a test set according to the power quality disturbance signal sample data matrix and the two-dimensional data set, wherein the training sets divided by the power quality disturbance signal sample data and the two-dimensional data set respectively comprise a first training set and a second training set;
constructing a first disturbance signal classification network model, wherein the disturbance signal classification network model comprises a VGGNet model, a GoogLeNet model and a ResNet model;
training a VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, and respectively training a GoogLeNet model and a ResNet model by using a second training set corresponding to a two-dimensional data set to obtain the well-trained VGGNet model, GoogLeNet model and ResNet model;
removing the fully connected layers and the Softmax function of the trained VGGNet model, the GoogLeNet model and the ResNet model, and accessing the pooling layer at the tail ends of the trained VGGNet model, the GoogLeNet model and the ResNet model into the feature fusion module and the fully connected layer classification module to obtain a second disturbance signal classification network model;
training a second disturbance signal classification network model by using a second training set corresponding to the power quality disturbance signal sample data and a second training set corresponding to the two-dimensional data set to obtain a trained second disturbance signal classification network model;
and preprocessing the power quality disturbance signals to be classified into two-dimensional data with time-frequency characteristics, and inputting the power quality disturbance signals to be classified and the preprocessed two-dimensional data with the time-frequency characteristics into a trained second disturbance signal classification network model to obtain a classification result of the power quality disturbance signals.
Optionally, the VGGNet model is trained by using a first training set corresponding to the power quality disturbance signal sample data, and the network parameters of the model are updated by using a momentum fraction step reduction optimizer when the google lenet model and the ResNet model are respectively trained by using a second training set corresponding to the two-dimensional data set.
Optionally, the ratio of the first training set, the second training set and the test set is 3:3:1 or 7:7: 3.
Optionally, the VGGNet model in the disturbance signal classification network model is an improved VGGNet model, and the network structure of the improved VGGNet model is as follows:
the VGG-16 network of the VGGNet is used as a main framework, the length of a first convolution kernel is 32, the last three sequentially connected continuous convolution layers of the VGG-16 network are replaced by a composite convolution layer consisting of two parallel convolution layers, the sizes of the convolution kernels of the composite convolution layer are 3 x 1, the last three fully-connected layers of the VGG-16 network are reduced to two fully-connected layers, the output of each convolution layer is connected with a BN layer, and a Relu layer is added behind each BN layer for activation.
Optionally, the google lenet model in the perturbation signal classification network model is an improved google lenet model, and a network structure of the improved google lenet model is as follows:
adding a 7 × 7 convolutional layer to the inclusion module on the basis of the GoogLeNet model, adding a 1 × 1 convolutional layer for down-sampling corresponding to the 7 × 7 convolutional layer, removing the last fully-connected layer of the GoogLeNet model, and replacing the last fully-connected layer with a global pooling layer.
Optionally, a ResNet model in the disturbance signal classification network model is an improved ResNet model, and a network structure of the improved ResNet model is as follows:
and taking the ResNet-50 network as a main framework, adding a BN layer in front of each convolution layer of each residual error unit in the ResNet-50 network, adding a Relu layer behind each BN layer for activation, and replacing the last average pooling layer of the ResNet-50 network with a pyramid pooling layer.
Optionally, constructing a two-dimensional data set with time-frequency characteristics according to the sample data matrix of the power quality disturbance signal, including:
constructing a multi-resolution wavelet decomposition model;
determining the number of frequency band divisions and wavelet basis functions;
sample data in the power quality disturbance signal sample data matrix is serialized to obtain signal waveform data sets corresponding to different types of disturbance signals;
and performing multi-resolution wavelet decomposition on the samples in the signal waveform data set by using a multi-resolution wavelet decomposition model according to the frequency band division number and the wavelet basis functions, and converting the time-frequency signals obtained by decomposition under different scales into a two-dimensional time-frequency graph to form a two-dimensional data set with time-frequency characteristics.
The second aspect of the present invention provides a power quality disturbance signal classification system, including:
the one-dimensional data sample construction module is used for constructing a power quality disturbance signal sample data matrix according to historical power quality disturbance signal data at nodes of the power system;
the two-dimensional data sample construction module is used for constructing a two-dimensional data set with time-frequency characteristics according to the power quality disturbance signal sample data matrix;
the system comprises a sample dividing module, a data processing module and a data processing module, wherein the sample dividing module is used for constructing a training set and a test set according to a power quality disturbance signal sample data matrix and a two-dimensional data set, and the training set divided by the power quality disturbance signal sample data and the two-dimensional data set respectively comprises a first training set and a second training set;
the first model building module is used for building a first disturbance signal classification network model, and the disturbance signal classification network model comprises a VGGNet model, a GoogLeNet model and a ResNet model;
the first training module is used for training the VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, and respectively training a GoogLeNet model and a ResNet model by using a second training set corresponding to the two-dimensional data set to obtain the well-trained VGGNet model, GoogLeNet model and ResNet model;
the second model building module is used for removing the full connection layers and the Softmax function of the trained VGGNet model, GoogLeNet model and ResNet model, and accessing the feature fusion module and the full connection layer classification module to the end pooling layers of the trained VGGNet model, GoogLeNet model and ResNet model to obtain a second disturbance signal classification network model;
the second training module is used for training a second disturbance signal classification network model by using a second training set corresponding to the power quality disturbance signal sample data and a second training set corresponding to the two-dimensional data set to obtain a trained second disturbance signal classification network model;
and the disturbance signal identification module is used for preprocessing the power quality disturbance signals to be classified into two-dimensional data with time-frequency characteristics, and inputting the power quality disturbance signals to be classified and the preprocessed two-dimensional data with the time-frequency characteristics into a trained second disturbance signal classification network model to obtain the classification results of the power quality disturbance signals.
Optionally, the two-dimensional data sample construction module is specifically configured to:
constructing a multi-resolution wavelet decomposition model;
determining the number of frequency band divisions and wavelet basis functions;
sample data in the power quality disturbance signal sample data matrix is serialized to obtain signal waveform data sets corresponding to different types of disturbance signals;
and performing multi-resolution wavelet decomposition on the samples in the signal waveform data set by using a multi-resolution wavelet decomposition model according to the frequency band division number and the wavelet basis functions, and converting the time-frequency signals obtained by decomposition under different scales into a two-dimensional time-frequency graph to form a two-dimensional data set with time-frequency characteristics.
