CN109359693A - A kind of Power Quality Disturbance Classification Method - Google Patents
A kind of Power Quality Disturbance Classification Method Download PDFInfo
- Publication number
- CN109359693A CN109359693A CN201811240639.1A CN201811240639A CN109359693A CN 109359693 A CN109359693 A CN 109359693A CN 201811240639 A CN201811240639 A CN 201811240639A CN 109359693 A CN109359693 A CN 109359693A
- Authority
- CN
- China
- Prior art keywords
- power quality
- disturbance
- convolutional neural
- neural networks
- classification
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
Abstract
The present invention problem difficult for electrical energy power quality disturbance identification, proposes a kind of Power Quality Disturbance Classification Method based on transfer learning.This method is using ImageNet data set as set of source data, firstly, passing through the intrinsic abstract characteristics of AlexNet convolutional neural networks model extraction POWER QUALITY DISTURBANCE WAVEFORM；Then, the AlexNet convolutional neural networks model trained is subjected to transfer learning, establishes new AlexNet convolutional neural networks model, the classification to POWER QUALITY DISTURBANCE WAVEFORM signal.And a large amount of single disturbance and dual disturbance waveform signal are established by matlab, the AlexNet model trained through POWER QUALITY DISTURBANCE WAVEFORM signal after migration is trained and is tested.The experimental results showed that this paper transfer learning method has good classification accuracy and noise immunity to electrical energy power quality disturbance.
Description
Technical field
The present invention relates to a kind of power quality analysis method, specially a kind of Power Quality Disturbance Classification Method.
Background technique
With the rapid development of China's power industry, going deep into for power system reform and opening gradually for electricity market, electricity
Power department increasingly payes attention to the safety and economy of electric system.The quality of power quality, directly affects electric system
Safety and stability and economical operation, also the moment affects the production of industrial or agricultural efficient stable and the safety fortune of other industrial equipments
Row.However as the development of smart grid, distributed generation unit and nonlinear-load are increasing, cause power grid to occur various
Electrical energy power quality disturbance problem brings great inconvenience to user, and electrical energy power quality disturbance becomes the hot spot of current power distribution network research
One of.Electrical energy power quality disturbance is many kinds of in electric system, affected by noise, but also possible a variety of disturbance loads, leads to electricity
Energy quality disturbance classification is difficult.Electrical energy power quality disturbance accurately identifies, and is the premise for carrying out power quality analysis and assessment
And basis, aid decision is provided for the management and improvement of power quality in power distribution network.
Electrical energy power quality disturbance identification process includes two links of feature extraction and pattern classification.Feature extraction refers to signal
Waveform is converted, then extracts the characteristic quantity that can reflect disturbance event.Common feature extracting method includes: discrete fourier
Convert (discrete Fourier transform, DFT), Short Time Fourier Transform (short-term Fourier
Transform, STFT), wavelet transformation (wavelet transform, WT), Hilbert-Huang transform (Hilbert-Huang
Transform, HHT), S-transformation, Kalman filter etc..DFT can only differentiate frequency domain, can be used for analyzing stable state disturbance, cannot analyze
Short time disturbance；The window of STFT is fixed, and time frequency resolution is also fixed, and is not suitable for analysis Transient Disturbance Signal；WT is in low frequency part
Frequency resolution with higher and lower temporal resolution, in high frequency section then temporal resolution with higher and lower
Frequency resolution.Therefore, WT has adaptivity, can obtain the mutation in Time And Frequency information, especially signal simultaneously
Information, but WT not can be carried out good detection to low-frequency excitation, and its operand is big, can not meet wanting for real-time detection
It asks；HHT noiseproof feature is preferable, and for non-linear, non-stationary signal has stronger processing capacity, can be accurately obtained signal
T/F-energy distribution characteristics, but there are boundary effects；S-transformation eliminates the selection of window function, and improves window
Wide fixed disadvantage, has good time frequency analysis characteristic, as a result more intuitive, and noiseproof feature is good, is current using most
More feature extraction algorithms, but information content of the disturbing signal after S-transformation is decomposed is too big, feature selecting extraction operation is difficult,
And its time frequency resolution is lower, so that feature extraction precision is restricted；Kalman filter sensibility is poor, smaller in amplitude
When, it is difficult to the state transition of identification amplitude in real time.Pattern classification is for determining disturbance generic, and main method has: decision tree
(decision tree, DT), artificial neural network (artificial neural network, ANN) and support vector machines
(support vector machine, SVM) etc..DT structure is simple, and the logical thinking for simulating the mankind realizes classification, is easy to manage
Solution, classification effectiveness is higher, but selection of its classification accuracy dependent on character subset and classification thresholds, and is easy to happen
Over-fitting；ANN can learn and store a large amount of input-output mode map relationships, close without understanding this mapping of description in advance
The math equation of system, but a large amount of training sample is needed, and the training time is long, easily falls into local optimum；SVM is based on statistics
The theories of learning are a kind of machine learning methods for solving few, non-linear and higher-dimension sample the pattern recognition problem of sample size,
Overcome the disadvantage that artificial neural network easily falls into locally optimal solution and training time length.
