CN109521270A - Harmonic detecting method based on modified wavelet neural network - Google Patents
Harmonic detecting method based on modified wavelet neural network Download PDFInfo
- Publication number
- CN109521270A CN109521270A CN201811186689.6A CN201811186689A CN109521270A CN 109521270 A CN109521270 A CN 109521270A CN 201811186689 A CN201811186689 A CN 201811186689A CN 109521270 A CN109521270 A CN 109521270A
- Authority
- CN
- China
- Prior art keywords
- network
- neural network
- hidden layer
- wavelet neural
- detecting method
- 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
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R23/00—Arrangements for measuring frequencies; Arrangements for analysing frequency spectra
- G01R23/16—Spectrum analysis; Fourier analysis
-
- 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
- G06N3/08—Learning methods
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Measurement Of Resistance Or Impedance (AREA)
Abstract
The invention discloses a kind of harmonic detecting methods based on modified wavelet neural network, comprising: establishes network structure, setting network initial value according to wavelet neural network principle, weighting parameter uses self-correlation correction method, determines weight wjk, warp parameter a is determined according to weightjWith translation parameters bj, hidden layer propagation function selects Morlet wavelet function, and input layer number determines by the sampling number of signal, and node in hidden layer is using calculating hidden layer node empirical equation；It determines algorithm for training network, using the training algorithm of momentum arithmetic come smooth weights learning path, currently once corrects excessive, the training algorithm of momentum arithmetic can be reduced current correction amount, and vice versa.Normal operation method is reviewed one's lessons by oneself present invention introduces network parameter, improves the convergence of network.Avoiding network training and falling into Local Minimum causes the low problem of precision of detection to improve Harmonic Detection precision using the smooth learning path of the training algorithm of momentum arithmetic.
Description
Technical field
The present invention relates to a kind of harmonic detecting methods, more particularly, to the harmonic detecting based on modified wavelet neural network
Method.
Background technique
Nowadays the problem of being widely used due to electronic device, pollution of the electric system by harmonic wave, is also more and more prominent
Out, the presence of harmonic wave all constitutes certain influence to the reliability of power equipment, safety and service life etc. how
Accurate detection harmonic content has become the hot spot in the field.The 6th interim " utilization of volume 26 in 2002 " electric power network technique "
The detection of Morlet continuous wavelet transform realization nonstoichiometric oxide " text, it is each that selection divides stringent wavelet function reduction harmonic wave
Aliasing between frequency band is detected nonstoichiometric oxide using the method for continuous integral wavelet transformation and made with Fourier's detection method
Comparison confirms the feasibility of method through emulation, but in view of frequency modulation(PFM) phenomenon and does not calculate cumbersome." motor work in 2003
Journey journal " a kind of 9 phases " high-precision Algorithms for Harmonic Analysis of Power Systems " text of volume 23, it analyzes plus the FFT of Hanning window is inserted
Value-based algorithm acquires the principle of system frequency and linear neural network detection harmonic wave, proposes one kind based on FFT-Adaline algorithm simultaneously
For harmonic detecting.Under using adding the FFT interpolation algorithm of Hanning window to obtain electric system accurate frequency, then use Adaline
Neural network model analyzes harmonic wave, tests through simulation comparison, it was confirmed that the feasibility of the algorithm, but the requirement to signal processor
It is relatively high.The 7th phase " time-variant harmonic signal inspection based on wavelet neural network of volume 28 in 2008 " Proceedings of the CSEE "
Survey " text, adaptive time-frequency of the small echo to signal is divided into characteristic, introduces neural network so as to improve network performance, but detect
Precision depend on the selection of network parameter.
Summary of the invention
In order to solve the shortcomings of the prior art, the invention discloses a kind of based on the humorous of modified wavelet neural network
Wave detecting method, first against setting improper the problem of causing network performance to decline of network parameter, introduce network parameter from
Correction algorithm improves the convergence of network.Next avoids network training from falling into the problem that Local Minimum causes the precision of detection low,
Using the smooth learning path of the training algorithm of momentum arithmetic, to improve detection accuracy.