Optionally, the first training module is specifically configured to:
training the VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, respectively training a GoogleLeNet model and a ResNet model by using a second training set corresponding to the two-dimensional data set to obtain a VGGNet model, a GoogleLeNet model and a ResNet model, training the VGGNet model by using the first training set corresponding to the power quality disturbance signal sample data, and updating network parameters of the models by using a momentum fractional order gradient descent optimizer when the GoogleLeNet model and the ResNet model are respectively trained by using the second training set corresponding to the two-dimensional data set.
According to the technical scheme, the embodiment of the invention has the following advantages:
the invention provides a power quality disturbance signal classification method, which is characterized in that one-dimensional disturbance signals are subjected to dimension expansion to construct a two-dimensional data sample, more detailed characteristics of the disturbance signals can be reflected, the noise immunity is improved, a VGGNet model, a GoogLeNet model and a ResNet model are respectively trained by using the one-dimensional data and the two-dimensional data, the characteristics of the disturbance signals are extracted from different dimensions, the accuracy of the extraction of the characteristics of the disturbance signals is ensured, finally, a plurality of disturbance signal characteristics are fused by using a characteristic fusion module, the robustness of the power quality disturbance signal classification is improved, and the technical problems that the accuracy and the interference immunity cannot be considered in the conventional power quality classification method and the reliability is lower are solved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other related drawings can be obtained according to these drawings without creative efforts.
Fig. 1 is a schematic flow chart of a power quality disturbance signal classification method provided in an embodiment of the present invention;
fig. 2 is another schematic flow chart of a power quality disturbance signal classification method according to an embodiment of the present invention;
FIG. 3 is a schematic structural diagram of a second perturbation signal classification network model provided in the embodiment of the present invention;
FIG. 4 is a schematic diagram of the structure of an improved VGGNet model VGGNet-C provided in embodiments of the invention;
FIG. 5 is a schematic structural diagram of the improved GoogleLeNet model GoogleLeNet-C provided in an embodiment of the present invention;
fig. 6 is a schematic structural view of a modified inclusion module inclusion-C provided in an embodiment of the present invention;
FIG. 7 is a schematic diagram of the improved ResNet model ResNet-C provided in an embodiment of the present invention;
FIG. 8 is a schematic structural diagram of a residual error unit Convk _ x-C of an improved ResNet model ResNet-C provided in an embodiment of the present invention;
fig. 9 is a schematic structural diagram of a power quality disturbance signal classification system provided in an embodiment of the present invention.
Detailed Description
In order to make the technical solutions of the present invention better understood by those skilled in the art, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, are within the scope of the present invention.
For easy understanding, referring to fig. 1, an embodiment of a method for classifying a power quality disturbance signal is provided in the present invention, including:
step 101, constructing a data matrix of electric energy quality disturbance signal samples according to historical electric energy quality disturbance signal data at nodes of an electric power system.
And acquiring a data sample set of various types of electric energy quality disturbance signals at different nodes in the electric power system from historical and real-time monitoring data. Selecting a sampling frequency, taking disturbance signal data of a cycle as a sample, dividing a data set of each disturbance signal into N samples (namely dividing the total number N of the sampling data set of each disturbance signal into a plurality of sections, wherein each section forms one sample, each section comprises a plurality of sampling data points, and the total number of the plurality of samples is N), and preprocessing the samples to form a power quality disturbance signal sample data matrix:
Figure BDA0003252774790000061
wherein x isi1～xijThe method comprises the steps that different types of power quality disturbance signal data of sampling nodes in a power system are respectively obtained, j is the type of a disturbance signal, and n is the total number of sampling data sets of each disturbance signal. At power frequency of 50Hz, the sampling frequency is set as fsThe number of samples N of each type of power quality disturbance signal is 50N/fs。
And 102, constructing a two-dimensional data set with time-frequency characteristics according to the power quality disturbance signal sample data matrix.
In order to exert the advantages of the deep convolutional network model on two-dimensional image processing and obtain the characteristics of more details of the power quality disturbance signal, the original disturbance signal is used in the inventionDecomposition of the dataset constitutes a two-dimensional dataset J with time-frequency characteristicspAs input to a portion of the channels of the multi-channel feature extraction module.
103, constructing a training set and a test set according to the power quality disturbance signal sample data matrix and the two-dimensional data set, wherein the training sets divided by the power quality disturbance signal sample data and the two-dimensional data set respectively comprise a first training set and a second training set.
Taking an original disturbance signal sample data set X and a corresponding class label as the input of a one-dimensional convolution neural network model, and taking a two-dimensional data set J with time-frequency characteristicspAnd corresponding class labels as inputs to a two-dimensional convolutional neural network model, the datasets X and J are scaled by 3:3:1 or 7:7:3pAnd the test set is divided into a first training set, a second training set and a test set and is used for model training and model testing of each convolutional neural network model and the full-connection layer classifier. The category label data of the power quality disturbance signals can be set to be 0,1,2,.. j, each number corresponds to one power quality disturbance signal, j is the category number of the total power quality disturbance signals, and the specific corresponding relation can be set according to actual conditions.
And 104, constructing a first disturbance signal classification network model, wherein the disturbance signal classification network model comprises a VGGNet model, a GoogLeNet model and a ResNet model.
And 105, training a VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, and respectively training a GoogLeNet model and a ResNet model by using a second training set corresponding to the two-dimensional data set to obtain the well-trained VGGNet model, GoogLeNet model and ResNet model.
One-dimensional disturbance data (namely an original disturbance signal sample data set X) in a first training set and label data corresponding to the one-dimensional disturbance data and the label data are sent to a VGGNet model for training, and two-dimensional disturbance data (namely a two-dimensional data set J with time-frequency characteristics) are sent to a VGGNet model for trainingp) And respectively sending the GoogLeNet model and the ResNet model, and training the GoogLeNet model and the ResNet model to obtain the trained VGGNet model, GoogLeNet model and ResNet model. Encoding method of category labelThe formula adopts One-hot coding.
And 106, removing the fully connected layers and the Softmax function of the trained VGGNet model, the GoogLeNet model and the ResNet model, and accessing the feature fusion module and the fully connected layer classification module to the end pooling layers of the trained VGGNet model, the GoogLeNet model and the ResNet model to obtain a second disturbance signal classification network model.