The method that the above-mentioned compound disturbance sorting algorithm feature extraction of power quality mostly uses mathematic(al) manipulation, complex steps are multiple
It is miscellaneous, computationally intensive, and be easy that original signal Partial Feature is caused to lose；The accuracy of identification of classifier is not ideal enough, and instructs
It is less to practice sample, causes accuracy of identification undesirable；Research object is mostly focused on single Power Quality Disturbance, and practical electric
Electrical energy power quality disturbance is often the compound disturbance of a variety of disturbance compositions in Force system, and feature overlapping, feature easily occurs in conventional method
The problems such as failure and nicety of grading decline, for the research also prematurity of compound disturbed depth problem.
Summary of the invention
The purpose of the present invention is to provide a kind of Power Quality Disturbance Classification Methods, can be by POWER QUALITY DISTURBANCE WAVEFORM figure
Classify by its disturbing cause.The present invention carries out the AlexNe convolutional neural networks model trained using ImageNet data set
Transfer learning, for being identified to POWER QUALITY DISTURBANCE WAVEFORM figure.By introducing transfer learning, of the invention overcomes tradition
The drawbacks described above of sorting algorithm, is greatly improved accuracy of identification.
In order to achieve the above object, the present invention proposes a kind of Power Quality Disturbance Classification Method, using Imagenet image
Divided data collection and AlexNet convolutional neural networks model classify Power Disturbance waveform collection according to disturbance Producing reason,
The AlexNet convolutional neural networks model first to layer 5 is convolutional layer, and the 6th to the 8th layer is full articulamentum, finally also
It is connected to a Softmax classification layer, comprising the following steps:
S1, using Imagenet image data set as set of source data, training obtains being extracted Imagenet image data set
The AlexNet convolutional neural networks model of intrinsic abstract characteristics；
S2, feature migration is carried out based on transfer learning principle, utilizes the AlexNet convolutional Neural net trained in step S1
Network model establishes new AlexNet convolutional neural networks model；
The parameter of S3, random initializtion Softmax classification layer；
S4, each Power Quality Disturbance waveform image to be sorted is converted into 227*227*3 pixel；Described it will turn
Input of the image as new AlexNet convolutional neural networks model after changing, to new AlexNet convolutional neural networks model
It is trained, and obtains the classification results of Power Quality Disturbance waveform from Softmax classification layer.
S5, the Power Quality Disturbance for generating a variety of perturbation schemes are verified through Power Quality Disturbance waveform image
The classification accuracy rate for the AlexNet convolutional neural networks model trained.
The Imagenet image score is according to collection, specially Imagenet image classification match data set ILSVRC- used
2011 data sets include 1,200,000 photos of having classified.
Established described in step S2 new AlexNet convolutional neural networks model, in particular to, will will be in step S1
The last one full articulamentum is replaced with one by the AlexNet convolutional neural networks model that Imagenet image data set is trained
A new full articulamentum.
Described in step S3 initialization Softmax classification layer parameter, in particular to: by the output classification of Softmax
It is changed to this paper disturbance type number, the parameter value of WeightLearnRateFactor and BiasLearnRateFactor are all set as
50.Initial learning rate is 0.0001, and maximum number cycle of training is 8.
Step S5 comprising the following steps:
S51: according to perturbation scheme, 9 kinds of single disturbances and 17 kinds of dual disturbances are randomly generated in MATLAB environment, altogether
Count 26 kinds of electrical energy power quality disturbance emulation signals；
S52, according to the single Disturbance Model expression formula of power quality, founding mathematical models divide the emulation signal
Class；
S53: the AlexNet convolutional neural networks model trained using step S4 is trained and is surveyed to emulation signal
The classification accuracy rate of new AlexNet convolutional neural networks model is verified in examination according to classification results.
The AlexNet convolutional neural networks model trained using step S5 is trained a variety of disturbing signals of generation
And test, classification results are obtained, the classification that new AlexNet convolutional neural networks model is verified according to the classification results is correct
Rate.