Technical solution provided by the invention is as follows:
Harmonic detecting method based on modified wavelet neural network, which comprises the following steps:
Network structure is established according to wavelet neural network principle, selects single input, single hidden layer, single export structure；
Setting network initial value, weighting parameter use self-correlation correction method, determine weight wjk, determined according to weight flexible
Parameter ajWith translation parameters bj, hidden layer propagation function select Morlet wavelet function, input layer number by signal sampled point
Number determines that node in hidden layer is using calculating hidden layer node empirical equation；
Determine algorithm for training network, using the training algorithm of momentum arithmetic come smooth weights learning path, currently once
Correct excessive, the training algorithm of momentum arithmetic can be reduced current correction amount, currently once correct it is too small, momentum arithmetic
Training algorithm can increase current correction amount.
Further, the self-correlation correction step of the weight are as follows:
Pay the random number on section [- 1,1] to weight wjk, then to wjkIt is normalized to obtain w'jk:
Multiplied by an auto-correlation coefficient c:
w″jk=cN1/m·w'jk (5)
The final selected weight for being input to hidden layer are as follows:
Flexible and translation parameters initial self-correlation correction are as follows:
xkmax、xkminNot to be maximum value and minimum value that sample is input to kth node, t*For wavelet function temporal center,
Δ t is time domain radius.
Further, the adjustment formula of the weight are as follows:
Gradient of the network output error to different parameters are as follows:
Further, the value range of the auto-correlation coefficient c is [2.3,2.6].
Further, the temporal center of the Morlet wavelet function is 0, and time domain radius is 0.7071.
Further, the expression formula of the Morlet wavelet function are as follows:
Further, the sampling number meets N > 2f in primitive period THT, fHFor harmonic signal highest frequency, N
For node in hidden layer.
Further, the calculating hidden layer empirical equation is N=2 × n+1, and n is the number of input layer.
Compared with prior art, the invention has the following advantages:
The present invention uses modified wavelet neural network, has saved a large amount of learning time, is optimized using parameter auto-correlation
Initial parameter is corrected, so that the selection of initial parameter has certain basis.For network training defect, with momentum arithmetic
Thought, effectively improve network performance.It is special that the segmentation of wavelet analysis time-frequency and the adaptive, self-adjusting of neural network etc. is utilized
Point improves the accuracy of detection.Finally, proving through emulation, which can effectively detect mains by harmonics.
Detailed description of the invention
Fig. 1 is one-dimensional WNN structural schematic diagram.
Fig. 2 is WNN training result schematic diagram.
Fig. 3 is BP training result schematic diagram.
Fig. 4 is WNN testing result schematic diagram.
Fig. 5 is BP neural network testing result schematic diagram.
Specific embodiment
The present invention is described in detail with reference to the accompanying drawing:
As shown in Figs. 1-5, the mathematic(al) representation of wavelet neural network are as follows:
Wherein:
H (t) is morther wavelet, | a |-1/2For normalization coefficient, b and a are respectively translation and the contraction-expansion factor of wavelet function.Letter
Number the change of partial structurtes resolution ratio can be realized by adjusting translation and contraction-expansion factor two parameters.
Wavelet function ψa,b(t), in the output of the network of t moment are as follows:
K=1,2 in formula, n；J=1,2, N；I=1,2, m.
According to wavelet neural network principle, single input, single hidden layer are selected, the network structure singly exported establishes network structure,
Input vector isExporting output vector isN and m is respectively input layer and defeated
Node layer number out, node in hidden layer are N number of.wjkAnd wijWeighted value between adjacent layer.
Setting network initial value, weighting parameter use self-correlation correction method, determine weight wjk:
Pay the random number on section [- 1,1] to weight wjk, then to wjkIt is normalized to obtain w'jk, such as formula (4) institute
Show:
Multiplied by an auto-correlation coefficient c:
w″jk=cN1/m·w'jk (5)
The value range of auto-correlation coefficient c is [2.3,2.6].
If the sample maximum for being input to kth node is xkmax, minimum value xkmin, finally selected to be input to hidden layer
Weight are as follows:
Flexible and translation parameters initial self-correlation correction are as follows:
t*For wavelet function temporal center, Δ t is time domain radius, and the temporal center of Morlet wavelet function is 0, time domain half
Diameter is 0.7071, and translation can be acquired and two initial parameter value ranges of stretching by substituting into formula (4).