And removing the output parts of the trained VGGNet model, GoogLeNet model and ResNet model, namely the full connection layer and the Softmax function part, and accessing the output of the final pooling layer of the trained VGGNet model, GoogLeNet model and ResNet model into a feature fusion module and a full connection layer classification module.
The feature fusion module is composed of a fusion layer and used for flattening the two-dimensional feature graph into one-dimensional feature vectors, and the plurality of one-dimensional feature vectors are spliced end to form fusion feature vectors which are used as input of the full-connection layer classification module.
The full-connection layer classification module is used for final feature identification and classification and comprises two full-connection layers and an output function, wherein the output function is a Softmax function, the number of output channels is j, and j is the number of categories of the power quality disturbance signals. The target function is a cross entropy loss function.
And 107, training a second disturbance signal classification network model by using a second training set corresponding to the power quality disturbance signal sample data and a second training set corresponding to the two-dimensional data set to obtain a trained second disturbance signal classification network model.
And training the full-connection-layer classification module by using a second training set, thereby completing the training of the whole disturbance signal classification network model and obtaining a second disturbance signal classification network model. And finally, testing the second disturbance signal classification network model by using a test set to verify the feasibility of the network model.
And 108, preprocessing the power quality disturbing signals to be classified into two-dimensional data with time-frequency characteristics, and inputting the power quality disturbing signals to be classified and the preprocessed two-dimensional data with the time-frequency characteristics into a trained second disturbing signal classification network model to obtain a classification result of the power quality disturbing signals.
Respectively processing the power quality disturbance signals to be classified into one-dimensional data and two-dimensional data according to the steps 101 and 102, inputting the two-dimensional data into a second disturbance signal classification network model, and obtaining an output sequence of the second disturbance signal classification network model: p ═ P1,p2,p3……pj]Wherein p is1～pjAnd j is the category number of the total disturbance signals. Find p1～pjThe element with the maximum value in the sequence P is output, the index value of the element in the sequence P is output, the index value is an integer value between (0, j-1), the integer value is compared with the label value corresponding to the disturbance signal which is set before, and the disturbance signal type corresponding to the label value which is equal to the index value is taken as the final identification result.
According to the method for classifying the power quality disturbance signals, provided by the embodiment of the invention, the one-dimensional disturbance signals are subjected to dimension expansion, two-dimensional data samples are constructed, more detailed characteristics of the disturbance signals can be reflected, the noise immunity is improved, the one-dimensional data and the two-dimensional data are utilized to respectively train a VGGNet model, a GoogLeNet model and a ResNet model, the characteristics of the disturbance signals are extracted from different dimensions, the accuracy of the characteristic extraction of the disturbance signals is ensured, finally, the characteristics of the disturbance signals are fused by utilizing a characteristic fusion module, the robustness of the classification of the power quality disturbance signals is improved, and the technical problems that the existing power quality classification method cannot give consideration to both accuracy and interference immunity and is low in reliability are solved.
For easy understanding, referring to fig. 2 and fig. 3, another embodiment of the method for classifying a power quality disturbance signal according to an embodiment of the present invention includes:
step 201, a power quality disturbance signal sample data matrix is constructed according to historical power quality disturbance signal data at nodes of a power system.
In the present invention, step 201 is identical to step 101 in the previous embodiment, and will not be described here.
Step 202, constructing a two-dimensional data set with time-frequency characteristics according to the power quality disturbance signal sample data matrix.
In order to exert the advantages of the deep convolutional network model in two-dimensional image processing and obtain the characteristics of more details of the power quality disturbance signal, the embodiment of the invention adopts a multi-resolution wavelet analysis method to decompose the original disturbance signal data set into a two-dimensional data set with time-frequency characteristics, and the two-dimensional data set is used as the input of partial channels of the multi-channel characteristic extraction module. The method specifically comprises the following steps:
(1) construction of multiresolution wavelet decomposition model
According to the functional space theory, when the sampling frequency of the signal to be analyzed satisfies the sampling theorem, the original signal sequence x (n) can be decomposed into two parts, i.e. a low-frequency part and a high-frequency part, by using an ideal low-pass filter and a high-pass filter respectively. Defining the frequency band occupied by the original sequence x (n) as space V0Then the subspace corresponding to the above-mentioned low frequency part is V1The subspace corresponding to the high-frequency part is W1. Therefore, after one decomposition, the original sequence x (n) corresponds to the frequency band space V0Is decomposed into two subspaces: v1And W1. This decomposition process can be continued, i.e. the low frequency part can be further decomposed into a new low frequency part and a high frequency part, corresponding to the subspace V1Is further decomposed into V2And W2Two subspaces. And decomposing step by step, wherein the corresponding subspace decomposition process is as follows:
Figure BDA0003252774790000091
Figure BDA0003252774790000101
in the formula, WjIs to reflect the low frequency subspace V of the same leveljHigh frequency subspace of signal details, VjThen it is reflected to the upper level of the low frequency subspace Vj-1A subspace of the spatial profile of (a). VjAnd WjReferred to as scale space at decomposition scale j and wavelet space, respectively. Accordingly, the multi-scale decomposition is carried out on the sample data set, and based on the Mallat algorithm, the multi-resolution decomposition formula is as follows:
Figure BDA0003252774790000102
wherein f (t) e L2(R) represents any type of disturbance signal waveform. The first part on the right of the above formula represents the perturbation signal f (t) in the scale space VjThe projection above, represents a smooth approximate approximation of the perturbation signal. The second part of the right hand side of the above equation represents the perturbation signal f (t) in wavelet space WjThe projection reflects the detail difference of the smooth approximate approximation corresponding to two adjacent scale spaces, namely the detail supplement to the disturbance signal. Phi is aj,k(t) and
Figure BDA0003252774790000103
respectively called scale function and wavelet function, respectively scale space VjAnd wavelet space WjThe orthonormal basis of (2) is directly obtained from the filter bank used in the decomposition process. Cj(k) And dj(k) The two decomposition coefficients for multi-resolution wavelet decomposition are called discrete smooth approximation and wavelet transform coefficients, respectively. Cj(k) And dj(k) The recurrence formula of (c) is:
Cj+1(k)＝∑h(m-2k)Cj(m)
dj+1(k)＝∑g(m-2k)Cj(m)
where h (k) and g (k) represent the unit sample response of the low-pass filter and the unit sample response of the high-pass filter used for the decomposition, respectively.
(2) Determining a number of band partitions