9 kinds of single disturbances described in step S51, in particular to temporarily liter, temporarily drop, interruption, harmonic wave, spike, cut mark, flickering, vibration
It swings and this nine kinds disturbances of pulse.
26 kinds of dual disturbances described in step S51, in particular to temporary liter plus harmonic wave, temporary rise add flickering, temporary rise to add oscillation, temporary liter
Adding pulse, temporary drop plus harmonic wave, temporary drop to add, flickering, temporarily drop adds oscillation, temporary drop plus pulse, interruption plus harmonic wave, interruption plus flickering, interrupts
Add oscillation, interrupt plus pulse, harmonic wave add flickering, harmonic wave add oscillation, harmonic wave add pulse, flickering add oscillation, flickering add pulse this 26
The dual disturbance of kind.
Compared with the prior art, the advantages of the present invention are as follows: use the ImageNet number comprising 1,200,000 trained pictures
According to collection, all more horn of plenty in image data amount and type, so that the AlexNet convolution mind trained through ImageNet data set
More, more accurate characteristics of image can be extracted through network model.The present invention will also be trained with ImageNet data set
AlexNet convolutional neural networks model moved on the problem of electrical energy power quality disturbance identification.Embodiment proves, the migration
AlexNet convolutional neural networks model can greatly improve the accuracy of identification of POWER QUALITY DISTURBANCE WAVEFORM, to electrical energy power quality disturbance wave
The classification of shape is more accurate, and noiseproof feature is good.
Detailed description of the invention
In order to illustrate more clearly of technical solution of the present invention, attached drawing needed in description will be made simply below
It introduces, it should be apparent that, the accompanying drawings in the following description is one embodiment of the present of invention, and those of ordinary skill in the art are come
It says, without creative efforts, is also possible to obtain other drawings based on these drawings:
Fig. 1 is flow chart of the invention.
Fig. 2 is the schematic diagram of transfer learning.
Fig. 3 is AlexNet model structure.
Fig. 4 is the specific method schematic diagram of progress transfer learning of the invention.
Fig. 5 is the single disturbance classification based training result of embodiment.
Fig. 6 is the dual disturbance classification based training result of embodiment.
Specific embodiment
The technical features, objects and effects for a better understanding of the present invention with reference to the accompanying drawing carry out more the present invention
To describe in detail.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to limit this
Patent of invention.It should be noted that being all made of very simplified form in these attached drawings and using non-accurate ratio, only use
In convenience, clearly aid in illustrating the invention patent.
The present invention proposes a kind of Power Quality Disturbance Classification Method, using Imagenet image score according to collection and AlexNet
Convolutional neural networks model classifies Power Disturbance waveform collection according to disturbance Producing reason, as shown in figure 3, AlexNet
Convolutional neural networks model first to layer 5 is convolutional layer, and the 6th to the 8th layer is full articulamentum, is finally further connected with one
Softmax classification layer.The convolutional layer carries out feature extraction work to input picture, is most important portion in convolutional neural networks
Point.The AlexNet convolutional neural networks model also includes three pond layers.Convolutional neural networks and general neural network are maximum
Difference be exactly that the former contains the feature extractor being made of convolutional layer and pond layer.With different convolution kernels to picture
Convolution is carried out to obtain the response in different characteristic, completes the extraction of a variety of local features.
In addition, being the perceptron of most original to avoid neural network from degenerating, each convolutional layer requires one activation of selection
Function makes the relationship that nonlinearity is kept between input and output.Usually using Simoid in traditional convolutional neural networks
FunctionAs activation primitive, but AlexNet model is used as using ReLU function phi (x)=max (0, x) and is swashed
The problem of overfitting problem can be effectively relieved in function living, and gradient disappears when improvement Simoid function backpropagation.
The pond layer can promote network to the ability of the anti-distortion such as input picture translation rotation.And image passes through convolution
The characteristic dimension obtained after operation is very high, if directly acting on classifier, it is existing to be easy to appear over-fitting in the training process
As.To avoid such phenomenon from occurring, pondization operation is carried out to convolution feature behind convolutional layer, characteristics of image dimension can be reduced,
And the number of parameter in neural network is reduced, improve robustness.The final layer of more sorter networks needs to export target to be sorted and exists
Prediction probability value of all categories exports each weighted value to network by Softmax function in AlexNet model and is calculated
Prior probability under final more classification.In addition, AlexNet model additionally uses ReLU activation primitive, Dropout strategy, part
Response normalization LRN etc. improves the performance of network.The shared training parameter that can reduce network of the weight of convolutional neural networks, makes
Its is more adaptable, and feature extraction and pattern classification carry out simultaneously, and generates and use in training simultaneously
Overlapping pool technology, can also be to avoid the generation of over-fitting.Therefore the feature extraction functions powerful using convolutional neural networks, mention
Taking the feature of Power Quality Disturbance waveform has feasibility.