The diversity of wavelet function makes the selection of wavelet basis function particularly critical, and the function difference of selection necessarily will cause
Difference on network output accuracy.According to some characteristics of the standard of choice of mother wavelet and the Harmonious Waves in Power Systems of analysis, choosing
Select propagation function of the Morlet small echo as hidden layer.The distinctive property of Morlet wavelet function can guarantee that signal reconstruction is not abnormal
Become, expression formula are as follows:
Input layer number is established by the sampling number of signal, by Nyquist theorem it is found that when sample frequency is sufficiently large,
The generation of block overlap of frequency bands could effectively be reduced.In primitive period T, sampling number must satisfy N > 2fHT, fHFor harmonic wave letter
Number highest frequency.Number of nodes selects inappropriate, and network performance must be affected, and the error reached is also not necessarily best.It is hidden
Selected take of the number containing node layer calculates hidden layer node empirical equation, as shown in formula (6):
N=2 × n+1 (9)
In formula, N is the number of hidden node, and n is the number of input layer.
The amendment of wavelet neural network weight is related with the single order local derviation of error function.It is non-linear due to activation primitive,
Error function is caused there are multiple minimum points, if first reaching local error minimum in training process, network weight and threshold value are just
It can not continue to adjust, global minima has not yet been reached in error at this time, and network falls into Local Minimum.It is calculated using the training of momentum arithmetic
Method comes smooth weights learning path, currently once corrects excessive, which can be reduced current correction amount, currently once correct
Small, the training algorithm of momentum arithmetic can increase current correction amount.The specific adjustment of parameter are as follows:
In formula, α is momentum term, and η is learning rate.
Gradient of the network output error to different parameters are as follows:
In the power system, alternating voltage and current signal are mainly fundamental frequency components, exist simultaneously other integral multiples
The high-order component of fundamental frequency.Due to the use of high power device, so that signal contains a large amount of harmonic waves.Overtone order and harmonic wave
Amplitude is inversely proportional, and now only considers to contain 3,5,7 subharmonic current component situations in harmonic current, other subharmonic can class according to this
It pushes away.If non-sine periodic signal are as follows:
If sample frequency fs=10kHz, signal highest subfrequency fH=350Hz, sample frequency fs> 2fH, satisfaction adopts
Sample requirement.Input layer number desirable 20, output layer number of nodes is harmonic wave number, can be taken as 4.Hidden layer node is 41.It learns
Practise efficiency eta=0.1, factor of momentum α=0.5.The initial position of first sample point of formation of the training sample of network is the moment
0, sample in 0.02s at 20 points, then the initial position of second sample point is moment Δ t, equally sample in 0.02s at 20 points,
The rest may be inferred.
Comparison diagram 2 and Fig. 3 can be clearly seen, pass through 92 times with the harmonic detecting method of modified wavelet neural network
After circuit training, close to the desired precision of error, a large amount of learning time has been saved, has been instructed with BP neural network training method
Experienced number is far longer than the training method of wavelet neural network.
Comparison diagram 4 and Fig. 5 can be clearly seen, the main lobe of each harmonic spectral line of wavelet neural network is very narrow and energy
It concentrates, detection accuracy is higher.There is Wavelet Neural Network adaptive and self-learning capability can inhibit to be permitted by training accordingly
The generation of more falseness frequencies.The multiresolution advantage for introducing wavelet analysis, has refined the frequency of signal, improves the essence of detection
Degree.Though BP neural network can detect some frequency informations, the function of refinement frequency is not had, so that containing in testing result
There is a large amount of pseudo- frequency, precision is greatly reduced.
Wavelet neural network combines the time-frequency multiresolution analysis of wavelet analysis and self-adjusting, the self study of neural network
The advantages that, largely improve accuracy in detection.Table 1 is WNN detection and BP neural network testing result.It can from table
Obtain that result is more accurate, and relative error also can control below 5% with WNN detection method out.
Table 1
Detection of the modified wavelet neural network to mains by harmonics.Firstly, utilizing the initial ginseng of parameter auto-correlation optimization amendment
Number, so that the selection of initial parameter has certain basis.Secondly, network training defect is directed to, with the think of of momentum arithmetic
Think, effectively improves network performance.Again, the segmentation of wavelet analysis time-frequency and the adaptive, self-adjusting of neural network etc. is utilized
Feature improves the accuracy of detection.Finally, proving through emulation, which can effectively detect mains by harmonics.