According to the multi-resolution decomposition theory, the disturbing signal can be divided into a plurality of parts according to the sampling frequency in a wireless mode, and the corresponding frequency band subspace can also be divided infinitely. For limiting fundamental frequency components to other sub-frequenciesInfluence of band component, let the sampling frequency of the disturbing signal be fs'The fundamental frequency of the disturbance signal is fbSo as to make the fundamental frequency f of the disturbing signalbAnd determining the theoretical frequency band division number as follows based on the principle of the lowest frequency sub-band center:
Figure BDA0003252774790000104
in the formula, Y represents the theoretical number of band divisions, and the actual number of band divisions is obtained by rounding the theoretical number.
(3) Determining wavelet basis functions
And selecting the Daubechies series wavelet as the wavelet class of the selected wavelet base by considering the tight support of the time domain and the frequency domain and the sensitivity to the irregular signal part. And comparing the series of wavelets, comprehensively considering the characteristics of orthogonality, high regularity, generalization and the like, and selecting the db4 wavelet with the shortest time window and excellent time resolution as the base wavelet of the multi-resolution wavelet analysis.
(4) And (4) continuously converting the sample data in the power quality disturbance signal sample data matrix to obtain signal waveform data sets corresponding to different kinds of disturbance signals.
And selecting N sample data of each type of power quality disturbance signal, and serializing the total Nxj sample data to obtain signal waveform data sets corresponding to different types of disturbance signals.
(5) Performing multi-resolution wavelet decomposition on samples in the signal waveform data set by using a multi-resolution wavelet decomposition model according to the frequency band division number and wavelet basis functions, converting time-frequency signals obtained by decomposition under different scales into two-dimensional time-frequency graphs to obtain N x J time-frequency graphs of J disturbance signals, adjusting pictures into sizes of 224 x 224 to form a two-dimensional data set J with time-frequency characteristicsp。
Step 203, constructing a training set and a test set according to the power quality disturbance signal sample data matrix and the two-dimensional data set, wherein the training sets divided by the power quality disturbance signal sample data and the two-dimensional data set respectively comprise a first training set and a second training set.
Step 203 in the embodiment of the present invention is the same as step 103 in the foregoing embodiment, and is not described herein again.
And 204, constructing a first disturbance signal classification network model, wherein the disturbance signal classification network model comprises an improved VGGNet model, an improved GoogLeNet model and an improved ResNet model.
The VGGNet is selected as the deep convolution neural network category of the one-dimensional feature extraction channel, the depth of the network and the accuracy of a prediction model are considered, and the VGG-16 network in the VGGNet is selected as a main framework of the model. And (3) improving the VGG-16 network by considering the length of the sequence data, the model training time and the feature extraction accuracy, and constructing a new VGGNet model, as shown in FIG. 4. To speed up the training speed, the length of the first convolution kernel of the VGGNet model is set to 32. In order to further accelerate the training speed and extract the detailed features of the sequence data, three sequentially connected continuous convolution layers of the VGG-16 model are replaced by three composite convolution layers consisting of two parallel convolution layers, and the sizes of convolution kernels are set to be 3 x 1. To further reduce the number of parameters, the last three fully connected layers in the VGG-16 are reduced to two. The window sizes of the pooling layers are set to be 2 x 1, and a Relu function is selected as an activation function in consideration of convergence rate and gradient diffusion effect. The formula for Relu function is:
Figure BDA0003252774790000111
in the formula (I), the compound is shown in the specification,
Figure RE-GDA0003350075110000112
represents the activation value of the kth neuron of the l layer,
Figure RE-GDA0003350075110000113
is the input value of the kth neuron of the l layer.
In order to accelerate the learning rate and improve the generalization capability of the network, the output of each convolution layer in the network structure is connected with a BN layer, a Relu layer is added behind each BN layer for activation, and the calculation formula of the BN layer is as follows:
Figure BDA0003252774790000121
Figure BDA0003252774790000122
in the formula (I), the compound is shown in the specification,
Figure BDA0003252774790000123
is the output of the current BN layer,
Figure BDA0003252774790000124
wherein, E is the content of the desired compound,
Figure BDA0003252774790000125
is the variance of the received signal and the received signal,
Figure BDA0003252774790000126
as scaling factors for the kth neuron of the l-th layer,
Figure BDA0003252774790000127
and epsilon is a constant term for the bias value of the current BN layer.
The output layer of the model adopts a Softmax activation function, and the calculation formula of the function is as follows:
Figure BDA0003252774790000128
where M is 1, 2.., M, and represents that the final output class of the model is M, pmRepresenting probability of class m, i.e. actual tag value, amRepresenting the neurons to be activated in the output layer.
The target function selects a cross entropy loss function, and the function calculation formula is as follows:
Figure BDA0003252774790000129
in the formula, ymIndicating the desired tag value.
The improved VGGNet model is denoted as VGGNet-C, as shown in FIG. 4.
And selecting GoogLeNet as a deep convolutional neural network main model architecture of a first two-dimensional feature extraction channel. And (3) improving the GoogLeNet model by considering the feature extraction accuracy, the training time and the characteristics of the power quality disturbance signal, and building a new GoogLeNet model as shown in FIG. 5.
Considering the sample number limit of the power quality disturbance signal, and in order to extract more potential features of the disturbance signal, the inclusion module in the original google lenet is improved: and adding a 7 × 7 convolutional layer on the basis of the original inclusion module, and adding a 1 × 1 convolutional layer for down-sampling corresponding to the convolutional layer.
In order to effectively reduce training parameters, prevent overfitting and enhance the corresponding relation between the mapping of the disturbance signal features and the corresponding categories, the last full-connection layer in the original model is removed and replaced by a global pooling layer, namely, global pooling is carried out on each feature map output by the last inclusion module. Considering that a multi-scale-based time-frequency image of a disturbance signal is input by a model, in order to weaken local interference of the time-frequency image and consider information of each scale of the time-frequency image, global average pooling is selected for pooling processing, and the formula of the global average pooling is as follows:
Figure BDA0003252774790000131
in the formula (I), the compound is shown in the specification,
Figure BDA0003252774790000132
values on the l characteristic diagram output by the last increment module are shown, m × n is the size of the characteristic diagram, y(l)And representing the characteristic value obtained after the ith characteristic diagram is subjected to global average pooling.
The two auxiliary classifiers in the original model do not output and process, and are only used for backward propagation of gradient signals and a certain degree of regularization. The activation function selects a Relu function, a BN layer is added in the network structure to improve the training speed and the training precision, the activation function of the output layer selects a Softmax function, and the target function selects a cross entropy loss function. The improved GooglLeNet model is marked as GooglLeNet-C, the structure of the GooglLeNet-C is shown in figure 5, and the structure of the improved inclusion module is shown in figure 6.
And selecting ResNet as the deep convolution neural network model type of a second two-dimensional feature extraction channel. Considering the universality, the number of training parameters and the training precision, selecting ResNet-50 as a main model architecture, improving the ResNet-50, and building a new ResNet model, as shown in FIG. 7.