It is of the invention to Power Quality Disturbance Classification Method the following steps are included:
S1, using Imagenet image data set as set of source data, training obtains being extracted Imagenet image data set
The AlexNet convolutional neural networks model of intrinsic abstract characteristics；The Imagenet image score is according to collection, specially Imagenet
Image classification match data set ILSVRC-2011 data set used, includes 1,200,000 photos of having classified.
S2, feature migration is carried out based on transfer learning principle, as shown in figure 4, establishing new AlexNet convolutional neural networks
Model, the AlexNet convolutional neural networks model that Imagenet image data set in step S1 is trained are complete by the last one
Articulamentum replaces with a new full articulamentum；
S3, random initializtion Softmax classification layer parameter, in particular to: by the output classification of Softmax be changed to herein
The parameter value of disturbance type number, WeightLearnRateFactor and BiasLearnRateFactor, are all set as 50.It is initial to learn
Habit rate is 0.0001, and maximum number cycle of training is 8.；
S4, each Power Quality Disturbance waveform image to be sorted is converted into 227*227*3 pixel；Described it will turn
Input of the image as new AlexNet convolutional neural networks model after changing, to new AlexNet convolutional neural networks model
It is trained, and obtains the classification results of Power Quality Disturbance waveform from Softmax classification layer.
S5, the Power Quality Disturbance for generating a variety of perturbation schemes are verified through Power Quality Disturbance waveform image
The classification accuracy rate for the AlexNet convolutional neural networks model trained.Comprising the following steps:
S51, according to perturbation scheme, in MATLAB environment, be randomly generated 9 kinds it is single disturbance and 17 kinds of dual disturbances,
The emulation signal of total 26 class Power Quality Disturbances, every kind respectively generates 2000 samples, and signal fundamental frequency f0 is 50Hz,
Sample frequency fs is 6.4kHz；
S52, mathematical transformation model is established according to the single Disturbance Model expression formula of table 1, to the emulation signal detection, divided
Class；9 kinds of single disturbances, in particular to temporarily rise, temporarily drop, interruption, harmonic wave, spike, cut mark, flickering, oscillation and pulse this nine
Kind disturbance.The principle that disturbing signal generates are as follows: single disturbance can be divided into four major class: voltage dip, electricity by disturbance definition and feature
Pressing temporary liter, voltage interruption, spike, cut mark is the disturbance of the 1st class；Oscillation and pulse are the disturbance of the 2nd class；Harmonic wave and flickering are respectively one
Class disturbance, the disturbance for belonging to a classification cannot be mutually mixed.According to the principle that disturbing signal generates, described 26 kinds dual to be disturbed
It is dynamic, in particular to temporarily rise plus harmonic wave, temporarily rise plus flickering, temporarily rise plus oscillation, temporarily rise plus pulse, temporarily drop plus harmonic wave, temporarily drop plus flickering,
Temporarily drop plus oscillation, temporarily drop plus pulse, interrupt plus harmonic wave, interrupt plus flickering, interrupt plus oscillation, interrupt plus pulse, harmonic wave add flickering,
Harmonic wave adds oscillation, harmonic wave to add pulse, flickering that oscillation, flickering is added to add this 26 kinds of dual disturbances of pulse.T is power frequency period, u in table 1
It (t) is unit jump function, t1 and t2 are respectively to disturb starting and end time, and sample of signal length was 10 periods.Due to reality
Border power quality data will receive the influence of noise, so being superimposed signal-to-noise ratio respectively on emulation signal is 20db, 30db, 40db
With the white Gaussian noise of 50db.
The single Disturbance Model expression formula (ω of table 10=2 π 50rads)
S53, the AlexNet convolutional neural networks model trained using step S5 carry out a variety of disturbing signals of generation
Training and test, obtain classification results, the classification of new AlexNet convolutional neural networks model are verified according to the classification results
Accuracy.The present invention is using momentum stochastic gradient descent method (SGDM) come training data.It is completed using ten folding cross validations to mould
The test of type takes in that is, every class 200 waveform samples as test data in turn, remaining 1800 sample is as training numbers
According to using the average value of 10 trained test results as final recognition accuracy.