Above-described embodiment is not to this just for the sake of clearly illustrating technical solution of the present invention example
The embodiment of invention limits, and for those of ordinary skill in the art, can also make on the basis of the above description
Other various forms of variations or variation are made any modification, equivalent replacement all within the spirits and principles of the present invention and are changed
Into, category protection scope of the present invention.
Claims (7)
1. the harmonic detecting method based on modified wavelet neural network, which comprises the following steps:
S1. network structure is established according to wavelet neural network principle, selects single input, single hidden layer, single export structure；
S2. setting network initial value, weighting parameter use self-correlation correction method, determine weight wjk, determined according to weight flexible
Parameter ajWith translation parameters bj, hidden layer propagation function select Morlet wavelet function, input layer number by signal sampled point
Number determines that node in hidden layer is using calculating hidden layer node empirical equation；
The self-correlation correction step of the weight are as follows:
Pay the random number on section [- 1,1] to weight wjk, then to wjkIt is normalized to obtain w'jk:
Multiplied by auto-correlation coefficient c:
w”jk=cN1/m·w'jk
The final selected weight for being input to hidden layer are as follows:
Flexible and translation parameters initial self-correlation correction are as follows:
xkmax、xkminThe maximum value and minimum value of kth node, t are input to for sample*For Morlet wavelet function temporal center, Δ
T is time domain radius；
S3. algorithm for training network is determined, using the smooth weights learning path of the training algorithm of momentum arithmetic.
2. the harmonic detecting method according to claim 1 based on modified wavelet neural network, which is characterized in that described
The adjustment formula of S3 weight are as follows:
Gradient of the network output error to different parameters are as follows:
3. the harmonic detecting method according to claim 2 based on modified wavelet neural network, which is characterized in that described
The value range of auto-correlation coefficient c is [2.3,2.6].
4. the harmonic detecting method according to claim 2 based on modified wavelet neural network, which is characterized in that described
The temporal center of Morlet wavelet function is 0, and time domain radius is 0.7071.
5. the harmonic detecting method according to claim 2 based on modified wavelet neural network, which is characterized in that described
The expression formula of Morlet wavelet function are as follows:
6. the harmonic detecting method according to claim 2 based on modified wavelet neural network, which is characterized in that described
Sampling number meets N > 2f in primitive period THT, fHFor harmonic signal highest frequency, N is node in hidden layer.
7. the harmonic detecting method according to claim 6 based on modified wavelet neural network, which is characterized in that described
Calculating hidden layer empirical equation is N=2 × n+1, and n is the number of input layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811186689.6A CN109521270A (en) | 2018-10-11 | 2018-10-11 | Harmonic detecting method based on modified wavelet neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811186689.6A CN109521270A (en) | 2018-10-11 | 2018-10-11 | Harmonic detecting method based on modified wavelet neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109521270A true CN109521270A (en) | 2019-03-26 |
Family
ID=65772465
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811186689.6A Pending CN109521270A (en) | 2018-10-11 | 2018-10-11 | Harmonic detecting method based on modified wavelet neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109521270A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111308260A (en) * | 2020-04-16 | 2020-06-19 | 山东卓文信息科技有限公司 | Electric energy quality monitoring and electric appliance fault analysis system based on wavelet neural network and working method thereof |
| CN111693774A (en) * | 2020-05-06 | 2020-09-22 | 南方电网科学研究院有限责任公司 | Harmonic wave measuring method and device for power transmission network |
| CN114217132A (en) * | 2021-11-11 | 2022-03-22 | 江苏弈赫能源科技有限公司 | Power system harmonic law calculation method based on WGAN |
| CN115062668A (en) * | 2022-06-28 | 2022-09-16 | 合肥工业大学 | Harmonic parameter detection method and system based on RAdam optimization width learning |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070282772A1 (en) * | 2006-05-19 | 2007-12-06 | Inha-Industry Partnership Institute | Separate Learning System and Method Using Two-Layered Neural Network Having Target Values for Hidden Nodes |
| CN103150594A (en) * | 2013-03-18 | 2013-06-12 | 山东电力集团公司济宁供电公司 | Transformer fault diagnosis method based on added momentum item BP (back propagation) neural network |
-
2018
- 2018-10-11 CN CN201811186689.6A patent/CN109521270A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070282772A1 (en) * | 2006-05-19 | 2007-12-06 | Inha-Industry Partnership Institute | Separate Learning System and Method Using Two-Layered Neural Network Having Target Values for Hidden Nodes |
| CN103150594A (en) * | 2013-03-18 | 2013-06-12 | 山东电力集团公司济宁供电公司 | Transformer fault diagnosis method based on added momentum item BP (back propagation) neural network |
Non-Patent Citations (5)
| Title |
|---|
| 湛从昌、陈新元 编著: "《液压可靠性最优化与智能故障诊断》", 31 October 2015 * |
| 熊欣 著: "《电力谐波分析及应用》", 31 May 2017 * |
| 薛梅: "开关磁阻电机非线性建模方法及智能控制策略研究", 《中国博士学位论文全文数据库》 * |
| 陈林 等: "变形预测的小波神经网络模型改进", 《测绘科学》 * |
| 黄海波 等: "基于EEMD与GA小波神经网络的传动系声品质预测", 《振动与冲击》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111308260A (en) * | 2020-04-16 | 2020-06-19 | 山东卓文信息科技有限公司 | Electric energy quality monitoring and electric appliance fault analysis system based on wavelet neural network and working method thereof |
| CN111693774A (en) * | 2020-05-06 | 2020-09-22 | 南方电网科学研究院有限责任公司 | Harmonic wave measuring method and device for power transmission network |
| CN114217132A (en) * | 2021-11-11 | 2022-03-22 | 江苏弈赫能源科技有限公司 | Power system harmonic law calculation method based on WGAN |
| CN115062668A (en) * | 2022-06-28 | 2022-09-16 | 合肥工业大学 | Harmonic parameter detection method and system based on RAdam optimization width learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109521270A (en) | Harmonic detecting method based on modified wavelet neural network | |
| CN103308804B (en) | Based on quick K-S converting electric power quality disturbance signal time and frequency parameter extracting method | |
| CN104391178B (en) | A kind of time shift phase difference stable state harmonic signal bearing calibration based on Nuttall windows | |
| Su et al. | Power harmonic and interharmonic detection method in renewable power based on Nuttall double‐window all‐phase FFT algorithm | |
| Yılmaz et al. | An improved automated PQD classification method for distributed generators with hybrid SVM-based approach using un-decimated wavelet transform | |
| CN109063613A (en) | Based on the synchronous nonstationary random response method for extracting transformation of Generalized parametering | |
| Yao et al. | Measurement of power system harmonic based on adaptive Kaiser self‐convolution window | |
| CN109782063A (en) | A kind of dynamic m-Acetyl chlorophosphonazo analysis method based on three spectral line interpolation FFT of Nuttall self-convolution window | |
| CN105137180B (en) | High-precision harmonic analysis method based on six four spectral line interpolations of Cosine Window | |
| CN103308766A (en) | Harmonic analysis method based on Kaiser self-convolution window dual-spectrum line interpolation FFT (Fast Fourier Transform) and device thereof | |
| Kumar et al. | Recognition of underlying causes of power quality disturbances using stockwell transform | |
| CN107643446B (en) | A kind of multiline interpolation harmonic analysis method and system based on main lobe width | |
| CN105486938B (en) | A kind of substation's mixed noise separation method | |
| Liu et al. | A new classification method for transient power quality combining spectral kurtosis with neural network | |
| CN109521275A (en) | A kind of synchronized phasor determines method, system, device and readable storage medium storing program for executing | |
| CN110095650A (en) | The complicated harmonic detecting analysis method of four spectral line interpolation FFTs based on five Rife-Vincent (I) windows | |
| CN106771586B (en) | A kind of the loop signal analysis method and device of DC control protection board | |
| CN102818930A (en) | Method for quickly calculating power harmonic parameters in high-accuracy mode | |
| CN108918964A (en) | A kind of harmonic analysis method of sparsity enhancing | |
| Liu et al. | Dual-channel convolutional network-based fault cause identification for active distribution system using realistic waveform measurements | |
| JP2012139067A (en) | Power fluctuation detection device and power fluctuation detection method | |
| Wang et al. | A fast algorithm to estimate phasor in power systems | |
| Kumar et al. | Online identification of underlying causes for multiple and multi-stage power quality disturbances using S-transform | |
| CN108761202A (en) | The harmonic detecting method that pole symmetric mode decomposition and Hilbert transform are combined | |
| Kasztenny | A new method for fast frequency measurement for protection applications |
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: 20190326 |
|
| RJ01 | Rejection of invention patent application after publication |