Considering the convergence stability and convergence speed of the network, and in order to prevent overfitting, a BN layer is added before each convolution layer of each residual unit in ResNet-50, and a Relu layer is added after each BN layer for activation. And (3) considering the characteristics of the disturbance signals extracted from different scales, replacing the last average pooling layer of the ResNet-50 network with a pyramid pooling layer, namely, outputting the characteristic map output from the last Conv5_ x residual unit in the ResNet-50 network after the processing of the three average pooling layers with different window sizes and step sizes. Based on the signature size output from the Conv5_ x unit, the pooling window sizes of the three pooling layers are set to 4, 7, 13, respectively, with corresponding step sizes of 3, 6, 13, respectively.
The activating function is a Relu function, the activating function of the output layer is a Softmax function, and the target function is a cross entropy loss function. The improved ResNet model and residual units are denoted as ResNet-C and Convk _ x-C, where x is 1,2,3,4,5, and each indicates a different residual unit, and ResNet-C is shown in fig. 7 and Convk _ x-C is shown in fig. 8.
And step 205, training the improved VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, and respectively training the improved GoogleLeNet model and the improved ResNet model by using a second training set corresponding to the two-dimensional data set to obtain the trained improved VGGNet model, improved GoogleLeNet model and improved ResNet model.
Sending the one-dimensional disturbance data (namely the original disturbance signal sample data set X) in the first training set and the label data corresponding to the one-dimensional disturbance data and the label data into an improved VGGNet model VGGNet-C for training, and sending the two-dimensional disturbance data (namely a two-dimensional data set J with time-frequency characteristics)p) And respectively sending an improved GoogLeNet model GoogLeNet-C and an improved ResNet model ResNet-C, and training the GoogLeNet-C model and the ResNet-C model to obtain a well-trained VGGNet-C model, GoogLeNet-C model and ResNet-C model. The encoding mode of the category label selects One-hot encoding.
In the embodiment of the invention, the momentum thought is added into the fractional order gradient descent algorithm to form the fractional order gradient descent optimizer based on the momentum thought, and the fractional order gradient descent optimizer is used for updating the parameters of the disturbance signal classification network. The parameter update formula based on this optimizer is:
Figure BDA0003252774790000141
Figure BDA0003252774790000142
in the formula (I), the compound is shown in the specification,
Figure BDA0003252774790000143
and
Figure BDA0003252774790000144
respectively representing the weight values of the kth connection of the ith neuron of the jth layer in the convolutional neural network at the next moment and the current moment,
Figure BDA0003252774790000145
and
Figure BDA0003252774790000146
respectively representing the offset values of the ith neuron of the jth layer in the convolutional neural network at the next moment and the current moment, v (n +1) andu (n +1) is a weight momentum term and a bias momentum term of the next time, and the update calculation formula of the information of the fractional gradient of the past time and the information v (n +1) and u (n +1) of the fractional gradient of the current time is as follows:
Figure BDA0003252774790000147
Figure BDA0003252774790000148
where Loss () is a Loss function, a function of weight w and offset b,
Figure BDA0003252774790000149
and
Figure BDA00032527747900001410
the values of beta are a weight value and a bias value of the previous moment of the network, beta is a momentum term coefficient, beta is more than 0 and less than 1, the value determines the influence degree of the gradient updating direction of the previous moment on the parameter updating direction of the current moment, rho is the learning rate of the network, rho is more than 0 and less than or equal to 1, delta is a tiny positive number, alpha represents the order of the fractional order, and alpha is more than 0 and less than 2.
According to the embodiment of the invention, the neural network is optimized by using the momentum fractional order gradient descent optimizer, so that the accuracy of classification is ensured while the training speed is improved.
And step 206, removing the fully connected layers and the Softmax function of the trained VGGNet model, the GoogLeNet model and the ResNet model, and accessing the feature fusion module and the fully connected layer classification module to the end pooling layers of the trained VGGNet model, the GoogLeNet model and the ResNet model to obtain a second disturbance signal classification network model.
And removing the output parts of the trained VGGNet model, GoogLeNet model and ResNet model, namely the full connection layer and the Softmax function part, and accessing the output of the final pooling layer of the trained VGGNet model, GoogLeNet model and ResNet model into a feature fusion module and a full connection layer classification module.
The feature fusion module is composed of a fusion layer and used for flattening the two-dimensional feature graph into one-dimensional feature vectors, and the plurality of one-dimensional feature vectors are spliced end to form fusion feature vectors which are used as input of the full-connection layer classification module.
The full-connection layer classification module is used for final feature identification and classification and comprises two full-connection layers and an output function, wherein the output function is a Softmax function, the number of output channels is j, and j is the number of categories of the power quality disturbance signals. The target function is a cross entropy loss function.
And step 207, training a second disturbance signal classification network model by using a second training set corresponding to the power quality disturbance signal sample data and a second training set corresponding to the two-dimensional data set to obtain a trained second disturbance signal classification network model.
And 208, preprocessing the power quality disturbance signals to be classified into two-dimensional data with time-frequency characteristics, and inputting the power quality disturbance signals to be classified and the preprocessed two-dimensional data with the time-frequency characteristics into a trained second disturbance signal classification network model to obtain a classification result of the power quality disturbance signals.
Step 207 and step 208 in the embodiment of the present invention are the same as step 107 and step 108 in the foregoing embodiment, and are not described again here.
The embodiment of the invention adopts original data as a one-dimensional data set, utilizes wavelet analysis to construct a disturbance signal two-dimensional data set, adopts a deep convolutional neural network method to construct convolutional neural network models suitable for different dimensions and carries out structural optimization improvement on the convolutional neural network models, adopts the deep convolutional neural network method, utilizes different improved convolutional neural network models to carry out disturbance signal characteristic extraction by training disturbance signal data sets of different dimensions to obtain a plurality of groups of disturbance signal characteristic vectors, constructs a momentum fractional gradient descent optimizer model, utilizes an optimizer to optimize weight values and bias values of various convolutional neural networks, utilizes disturbance signal characteristic vectors obtained by training of different convolutional neural network models to splice the various quantities to obtain fusion characteristic vectors, adopts the convolutional neural network method to construct a full-connection layer classifier, training a full-connection layer classifier by utilizing the fusion characteristic vector to obtain a neural network classifier for identifying and classifying the power quality disturbance signals, and classifying the power quality disturbance signals by utilizing a trained convolutional neural network model suitable for different dimensions, a momentum fraction step reduction optimizer and the full-connection layer classifier. According to the electric energy quality disturbance signal classification method provided by the embodiment of the invention, the detail characteristics of the disturbance signals are extracted from different scales and dimensions, the accuracy is ensured, meanwhile, the disturbance signals are quickly identified, and necessary technical support is provided for fault elimination of a power system, formulation of compensation measures and stable operation.