Transfer learning described in step S2 is machine learning field field more popular in recent years, by the wide of scholar
General concern, and achieve a series of research achievements.The purpose is to come the knowledge migration acquired in source domain to target domain
Solve goal task.Fig. 2 is transfer learning schematic diagram, and different from machine learning algorithm, transfer learning can be applied in less sample
This situation, and do not need to do same distributional assumption, convolutional Neural net is used mainly for the specific area collection of limited sample size
Over-fitting is easy to produce when network and leads to not the problem of training is with study.Transfer learning may be defined as: give a source domain DS
With originating task TS, an aiming field DTWith goal task TT, transfer learning makes DSAnd TSKnowledge can help to solve TT, wherein
DS≠TSOr DT≠TT.Domain D is defined as a binary to { x, P (X) }, and wherein x is characterized space, and P (X) is the edge distribution of X, X
={ x1,x2,...,xn∈ x, task T is also a binary to { y, f (x) }, and wherein y is Label space, and y=f (x) is from instruction
Practice sample { xi,yi}(xi∈X,yi∈ y) objective function that learns.Number of training in source domain is denoted as nS, in aiming field
Be denoted as nT。
Traditional feature learning is to learn a disaggregated model on the basis of giving and training up sample, then utilizes
The disaggregated model practised classifies to test sample.The support that this process needs largely to have marked training sample is completed,
If feature learning effect can be made undesirable or even be unable to complete study without a large amount of labeled data.And in power quality
It in field, has largely marked POWER QUALITY DISTURBANCE WAVEFORM and has been extremely difficult to achieve, needed greatly because mark is trained with test sample
The manpower and material resources of amount, and requirement of the study of a large amount of training samples to computer hardware is very high, realizes difficult.In face of above-mentioned
Difficulty, transfer learning provide a kind of new method solved the problems, such as, it allows to migrate existing knowledge to solve in target domain
The problem concerning study of sample data has only been marked on a small quantity, that is, different but related fields problem has been carried out with existing knowledge
It solves.On the whole, transfer learning is very important a part in machine learning, and solves electrical energy power quality disturbance identification and ask
The effective ways of topic.
In electrical energy power quality disturbance identification, the tape label data that can be collected into are very limited.If directly used
Sample size less data collection one convolutional neural networks of training easily generated quasi- since the number of parameters of network is larger
Close phenomenon.In view of the above-mentioned problems, the present invention combines transfer learning with AlexNet model, i.e. source domain DS is ImageNet figure
As data set, DT is POWER QUALITY DISTURBANCE WAVEFORM collection.The identification of POWER QUALITY DISTURBANCE WAVEFORM will be converted into the automatic of figure
It identifies classification problem, the feature on the ImageNet data set for being all image data in the AlexNet model of pre-training is mentioned
Layer is taken to move in the classification task of POWER QUALITY DISTURBANCE WAVEFORM data set, then the ginseng of random initializtion Softmax classification layer
Number is simultaneously trained with electric energy quality waveform, completes the identification to POWER QUALITY DISTURBANCE WAVEFORM.Relative to electrical energy power quality disturbance wave
Graphic data collection, ImageNet data set is all very rich in data volume and type, the image classification data collection of ImageNet
Training set includes 1229413 pictures in ILSVRC2011, and it includes 50000 pictures that verifying, which is concentrated, includes in test set
100000 pictures.Although without a large amount of POWER QUALITY DISTURBANCE WAVEFORM image in ImageNet, since power quality is disturbed
Dynamic identification is equally using characteristics of image such as edge, textures, and therefore, the migration of transfer learning model is that convolutional neural networks are powerful
Ability in feature extraction.Rich and varied image space characteristic information can be extracted as set of source data using ImageNet,
And then on the problem of moving to electrical energy power quality disturbance identification.
A kind of Power Quality Disturbance Classification Method based on transfer learning of the invention.In view of electric energy matter in practical application
It is less to measure disturbance waveform, it is difficult to obtain enough samples, present invention introduces the methods of transfer learning, will use ImageNet data
Collect on the problem of AlexNet convolutional neural networks model trained moves to electrical energy power quality disturbance identification, by being extracted figure
As the convolutional neural networks automatic identification Power Quality Disturbance waveform of feature, reduce data prediction, feature extraction, spy
The complicated manual operations such as sign expression.Embodiment the result shows that, disaggregated model proposed by the invention can effectively improve electric energy
The nicety of grading of quality disturbance waveform is significantly better than traditional classification model.
Table 2 is single disturbance classification results.The training result of single disturbance is as shown in Figure 5.As can be seen from Table 2, this hair
Bright method and traditional mathematics transform method have higher accuracy rate for the identification of single disturbance, and model of the present invention becomes compared with mathematics
Accuracy rate is changed slightly to improve.Since single disturbing signal is fairly simple, is classified by mathematic(al) manipulation extraction feature and had
There is good effect, and the advantage of model of the present invention is more embodied in the classification to complicated compound disturbance.
The single disturbance classification results of table 2
It is as shown in Figure 6 to the training result of 17 kinds of dual disturbances.Classification results are as shown in Table 3, 4.
The dual disturbance the simulation results of table 3
Table 4 disturbs classification results
From table 3,4 it is found that the classification accuracy of disaggregated model of the present invention is significantly higher than mathematical transformation model, Average Accuracy
Up to 99.1%.When compound disturbance, due to being interfered with each other between multiple disturbances, so that explicitly manually being extracted by mathematic(al) manipulation
Feature is difficult, and complex steps are complicated, cause classification accuracy low, and transfer learning model of the present invention can be implicitly
Learnt from training data, the shortcomings that wave character is extracted in traditional mathematics transformation is overcome, so that accuracy rate greatly improves.
Moreover, model of the present invention still has higher recognition accuracy, illustrates the present invention when being superimposed the white Gaussian noise of different signal-to-noise ratio
Model noiseproof feature is good.
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
Those familiar with the art in the technical scope disclosed by the present invention, can readily occur in various equivalent modifications or replace
It changes, these modifications or substitutions should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with right
It is required that protection scope subject to.
Claims (7)
1. a kind of Power Quality Disturbance Classification Method, using Imagenet image score according to collection and AlexNet convolutional neural networks
Model classifies Power Disturbance waveform collection according to disturbance Producing reason, the AlexNet convolutional neural networks model the
One to layer 5 is convolutional layer, and the 6th to the 8th layer is full articulamentum, is finally further connected with Softmax classification layer, feature
It is, comprising the following steps:
S1, using Imagenet image data set as set of source data, training obtains being extracted Imagenet image data set intrinsic
The AlexNet convolutional neural networks model of abstract characteristics；
S2, feature migration is carried out based on transfer learning principle, utilizes the AlexNet convolutional neural networks mould trained in step S1
Type establishes new AlexNet convolutional neural networks model；
The parameter of S3, random initializtion Softmax classification layer；
S4, each Power Quality Disturbance waveform image to be sorted is converted into 227*227*3 pixel；After the conversion
Input of the image as new AlexNet convolutional neural networks model, new AlexNet convolutional neural networks model is carried out
Training, and the classification results of Power Quality Disturbance waveform are obtained from Softmax classification layer；
S5, the Power Quality Disturbance for generating a variety of perturbation schemes, verifying is through the training of Power Quality Disturbance waveform image
The classification accuracy rate for the AlexNet convolutional neural networks model crossed.
2. Power Quality Disturbance Classification Method as described in claim 1, which is characterized in that the Imagenet image score evidence
Collection, specially Imagenet image classification match data set ILSVRC-2011 data set used, include 1,200,000 photographs of having classified
Piece.
3. Power Quality Disturbance Classification Method as described in claim 1, which is characterized in that established newly described in step S2
AlexNet convolutional neural networks model, in particular to, by what Imagenet image data set in step S1 was trained
AlexNet convolutional neural networks model, the last one full articulamentum replace with a new full articulamentum.
4. Power Quality Disturbance Classification Method as described in claim 1, which is characterized in that initial described in the step S3
Change Softmax classification layer parameter, in particular to: the output classification of Softmax is changed to this paper disturbance type number,
The parameter value of WeightLearnRateFactor and BiasLearnRateFactor, is all set as 50, and initial learning rate is
0.0001, maximum number cycle of training is 8.
5. a kind of Power Quality Disturbance Classification Method as described in claim 1, which is characterized in that step S5 specifically includes following
Step:
S51: according to perturbation scheme, 9 kinds of single disturbances and 17 kinds of dual disturbances are randomly generated in MATLAB environment, amount to 26
Kind electrical energy power quality disturbance emulates signal；
S52, according to the single Disturbance Model expression formula of power quality, founding mathematical models classify to the emulation signal；
S53: the AlexNet convolutional neural networks model trained using step S4 is trained and is tested to emulation signal, according to
The classification accuracy rate of new AlexNet convolutional neural networks model is verified according to classification results.
6. a kind of Power Quality Disturbance Classification Method as claimed in claim 5, which is characterized in that described in step S51 9 kinds it is single
This nine kinds disturbance, in particular to temporarily liter, temporarily drop, interruption, harmonic wave, spike, cut mark, flickering, oscillation and pulse disturbances.
7. a kind of Power Quality Disturbance Classification Method as claimed in claim 5, which is characterized in that 26 kinds pairs described in step S51
Disturbance, in particular to temporary liter plus harmonic wave, temporary rise add flickering, temporary rise to add oscillation, temporary liter plus pulse, temporary drop that harmonic wave, temporary drop is added to add again
Flickering, temporary drop add oscillation, temporary drop to add pulse, interruption plus harmonic wave, interruption that flickering, interruption is added to add oscillation, interruption that pulse, harmonic wave is added to add
Flickering, harmonic wave add oscillation, harmonic wave to add pulse, flickering that oscillation, flickering is added to add this 26 kinds of dual disturbances of pulse.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811240639.1A CN109359693A (en) | 2018-10-24 | 2018-10-24 | A kind of Power Quality Disturbance Classification Method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811240639.1A CN109359693A (en) | 2018-10-24 | 2018-10-24 | A kind of Power Quality Disturbance Classification Method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109359693A true CN109359693A (en) | 2019-02-19 |
Family
ID=65346404
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811240639.1A Pending CN109359693A (en) | 2018-10-24 | 2018-10-24 | A kind of Power Quality Disturbance Classification Method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109359693A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110399796A (en) * | 2019-09-02 | 2019-11-01 | 国网上海市电力公司 | A kind of electrical energy power quality disturbance recognition methods based on improvement deep learning algorithm |
| CN110672905A (en) * | 2019-09-16 | 2020-01-10 | 东南大学 | CNN-based self-supervision voltage sag source identification method |
| CN110703006A (en) * | 2019-09-04 | 2020-01-17 | 国网浙江省电力有限公司金华供电公司 | Three-phase power quality disturbance detection method based on convolutional neural network |
| CN110728195A (en) * | 2019-09-18 | 2020-01-24 | 武汉大学 | Power quality disturbance detection method based on YOLO algorithm |
| CN110991603A (en) * | 2019-10-23 | 2020-04-10 | 广州市智能软件产业研究院 | Local robustness verification method of neural network |
| CN111105396A (en) * | 2019-12-12 | 2020-05-05 | 山东浪潮人工智能研究院有限公司 | Printed matter quality detection method and system based on artificial intelligence |
| CN111525587A (en) * | 2020-04-01 | 2020-08-11 | 中国电力科学研究院有限公司 | Reactive load situation-based power grid reactive voltage control method and system |
| CN111666984A (en) * | 2020-05-20 | 2020-09-15 | 海南电网有限责任公司电力科学研究院 | Intelligent overvoltage identification method based on transfer learning |
| CN112801049A (en) * | 2021-03-26 | 2021-05-14 | 北京至真互联网技术有限公司 | Image classification method, device and equipment |
| CN113780160A (en) * | 2021-09-08 | 2021-12-10 | 广东电网有限责任公司广州供电局 | Electric energy quality disturbance signal classification method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107704610A (en) * | 2017-10-18 | 2018-02-16 | 国网上海市电力公司 | A kind of power distribution network operation data event correlation analysis system and analysis method |
| CN108390382A (en) * | 2018-02-27 | 2018-08-10 | 杭州电力设备制造有限公司 | A kind of suppressing method and Research on Unified Power Quality Conditioner of electrical energy power quality disturbance |
| CN108664950A (en) * | 2018-05-22 | 2018-10-16 | 天津大学 | A kind of electrical energy power quality disturbance identification and sorting technique based on deep learning |
-
2018
- 2018-10-24 CN CN201811240639.1A patent/CN109359693A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107704610A (en) * | 2017-10-18 | 2018-02-16 | 国网上海市电力公司 | A kind of power distribution network operation data event correlation analysis system and analysis method |
| CN108390382A (en) * | 2018-02-27 | 2018-08-10 | 杭州电力设备制造有限公司 | A kind of suppressing method and Research on Unified Power Quality Conditioner of electrical energy power quality disturbance |
| CN108664950A (en) * | 2018-05-22 | 2018-10-16 | 天津大学 | A kind of electrical energy power quality disturbance identification and sorting technique based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 王知芳 等: "《基于卷积神经网络的电能质量扰动分类》", 《第九届电能质量研讨会论文集》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110399796A (en) * | 2019-09-02 | 2019-11-01 | 国网上海市电力公司 | A kind of electrical energy power quality disturbance recognition methods based on improvement deep learning algorithm |
| CN110703006A (en) * | 2019-09-04 | 2020-01-17 | 国网浙江省电力有限公司金华供电公司 | Three-phase power quality disturbance detection method based on convolutional neural network |
| CN110703006B (en) * | 2019-09-04 | 2022-06-21 | 国网浙江省电力有限公司金华供电公司 | Three-phase power quality disturbance detection method based on convolutional neural network |
| CN110672905A (en) * | 2019-09-16 | 2020-01-10 | 东南大学 | CNN-based self-supervision voltage sag source identification method |
| CN110728195A (en) * | 2019-09-18 | 2020-01-24 | 武汉大学 | Power quality disturbance detection method based on YOLO algorithm |
| CN110991603A (en) * | 2019-10-23 | 2020-04-10 | 广州市智能软件产业研究院 | Local robustness verification method of neural network |
| CN110991603B (en) * | 2019-10-23 | 2023-11-28 | 广州市智能软件产业研究院 | Local robustness verification method of neural network |
| CN111105396A (en) * | 2019-12-12 | 2020-05-05 | 山东浪潮人工智能研究院有限公司 | Printed matter quality detection method and system based on artificial intelligence |
| CN111525587B (en) * | 2020-04-01 | 2022-10-25 | 中国电力科学研究院有限公司 | Reactive load situation-based power grid reactive voltage control method and system |
| CN111525587A (en) * | 2020-04-01 | 2020-08-11 | 中国电力科学研究院有限公司 | Reactive load situation-based power grid reactive voltage control method and system |
| CN111666984A (en) * | 2020-05-20 | 2020-09-15 | 海南电网有限责任公司电力科学研究院 | Intelligent overvoltage identification method based on transfer learning |
| CN111666984B (en) * | 2020-05-20 | 2023-08-25 | 海南电网有限责任公司电力科学研究院 | Overvoltage intelligent identification method based on transfer learning |
| CN112801049A (en) * | 2021-03-26 | 2021-05-14 | 北京至真互联网技术有限公司 | Image classification method, device and equipment |
| WO2022198898A1 (en) * | 2021-03-26 | 2022-09-29 | 北京至真互联网技术有限公司 | Picture classification method and apparatus, and device |
| CN112801049B (en) * | 2021-03-26 | 2021-07-23 | 北京至真互联网技术有限公司 | Image classification method, device and equipment |
| CN113780160B (en) * | 2021-09-08 | 2023-01-24 | 广东电网有限责任公司广州供电局 | Electric energy quality disturbance signal classification method and system |
| CN113780160A (en) * | 2021-09-08 | 2021-12-10 | 广东电网有限责任公司广州供电局 | Electric energy quality disturbance signal classification method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109359693A (en) | A kind of Power Quality Disturbance Classification Method | |
| CN108664893A (en) | A kind of method for detecting human face and storage medium | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN106897738B (en) | A kind of pedestrian detection method based on semi-supervised learning | |
| CN109948647A (en) | A kind of electrocardiogram classification method and system based on depth residual error network | |
| CN106973057A (en) | A kind of sorting technique suitable for intrusion detection | |
| CN107846326A (en) | A kind of adaptive semi-supervised net flow assorted method, system and equipment | |
| CN110222953A (en) | A kind of power quality hybrid perturbation analysis method based on deep learning | |
| CN109002810A (en) | Model evaluation method, Radar Signal Recognition method and corresponding intrument | |
| CN107679453A (en) | Weather radar electromagnetic interference echo recognition methods based on SVMs | |
| CN109948726A (en) | A kind of Power Quality Disturbance Classification Method based on depth forest | |
| CN112183643B (en) | Hard rock tension-shear fracture identification method and device based on acoustic emission | |
| CN106874950A (en) | A kind of method for identifying and classifying of transient power quality recorder data | |
| Davies et al. | Deep neural networks for appliance transient classification | |
| CN110147760A (en) | A kind of efficient electrical energy power quality disturbance image characteristics extraction and identification new method | |
| CN111709468B (en) | Training method and device for directional artificial intelligence and storage medium | |
| CN107682109A (en) | A kind of interference signal classifying identification method suitable for UAV Communication system | |
| CN112200238B (en) | Hard rock pulling shear rupture identification method and device based on sound characteristics | |
| CN112200038B (en) | CNN-based quick identification method for oscillation type of power system | |
| CN110175657A (en) | A kind of image multi-tag labeling method, device, equipment and readable storage medium storing program for executing | |
| CN115270872A (en) | Radar radiation source individual small sample learning and identifying method, system, device and medium | |
| CN103336830B (en) | Image search method based on structure semantic histogram | |
| CN110321390A (en) | Based on the load curve data visualization method for thering is supervision and unsupervised algorithm to combine | |
| CN106372618A (en) | Road extraction method and system based on SVM and genetic algorithm | |
| CN113111786A (en) | Underwater target identification method based on small sample training image convolutional 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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190219 |
|
| RJ01 | Rejection of invention patent application after publication |