Compared with the prior art, the method for classifying the power quality disturbance signals provided by the embodiment of the invention has the following advantages:
(1) at present, the classification method which separates the feature extraction and the disturbance classification identification of the disturbance signal for respective operation usually needs to design and construct two parts of contents, the process is complicated, and the time consumption is too long.
(2) The accuracy, the rapidity and the robustness of a single convolutional neural network in classifying the multi-characteristic power quality disturbance signals are often insufficient. According to the embodiment of the invention, through the classification process of splitting the convolutional neural network, the convolutional neural network feature extraction fusion module and the convolutional neural network classifier which are suitable for signals with different dimensionalities are built, and the efficient, accurate and stable classification of the electric energy quality disturbance based on the improved multi-channel convolutional neural network is provided.
(3) The complexity, irregularity, possibility of feature aliasing and limitation of one-dimensional signals of the power quality disturbance signals are not favorable for feature identification and disturbance classification of the convolutional neural network. According to the embodiment of the invention, wavelet analysis is utilized to transform the original disturbance signal, a two-dimensional disturbance signal data set capable of reflecting more characteristics of the power quality disturbance signal is obtained and is used as the input of the convolutional neural network, and the reliability of classification of the convolutional neural network is improved.
(4) The number of layers of the deep convolutional neural network and the complexity of each improved network model are increased, so that the convergence speed of the network is too low.
For easy understanding, referring to fig. 9, an embodiment of a power quality disturbance signal classification system according to the present invention includes:
a one-dimensional data sample construction module 901, configured to construct a power quality disturbance signal sample data matrix according to historical power quality disturbance signal data at a node of the power system;
a two-dimensional data sample construction module 902, configured to construct a two-dimensional data set with time-frequency characteristics according to the power quality disturbance signal sample data matrix;
the sample dividing module 903 is configured to construct a training set and a test set according to the power quality disturbing signal sample data matrix and the two-dimensional data set, where the training sets divided by the power quality disturbing signal sample data and the two-dimensional data set respectively include a first training set and a second training set;
a first model building module 904, configured to build a first disturbance signal classification network model, where the disturbance signal classification network model includes a VGGNet model, a google lenet model, and a ResNet model;
the first training module 905 is configured to train the VGGNet model using a first training set corresponding to the power quality disturbance signal sample data, and train a google lenet model and a ResNet model using a second training set corresponding to the two-dimensional data set, so as to obtain a trained VGGNet model, a google lenet model and a ResNet model;
a second model building module 906, configured to remove the fully connected layers and the Softmax function of the trained VGGNet model, google lenet model, and ResNet model, and access the feature fusion module and the fully connected layer classification module to the end pooling layers of the trained VGGNet model, google lenet model, and ResNet model to obtain a second disturbance signal classification network model;
the second training module 907 is configured to train a second disturbance signal classification network model by using a second training set corresponding to the power quality disturbance signal sample data and a second training set corresponding to the two-dimensional data set, so as to obtain a trained second disturbance signal classification network model;
the disturbance signal identification module 908 is configured to preprocess the power quality disturbance signal to be classified into two-dimensional data with time-frequency characteristics, and input the power quality disturbance signal to be classified and the preprocessed two-dimensional data with time-frequency characteristics into a trained second disturbance signal classification network model to obtain a classification result of the power quality disturbance signal.
The two-dimensional data sample construction module 902 is specifically configured to:
constructing a multi-resolution wavelet decomposition model;
determining the number of frequency band divisions and wavelet basis functions;
sample data in the power quality disturbance signal sample data matrix is serialized to obtain signal waveform data sets corresponding to different types of disturbance signals;
and performing multi-resolution wavelet decomposition on the samples in the signal waveform data set by using a multi-resolution wavelet decomposition model according to the frequency band division number and the wavelet basis functions, and converting the time-frequency signals obtained by decomposition under different scales into a two-dimensional time-frequency graph to form a two-dimensional data set with time-frequency characteristics.
The first training module 905 is specifically configured to:
training the VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, respectively training a GoogleLeNet model and a ResNet model by using a second training set corresponding to the two-dimensional data set to obtain a VGGNet model, a GoogleLeNet model and a ResNet model, training the VGGNet model by using the first training set corresponding to the power quality disturbance signal sample data, and updating network parameters of the models by using a momentum fractional order gradient descent optimizer when the GoogleLeNet model and the ResNet model are respectively trained by using the second training set corresponding to the two-dimensional data set.
According to the electric energy quality disturbance signal classification system provided by the embodiment of the invention, the one-dimensional disturbance signal is subjected to dimension expansion, a two-dimensional data sample is constructed, more detailed characteristics of the disturbance signal can be reflected, the noise immunity is improved, the one-dimensional data and the two-dimensional data are utilized to respectively train a VGGNet model, a GoogLeNet model and a ResNet model, the characteristics of the disturbance signal are extracted from different dimensions, the accuracy of the characteristic extraction of the disturbance signal is ensured, finally, the characteristics of a plurality of disturbance signals are fused by utilizing a characteristic fusion module, the robustness of electric energy quality disturbance signal classification is improved, and the technical problems that the existing electric energy quality classification method cannot give consideration to both accuracy and interference immunity and is low in reliability are solved.
The electric energy quality disturbing signal classification system provided in the embodiment of the present invention is configured to execute any one of the electric energy quality disturbing signal classification methods in the foregoing electric energy quality disturbing signal classification method embodiments, and can obtain the same technical effect as the electric energy quality disturbing signal classification method in the foregoing electric energy quality disturbing signal classification method embodiments, which is not described herein again.
The above-mentioned embodiments are only used for illustrating the technical solutions of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims (10)

1. A method for classifying power quality disturbance signals is characterized by comprising the following steps:
constructing a power quality disturbance signal sample data matrix according to historical power quality disturbance signal data at nodes of the power system;
constructing a two-dimensional data set with time-frequency characteristics according to the power quality disturbance signal sample data matrix;
constructing a training set and a test set according to the power quality disturbance signal sample data matrix and the two-dimensional data set, wherein the training sets divided by the power quality disturbance signal sample data and the two-dimensional data set respectively comprise a first training set and a second training set;
constructing a first disturbance signal classification network model, wherein the disturbance signal classification network model comprises a VGGNet model, a GoogLeNet model and a ResNet model;
training a VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, and respectively training a GoogLeNet model and a ResNet model by using a second training set corresponding to a two-dimensional data set to obtain the well-trained VGGNet model, GoogLeNet model and ResNet model;
removing the fully connected layers and the Softmax function of the trained VGGNet model, GoogLeNet model and ResNet model, and accessing a feature fusion module and a fully connected layer classification module to the end pooling layers of the trained VGGNet model, GoogLeNet model and ResNet model to obtain a second disturbance signal classification network model;
training a second disturbance signal classification network model by using a second training set corresponding to the power quality disturbance signal sample data and a second training set corresponding to the two-dimensional data set to obtain a trained second disturbance signal classification network model;
and preprocessing the power quality disturbance signals to be classified into two-dimensional data with time-frequency characteristics, and inputting the power quality disturbance signals to be classified and the preprocessed two-dimensional data with the time-frequency characteristics into a trained second disturbance signal classification network model to obtain a classification result of the power quality disturbance signals.
2. The method for classifying the power quality disturbance signals according to claim 1, wherein a VGGNet model is trained by using a first training set corresponding to the sample data of the power quality disturbance signals, and when a GoogLeNet model and a ResNet model are respectively trained by using a second training set corresponding to a two-dimensional data set, a momentum score step reduction optimizer is used for updating the network parameters of the models.
3. The power quality disturbance signal classification method according to claim 1, wherein the ratio of the first training set, the second training set and the test set is 3:3:1 or 7:7: 3.
4. The method for classifying the power quality disturbance signals according to claim 1, wherein the VGGNet model in the disturbance signal classification network model is an improved VGGNet model, and the network structure of the improved VGGNet model is as follows:
the VGG-16 network of the VGGNet is used as a main framework, the length of a first convolution kernel is 32, the last three sequentially connected continuous convolution layers of the VGG-16 network are replaced by a composite convolution layer consisting of two parallel convolution layers, the sizes of the convolution kernels of the composite convolution layer are 3 x 1, the last three fully-connected layers of the VGG-16 network are reduced to two fully-connected layers, the output of each convolution layer is connected with a BN layer, and a Relu layer is added behind each BN layer for activation.
5. The power quality disturbance signal classification method according to claim 1, wherein the google lenet model in the disturbance signal classification network model is an improved google lenet model, and the network structure of the improved google lenet model is as follows:
adding a 7 × 7 convolutional layer to the inclusion module on the basis of the GoogLeNet model, adding a 1 × 1 convolutional layer for down-sampling corresponding to the 7 × 7 convolutional layer, removing the last fully-connected layer of the GoogLeNet model, and replacing the last fully-connected layer with a global pooling layer.
6. The method for classifying the power quality disturbance signals according to claim 1, wherein a ResNet model in the disturbance signal classification network model is an improved ResNet model, and a network structure of the improved ResNet model is as follows:
and taking the ResNet-50 network as a main framework, adding a BN layer in front of each convolution layer of each residual error unit in the ResNet-50 network, adding a Relu layer behind each BN layer for activation, and replacing the last average pooling layer of the ResNet-50 network with a pyramid pooling layer.
7. The method for classifying the power quality disturbance signals according to claim 1, wherein a two-dimensional data set with time-frequency characteristics is constructed according to a power quality disturbance signal sample data matrix, and the method comprises the following steps:
constructing a multi-resolution wavelet decomposition model;
determining the number of frequency band divisions and wavelet basis functions;
sample data in the power quality disturbance signal sample data matrix is serialized to obtain signal waveform data sets corresponding to different types of disturbance signals;
and performing multi-resolution wavelet decomposition on the samples in the signal waveform data set by using a multi-resolution wavelet decomposition model according to the frequency band division number and the wavelet basis functions, and converting the time-frequency signals obtained by decomposition under different scales into a two-dimensional time-frequency graph to form a two-dimensional data set with time-frequency characteristics.
8. A power quality disturbance signal classification system, comprising:
the one-dimensional data sample construction module is used for constructing a power quality disturbance signal sample data matrix according to historical power quality disturbance signal data at nodes of the power system;
the two-dimensional data sample construction module is used for constructing a two-dimensional data set with time-frequency characteristics according to the power quality disturbance signal sample data matrix;
the system comprises a sample dividing module, a data processing module and a data processing module, wherein the sample dividing module is used for constructing a training set and a test set according to a power quality disturbance signal sample data matrix and a two-dimensional data set, and the training set divided by the power quality disturbance signal sample data and the two-dimensional data set respectively comprises a first training set and a second training set;
the first model building module is used for building a first disturbance signal classification network model, and the disturbance signal classification network model comprises a VGGNet model, a GoogLeNet model and a ResNet model;
the first training module is used for training the VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, and respectively training a GoogLeNet model and a ResNet model by using a second training set corresponding to the two-dimensional data set to obtain the well-trained VGGNet model, GoogLeNet model and ResNet model;
the second model building module is used for removing the full connection layers and the Softmax function of the trained VGGNet model, GoogLeNet model and ResNet model, and accessing the feature fusion module and the full connection layer classification module to the end pooling layers of the trained VGGNet model, GoogLeNet model and ResNet model to obtain a second disturbance signal classification network model;
the second training module is used for training a second disturbance signal classification network model by using a second training set corresponding to the power quality disturbance signal sample data and a second training set corresponding to the two-dimensional data set to obtain a trained second disturbance signal classification network model;
and the disturbance signal identification module is used for preprocessing the power quality disturbance signals to be classified into two-dimensional data with time-frequency characteristics, and inputting the power quality disturbance signals to be classified and the preprocessed two-dimensional data with the time-frequency characteristics into a trained second disturbance signal classification network model to obtain the classification results of the power quality disturbance signals.
9. The power quality disturbance signal classification system according to claim 8, wherein the two-dimensional data sample construction module is specifically configured to:
constructing a multi-resolution wavelet decomposition model;
determining the number of frequency band divisions and wavelet basis functions;
sample data in the power quality disturbance signal sample data matrix is serialized to obtain signal waveform data sets corresponding to different types of disturbance signals;
and performing multi-resolution wavelet decomposition on the samples in the signal waveform data set by using a multi-resolution wavelet decomposition model according to the frequency band division number and the wavelet basis functions, and converting the time-frequency signals obtained by decomposition under different scales into a two-dimensional time-frequency graph to form a two-dimensional data set with time-frequency characteristics.
10. The method for classifying the power quality disturbance signal according to claim 1, wherein the first training module is specifically configured to:
training the VGGNet model by using a first training set corresponding to the power quality disturbance signal sample data, respectively training a GoogleLeNet model and a ResNet model by using a second training set corresponding to the two-dimensional data set to obtain a VGGNet model, a GoogleLeNet model and a ResNet model, training the VGGNet model by using the first training set corresponding to the power quality disturbance signal sample data, and updating network parameters of the models by using a momentum score gradient descent optimizer when the GoogleLeNet model and the ResNet model are respectively trained by using the second training set corresponding to the two-dimensional data set.
CN202111051091.8A 2021-09-08 2021-09-08 Electric energy quality disturbance signal classification method and system Active CN113780160B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111051091.8A CN113780160B (en) 2021-09-08 2021-09-08 Electric energy quality disturbance signal classification method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111051091.8A CN113780160B (en) 2021-09-08 2021-09-08 Electric energy quality disturbance signal classification method and system

Publications (2)

Publication Number Publication Date
CN113780160A true CN113780160A (en) 2021-12-10
CN113780160B CN113780160B (en) 2023-01-24

Family

ID=78841834

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111051091.8A Active CN113780160B (en) 2021-09-08 2021-09-08 Electric energy quality disturbance signal classification method and system

Country Status (1)

Country Link
CN (1) CN113780160B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114936947A (en) * 2022-07-18 2022-08-23 四川轻化工大学 High-voltage direct-current transmission line fault diagnosis method based on GADF-VGG16
CN117240624A (en) * 2023-11-14 2023-12-15 长春大学 Method and device for generating and testing anti-attack sample based on black box scene

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359693A (en) * 2018-10-24 2019-02-19 国网上海市电力公司 A kind of Power Quality Disturbance Classification Method
CN111783558A (en) * 2020-06-11 2020-10-16 上海交通大学 Satellite navigation interference signal type intelligent identification method and system

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359693A (en) * 2018-10-24 2019-02-19 国网上海市电力公司 A kind of Power Quality Disturbance Classification Method
CN111783558A (en) * 2020-06-11 2020-10-16 上海交通大学 Satellite navigation interference signal type intelligent identification method and system

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
王维博等: "基于特征融合一维卷积神经网络的电能质量扰动分类", 《电力系统保护与控制》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114936947A (en) * 2022-07-18 2022-08-23 四川轻化工大学 High-voltage direct-current transmission line fault diagnosis method based on GADF-VGG16
CN117240624A (en) * 2023-11-14 2023-12-15 长春大学 Method and device for generating and testing anti-attack sample based on black box scene
CN117240624B (en) * 2023-11-14 2024-01-23 长春大学 Method and device for generating and testing anti-attack sample based on black box scene

Also Published As

Publication number Publication date
CN113780160B (en) 2023-01-24

Similar Documents

Publication Publication Date Title
CN109492830B (en) Mobile pollution source emission concentration prediction method based on time-space deep learning
AU2020214409A1 (en) Structural damage identification method based on ensemble empirical mode decomposition and convolution neural network
Xia et al. Multi-stage fault diagnosis framework for rolling bearing based on OHF Elman AdaBoost-Bagging algorithm
CN110751108B (en) Subway distributed vibration signal similarity determination method
CN113780160B (en) Electric energy quality disturbance signal classification method and system
CN109886464B (en) Low-information-loss short-term wind speed prediction method based on optimized singular value decomposition generated feature set
CN111126868B (en) Road traffic accident occurrence risk determination method and system
CN110334865B (en) Power equipment fault rate prediction method and system based on convolutional neural network
CN112798888B (en) Non-invasive fault diagnosis method for vehicle-mounted electrical system of unmanned train
CN114355240A (en) Power distribution network ground fault diagnosis method and device
CN113295413B (en) Traction motor bearing fault diagnosis method based on indirect signals
CN114548199A (en) Multi-sensor data fusion method based on deep migration network
CN114021424B (en) PCA-CNN-LVQ-based voltage sag source identification method
CN112085062A (en) Wavelet neural network-based abnormal energy consumption positioning method
CN114880538A (en) Attribute graph community detection method based on self-supervision
CN116106672A (en) Vehicle network resonance detection method and device based on data driving and engineering knowledge
CN115015719A (en) Sleeve partial discharge recognition algorithm based on optical fiber ultrasonic sensing signal
CN111141879A (en) Deep learning air quality monitoring method, device and equipment
CN117689082A (en) Short-term wind power probability prediction method, system and storage medium
CN115619999A (en) Real-time monitoring method and device for power equipment, electronic equipment and readable medium
CN117493930A (en) Multi-element time sequence anomaly detection method based on contrast generation countermeasure network
CN113740671A (en) Fault arc identification method based on VMD and ELM
CN109239006B (en) Substance identification method and device based on humidity compensation model and storage medium
CN116358872A (en) Bearing residual life prediction method based on deep learning
CN115936196A (en) Monthly rainfall model prediction method based on time sequence convolution network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant