CN108562811B - Bidirectional long-short term memory-based complex power quality disturbance analysis method - Google Patents
Bidirectional long-short term memory-based complex power quality disturbance analysis method Download PDFInfo
- Publication number
- CN108562811B CN108562811B CN201810198781.8A CN201810198781A CN108562811B CN 108562811 B CN108562811 B CN 108562811B CN 201810198781 A CN201810198781 A CN 201810198781A CN 108562811 B CN108562811 B CN 108562811B
- Authority
- CN
- China
- Prior art keywords
- short term
- term memory
- neural network
- sequence
- data
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Complex Calculations (AREA)
- Measurement Of Current Or Voltage (AREA)
Abstract
The invention discloses a complicated power quality disturbance analysis method based on bidirectional long-term and short-term memory, which specifically comprises the following steps: collecting a plurality of voltage or current signals in a power system to be detected or obtaining a total sample of 7 types of basic PQDs and complex PQDs formed by different combinations of the basic PQDs by using a mathematical model; labeling and converting a sample into a sequence form, and then dividing the sample into a training set and a testing set; constructing a bidirectional long-short term memory neural network model and training; then, performing overfitting judgment, if overfitting occurs, adjusting the hyper-parameters, then retraining, and repeating the steps until overfitting does not occur; and performing PQD judgment by using the retrained neural network model, wherein input data is signal sequence data, and output data is an electric energy type corresponding to each data in the sequence. The invention solves the problems of low identification accuracy, complex realization process, poor real-time performance and incapability of accurately positioning disturbance starting and stopping moments in the prior art.
Description
Technical Field
The invention belongs to the technical field of power quality analysis and detection methods, and relates to a complex power quality disturbance analysis method based on bidirectional long-term and short-term memory.
Background
Accurate identification of Disturbance type is a prerequisite and basis for complex Power Quality Disturbance (PQD) analysis. Power Quality Disturbance (PQD) can be divided into basic PQD and complex PQD. The basic PQD is further divided into steady-state disturbances (mainly including harmonics/inter-harmonics, fluctuations, etc.) and transient disturbances (mainly including dips, interrupts, oscillatory transients, impulsive transients, etc.) according to the temporal characteristics of the disturbances. The complex PQD is formed by compounding a plurality of basic PQDs with different disturbance types, different disturbance intensities and different starting and stopping moments, and particularly relates to disturbance with superimposed transient components.
At present, a shallow learning-based method is mainly used for researching the problem of electric energy quality disturbance type identification, and can be summarized into two links of characteristic quantity extraction and pattern identification, namely, the disturbance signal characteristic quantity is extracted from an original signal after transformation and reconstruction, and then disturbance type identification is performed by adopting shallow models such as a neural network or a support vector machine, and the like, and the following common problems exist in the methods:
(1) most of the complex PQDs formed by basic PQDs or two basic PQDs (most of which are harmonic waves and other basic PQDs) are only considered, but with the increase of basic PQD types contained in the complex PQDs, the phenomenon of cross coupling among characteristic quantities is serious, so that the identification accuracy is greatly reduced or the problem of identifying the disturbance types of the complex PQDs cannot be solved at all;
(2) the method has good effect on steady-state disturbance, but for transient disturbance, the amplitude and frequency characteristics of the transient disturbance cannot be reflected well by simply extracting the amplitude and frequency characteristics, so that the accuracy is low;
(3) the selection of what features and how to select features need to be deeply understood about signal characteristics or tried according to rich engineering practice experience of experts in the field, so that the manual perturbation feature extraction and selection process is complex. In addition, the method is confined to the inherent mode of 'feature extraction before mode identification', the relationship between the two modes is viewed in a cutting way, the identification accuracy rate is seriously dependent on the disturbance feature quantity which is manually designed in advance, the complexity and the redundancy of the disturbance type identification process are increased, the real-time performance of the identification process is difficult to ensure, and the method is not suitable for online application.
Disclosure of Invention
The invention aims to provide a complicated power quality disturbance analysis method based on bidirectional long-short term memory, and solves the problems of low identification accuracy, complex implementation process, poor real-time performance and incapability of accurately positioning the disturbance starting and stopping time in the prior art.
The technical scheme adopted by the invention is that the method for analyzing the complicated power quality disturbance based on the bidirectional long-term and short-term memory is implemented according to the following steps:
step 1, acquiring a plurality of voltage or current signals in a power system to be detected by using a measuring instrument or adopting a mathematical model shown in formula (1) to obtain a series of 7 types of basic PQDs including temporary rising, temporary falling, interruption, oscillation transient, pulse transient, harmonic/inter-harmonic and fluctuation and a complex PQD total sample formed by different combinations of the basic PQDs;
the formula (1) of the unified parameterized analytic mathematical model of the complex PQD signal s (t) including the transient rise, the transient fall, the interruption, the oscillation transient, the pulse transient, the harmonic/inter-harmonic and the fluctuation is shown as the following formula:
wherein, δ (t), A, f0(t)、θ1(t)、N、an(t)、fn(t)、ψn(t) represents the amplitude-dependent disturbance: δ (t) is a time-varying function, A, f0(t)、θ1(t) amplitude, frequency, phase of the fundamental signal, N, an(t)、fn(t)、ψn(t) respectively representing the number of the fluctuation envelope components, and the amplitude, the frequency and the phase of the envelope signal for n times; H. h, bh(t)、θh(t) represents integer order harmonic perturbations related to the fundamental frequency: H. h, bh(t)、θh(t) represents the number of harmonic components, the harmonic frequency, the amplitude and the phase of the h-th harmonic respectively; K. c. Ck(t)、fk(t)、
Represents inter-harmonic disturbances independent of the fundamental frequency: K. c. Ck(t)、fk(t)、
Respectively representing the number of inter-harmonic components, the amplitude, frequency and phase of k-th inter-harmonic, M, αm、βm、dm、fm、θm(t)、τmRepresenting a transient disturbance M, αm、βm、dm、fm、θm(t)、τmRespectively representing the number of oscillation transient components, the starting time, the ending time, the amplitude, the frequency, the phase and the attenuation factor of the m oscillation transients; μ (t) is a noise component;
step 2, labeling samples
Respectively defining and labeling the PQD samples acquired in the step 1 according to disturbance types contained in the signals;
step 3, converting the PQD sample marked in the step 2 into a sequence form;
step 4, dividing the samples serialized in the step 3 into a training set and a testing set, wherein the training set accounts for 70% of the total samples, and the testing set data accounts for 30% of the total samples;
step 5, constructing a bidirectional long-short term memory neural network model, which comprises an input layer part, a hidden layer part and an output layer part which are sequentially connected from bottom to top;
step 6, training the bidirectional long and short term memory neural network model constructed in the step 5, traversing each training data in the training set each time, wherein each traversal is called a generation, and performing a plurality of generations of training on the neural network model, namely obtaining the trained bidirectional long and short term memory neural network model after a plurality of generations;
step 7, overfitting judgment
Using the two-way long and short term memory neural network model trained in the step 6, testing by using 80% of data in the test set to obtain data accuracy, then testing by using the rest 20% of test set data, if the accuracy is greatly reduced, generating an overfitting phenomenon, adjusting the hyper-parameters of the two-way long and short term memory neural network model, then retraining the two-way long and short term memory neural network model, executing overfitting judgment again after training, and repeating the steps until overfitting does not occur, thereby obtaining the two-way long and short term memory neural network model with good generalization;
and 8, performing PQD judgment by using the two-way long and short term neural network model trained again in the step 7, wherein input data are signal sequence data, and output data are the electric energy types corresponding to each data in the sequence.
The present invention is also characterized in that,
the complex PQD in step 1 includes all combinations of two or more basic PQDs.
When the formula (1) is adopted in the step 1, the following principle should be followed when the complex PQD is combined: a. mutation in two different ways cannot occur simultaneously with the same parameter; b. different parameters may mutate simultaneously; c. the presence of additive perturbations is not limited by parameter variations.
The sequence in the step 3 comprises two parts, wherein the first part is a signal sequence, and the sampling interval is set according to the sampling time; the second part is a label sequence, the type of each element is marked by the label sequence, and the label sequence corresponds to the elements in the sampling sequence one by one.
The input layer part in the step 5 only comprises an input layer, the hidden layer part comprises a plurality of hidden layers, the hidden layers comprise a bidirectional long-short term memory layer, a full connection layer and a discarding layer, and the output layer part is a Soft-Max layer.
The input data format of the input layer is [ samples, sequence-length, dim ], wherein samples are input data sample size and are the same as the number of sequences, sequence-length is input sequence length, and dim is data dimension.
The rest of the neural network layers except the first input layer are connected with the previous neural network layer through the activation function.
The invention has the advantages that
(1) The method can directly and automatically learn the disturbed feature information from the original bottom data, can extract the complex implicit features which are difficult to quantify from the simple explicit features, more comprehensively describes the complex PQD signals, ensures the integrity of the complex PQD information to the maximum extent, avoids the complicated manual feature extraction process in the traditional shallow learning, and improves the rapidity and the accuracy of disturbance type identification and the positioning precision of disturbance starting and stopping moments.
(2) The noise is not required to be subjected to any hypothesis or processing in advance, unnecessary errors are avoided, the method has stronger background noise interference resistance, in addition, the organic unification of feature extraction and pattern recognition can be realized, the feature extraction and the pattern recognition are simultaneously carried out and are simultaneously generated in training, the end-to-end online real-time processing can be realized, and the algorithm complexity is effectively reduced.
(3) The bidirectional long and short term memory neural network not only refers to elements before the current element but also refers to elements after the current element in the judging process to realize bidirectional reference, so that the problem of judging distortion in the traditional long and short term memory neural network can be solved, and accurate starting and stopping moments of each basic power quality disturbance are obtained.
Drawings
FIG. 1 is a diagram of an embodiment of a complicated power quality disturbance analysis method based on bidirectional long-short term memory according to the present invention;
FIG. 2 is a data diagram of a disturbance waveform of 0.4-2.6s detected by a noisy analog power quality monitoring terminal according to an embodiment of the present invention;
FIG. 3 is a data diagram of a disturbance waveform of 3.0-5.4s detected by a noise-containing analog power quality monitoring terminal according to an embodiment of the present invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention relates to a complex power quality disturbance analysis method based on bidirectional long-term and short-term memory, which is implemented according to the following steps:
step 1, acquiring a plurality of voltage or current signals in a power system to be detected by using a measuring instrument or obtaining a series of total samples of complex PQDs formed by 7 types of basic PQDs including temporary rise, temporary fall, interruption, oscillation transient, pulse transient, harmonic/inter-harmonic and fluctuation and different combinations thereof by adopting a mathematical model shown in formula (1), wherein the complex PQDs include all combination forms formed by two or more basic PQDs, and the following principle should be followed when the complex PQDs are combined: a. mutation in two different ways cannot occur simultaneously with the same parameter; b. different parameters may mutate simultaneously; c. the existence of additive disturbance is not limited by parameter change;
the formula (1) of the unified parameterized analytic mathematical model of the complex PQD signal s (t) including the transient rise, the transient fall, the interruption, the oscillation transient, the pulse transient, the harmonic/inter-harmonic and the fluctuation is shown as the following formula:
wherein, δ (t), A, f0(t)、θ1(t)、N、an(t)、fn(t)、ψn(t) represents the amplitude-dependent disturbance: δ (t) is a time-varying function, A, f0(t)、θ1(t) amplitude, frequency, phase of the fundamental signal, N, an(t)、fn(t)、ψn(t) respectively representing the number of the fluctuation envelope components, and the amplitude, the frequency and the phase of the envelope signal for n times; H. h, bh(t)、θh(t) represents integer order harmonic perturbations related to the fundamental frequency: H. h, bh(t)、θh(t) represents the number of harmonic components, the harmonic frequency, the amplitude and the phase of the h-th harmonic respectively; K. c. Ck(t)、fk(t)、
Represents inter-harmonic disturbances independent of the fundamental frequency: K. c. Ck(t)、fk(t)、
Respectively representing the number of inter-harmonic components, the amplitude, frequency and phase of k-th inter-harmonic, M, αm、βm、dm、fm、θm(t)、τmRepresenting a transient disturbance M, αm、βm、dm、fm、θm(t)、τmRespectively representing the number of oscillation transient components, the starting time, the ending time, the amplitude, the frequency, the phase and the attenuation factor of the m oscillation transients; μ (t) is a noise component;
step 2, labeling samples
Respectively defining and labeling the PQD samples acquired in the step 1 according to disturbance types contained in the signals;
step 3, converting the PQD sample marked in the step 2 into a sequence form, wherein the sequence comprises two parts, the first part is a signal sequence, and the sampling interval is set according to the sampling time; the second part is a label sequence, the type of each element is marked by the label sequence, and the label sequence corresponds to the elements in the sampling sequence one by one;
step 4, dividing the samples serialized in the step 3 into a training set and a testing set, wherein the training set accounts for 70% of the total samples, and the testing set data accounts for 30% of the total samples;
step 5, constructing a bidirectional long and short term memory neural network model, which comprises an input layer part, a hidden layer part and an output layer part which are sequentially connected from bottom to top, wherein the input layer part only comprises an input layer, the hidden layer part comprises a plurality of hidden layers, the hidden layers comprise a bidirectional long and short term memory layer, a full connection layer and a discarding layer, and the output layer part is a Soft-Max layer; the input data format of the input layer is [ samples, sequence-length, dim ], wherein samples are input data sample size and are the same as the number of sequences, sequence-length is input sequence length, and dim is data dimension; the rest of the neural network layers except the first input layer are connected with the previous neural network layer through the activation function.
Step 6, training the bidirectional long and short term memory neural network model constructed in the step 5, traversing each training data in the training set each time, wherein each traversal is called a generation, and performing a plurality of generations of training on the neural network model, namely obtaining the trained bidirectional long and short term memory neural network model after a plurality of generations;
step 7, overfitting judgment
Using the two-way long and short term memory neural network model trained in the step 6, testing by using 80% of data in the test set to obtain data accuracy, then testing by using the rest 20% of test set data, if the accuracy is greatly reduced, generating an overfitting phenomenon, adjusting the hyper-parameters of the two-way long and short term memory neural network model, then retraining the two-way long and short term memory neural network model, executing overfitting judgment again after training, and repeating the steps until overfitting does not occur, thereby obtaining the two-way long and short term memory neural network model with good generalization;
and 8, performing PQD judgment by using the two-way long and short term neural network model trained again in the step 7, wherein input data are signal sequence data, and output data are the electric energy types corresponding to each data in the sequence.
Examples
The method for analyzing the complicated power quality disturbance based on the bidirectional long-short term memory is implemented according to the following steps:
step 1, acquiring a plurality of voltage or current signals in a power system to be detected by using a measuring instrument or obtaining a series of total samples of complex PQDs formed by 7 types of basic PQDs including temporary rise, temporary fall, interruption, oscillation transient, pulse transient, harmonic/inter-harmonic and fluctuation and different combinations thereof by adopting a mathematical model shown in a formula (1), wherein the complex PQDs include all combination forms formed by two or more basic PQDs, and the following principle is followed when the complex PQDs are combined by adopting the formula (1) in the step 1: a. mutation in two different ways cannot occur simultaneously with the same parameter; b. different parameters may mutate simultaneously; c. the existence of additive disturbance is not limited by parameter change;
the formula (1) of the unified parameterized analytic mathematical model of the complex PQD signal s (t) including the transient rise, the transient fall, the interruption, the oscillation transient, the pulse transient, the harmonic/inter-harmonic and the fluctuation is shown as the following formula:
wherein, δ (t), A, f0(t)、θ1(t)、N、an(t)、fn(t)、ψn(t) represents the amplitude-dependent disturbance: δ (t) is a time-varying function, A, f0(t)、θ1(t) amplitude (V), frequency (Hz), phase (rad/s) of the fundamental wave signal, N, an(t)、fn(t)、ψn(t) respectively representing the number of fluctuation envelope components, the amplitude (V), the frequency (Hz) and the phase (rad/s) of the envelope signals for n times; H. h, bh(t)、θh(t) represents integer order harmonic perturbations related to the fundamental frequency: H. h, bh(t)、θh(t) represents the number of harmonic components, the harmonic frequency, the amplitude (V) of the h-th harmonic and the phase (rad/s) respectively; K. c. Ck(t)、fk(t)、
Represents inter-harmonic disturbances independent of the fundamental frequency: K. c. Ck(t)、fk(t)、
Respectively representing the number of inter-harmonic components, the amplitude (V), frequency (Hz), and phase (rad/s) of k-th inter-harmonic, M, αm、βm、dm、fm、θm(t)、τmRepresenting transient disturbances, such as oscillatory transients, impulsive transients, different disturbance parameter ranges may represent different disturbance types M, αm、βm、dm、fm、θm(t)、τmRespectively representing the number of the oscillation transient components, the starting time(s), the ending time(s), the amplitude (V), the frequency (Hz), the phase (rad/s) and the attenuation factor(s) of the m oscillation transients; μ (t) is a noise component;
step 2, labeling samples
Respectively defining and labeling the PQD samples acquired in the step 1 according to disturbance types contained in the signals;
step 3, converting the PQD sample marked in the step 2 into a sequence form, wherein the sequence comprises two parts, the first part is a signal sequence, and the sampling interval is set according to the sampling time; the second part is a label sequence, the type of each element is marked by the label sequence, and the label sequence corresponds to the elements in the sampling sequence one by one; in this example, the PQD signal per second contains 1000 samples, i.e., each signal sequence contains 1000 elements; the label sequence marks the type of each element, and the label sequence elements correspond to the sampling sequence elements one to one.
Step 4, dividing the samples serialized in the step 3 into a training set and a testing set, wherein the training set accounts for 70% of the total samples, and the testing set data accounts for 30% of the total samples;
step 5, constructing a bidirectional long and short term memory neural network model, which comprises an input layer part, a hidden layer part and an output layer part which are sequentially connected from bottom to top, wherein the input layer part only comprises an input layer, the hidden layer part comprises a plurality of hidden layers, the hidden layers comprise a bidirectional long and short term memory layer, a full connection layer and a discard layer (dropout layer), and the output layer part is a Soft-Max layer; outputting a judgment result of the sequence through the last Soft-Max layer by the trained bidirectional long-short term memory neural network; the rest of the neural network layers except the first input layer are connected with the previous neural network layer through the activation function. For example: as shown in fig. 1, one input layer is used, eight hidden layers are used, and the hidden layers are sequentially a bidirectional long and short term memory layer, a discarding layer, a bidirectional long and short term memory layer, and a full connection layer; then an output layer part is provided, and the part is a Soft-Max layer; the input data format of the input layer is [ samples, sequence-length, dim ], wherein samples are input data sample size and are the same as the number of sequences, sequence-length is input sequence length, and dim is data dimension; in this example, samples is set to the total number of sequences, sequence-length is 1000, dim is set to 1, i.e. the input data has only one dimension. The output sequence is in the same form as the input sequence. The output content is a type tag corresponding to the input signal sequence. In the hidden layer, the bidirectional long-short term memory neural network unit layer is set to 500 hidden units, and the n-steps size of the bidirectional long-short term memory layer is the same as the sequence-length size of input data, and 1000 is used here. The discard layer discard rate is 0.3 and the full connection layer is 1000 units. The output layer is a Soft-max layer which is connected with the full connecting layer of the last layer. The activation function uses the Relu activation function. The output data of each layer is normalized using batch normalization.
Step 6, training the bidirectional long and short term memory neural network model constructed in the step 5, traversing each training data in the training set each time, wherein each traversal is called a generation, and performing a plurality of generations of training on the neural network model, namely obtaining the trained bidirectional long and short term memory neural network model after a plurality of generations; in this example, a global random initialization mode is used for parameter initialization, adam optimizer training or Momentum optimizer or SGD optimizer is used, 5000 generations are trained, the learning rate is 0.01, the loss function uses cross entropy as the loss function or uses the mean square error or the mean difference of the cross entropy loss function as the loss function.
Step 7, overfitting judgment
Using the two-way long and short term memory neural network model trained in the step 6, testing by using 80% of data in the test set to obtain data accuracy, then testing by using the rest 20% of test set data, if the accuracy is greatly reduced, generating an overfitting phenomenon, adjusting the hyper-parameters of the two-way long and short term memory neural network model, then retraining the two-way long and short term memory neural network model, executing overfitting judgment again after training, and repeating the steps until overfitting does not occur, thereby obtaining the two-way long and short term memory neural network model with good generalization; if overfitting occurs, adjusting the hyper-parameters, such as modifying the discarding rate of the discarding layer, modifying the number of the full-connection layers, modifying the learning rate, modifying the training generation, adjusting the number of the hidden layers (such as reducing the bidirectional long-short term memory layer),
step 8, using the bidirectional long and short term neural network model trained again in the step 7 to judge PQD, wherein input data are signal sequence data, and output data are electric energy types corresponding to each data in the sequence; at this time, the sequence-length of the input data is changed to 1, that is, the input data is signal sequence data one by one, and the output data is the electric energy type corresponding to each data in the sequence.
The number of the hidden layers in the neural network model can be modified, and the number of the bidirectional long-short term memory layer, the discarding layer, the full connection layer and the modified full connection layer units can be increased or reduced.
In the invention, the rest of the neural network layers except the first input layer are connected with the previous neural network layer through an activation function, wherein the activation function can be ReLU or Leaky Relu or Sigmoid or tanh as the activation function.
The neural network hyper-parameters in the invention can be adjusted according to the actual conditions, such as the number of training generations, the size of sequence n-steps, the learning rate, the length of input sequence, the dimension of input sequence and the size of sequence window in the bidirectional long and short term memory layer.
Simulation comparison is given below to prove the detection effect of the method provided by the invention:
according to the model shown in formula (1), all parameters in the above formula are randomly changed within a range, 300 data samples are generated for each type of disturbance signals, 200 of the data samples are used as training samples, 100 of the data samples are used as testing samples, and white Gaussian noise with the signal-to-noise ratio of 20dB is superposed.
Table 1 shows the results of adaptive identification of disturbance types of 48 PQD signals based on a bidirectional long-short time memory neural network under two working conditions of no noise and noise with a superimposed signal-to-noise ratio of 20 dB;
TABLE 1 disturbance type identification of complex PQD in noiseless and noisy environments
Table 1 has 5 columns of data, the first column shows the identification accuracy of 7 basic PQDs, the second column shows the identification accuracy of complex PQD composed of 2 basic PQDs, one disturbance is from the basic PQD shown in column 1, the other disturbance is from the basic PQD shown in column 2, and by analogy, columns 3 to 5 respectively show the identification accuracy of complex PQD disturbance types composed of 3 to 5 basic PQDs. For example, the fifth column, first row plus bold identification data fluctuation (98, 97) in the table can be understood as: the identification accuracy of the complex PQD composed of 5 basic PQDs (sag, harmonic/inter-harmonic, oscillation transient, pulse and fluctuation) is 98% under the noiseless working condition, and the identification accuracy of the disturbance type under the working condition that the signal-to-noise ratio is 20dB is 97%.
Looking at table 1, the following conclusions can be drawn: 1) the method overcomes the defects of the traditional shallow learning in feature extraction and disturbance type identification, the influence of the number of basic PQD types contained in the complex PQD on the disturbance identification accuracy is small, the identification accuracy is greatly improved, and the complex PQD disturbance type identification accuracy is higher than 99% under the noiseless condition. 2) When background noise with the signal-to-noise ratio of 20dB is superposed, the accuracy rate of complex PQD disturbance type identification is higher than 97%, so that the disturbance type identification method has strong anti-noise capability and does not need to carry out noise elimination processing on original signals.
In addition, it is particularly pointed out that the trained bidirectional long-short term neural network model also has the advantage of high recognition speed, and can meet the real-time requirement of an actual power quality online monitoring system, and the requirement is often difficult to meet in a complex PQD disturbance type recognition method based on traditional shallow learning.
As shown in fig. 2-3, for disturbance waveform data detected by a noise-containing simulated power quality monitoring terminal, the start-stop times of all basic events included in the disturbance waveform data are positioned by using a bidirectional long-short term neural network model, and the obtained detection results are shown in table 2.
TABLE 2 detection results of disturbance start-stop time
Claims (7)
1. The method for analyzing the complicated power quality disturbance based on the two-way long-short term memory is characterized by comprising the following steps:
step 1, acquiring a plurality of voltage or current signals in a power system to be detected by using a measuring instrument or adopting a mathematical model of a formula (1) to obtain a series of 7 types of basic PQDs comprising a temporary rising, a temporary falling, an interruption, an oscillation transient state, a pulse transient state, a harmonic/inter-harmonic wave and a fluctuation and a complex PQD total sample formed by different combinations of the basic PQDs;
the formula (1) of the unified parameterized analytic mathematical model of the complex PQD signal s (t) including the transient rise, the transient fall, the interruption, the oscillation transient, the pulse transient, the harmonic/inter-harmonic and the fluctuation is shown as the following formula:
wherein, δ (t), A, f0(t)、θ1(t)、N、an(t)、fn(t)、ψn(t) represents the amplitude-dependent disturbance: δ (t) is a time-varying function, A, f0(t)、θ1(t) amplitude, frequency, phase of the fundamental signal, N, an(t)、fn(t)、ψn(t) respectively representing the number of the fluctuation envelope components, and the amplitude, the frequency and the phase of the envelope signal for n times; H. h, bh(t)、θh(t) represents integer order harmonic perturbations related to the fundamental frequency: H. h, bh(t)、θh(t) represents the number of harmonic components, the harmonic frequency, the amplitude and the phase of the h-th harmonic respectively; K. c. Ck(t)、fk(t)、
Represents inter-harmonic disturbances independent of the fundamental frequency: K. c. Ck(t)、fk(t)、
Respectively representing the number of inter-harmonic components, the amplitude, frequency and phase of k-th inter-harmonic, M, αm、βm、dm、fm、
τmRepresenting a transient disturbance M, αm、βm、dm、fm、
τmRespectively representing the number of oscillation transient components, the starting time and the end of the m oscillation transientsDead time, amplitude, frequency, phase, attenuation factor; μ (t) is a noise component;
step 2, labeling samples
Respectively defining and labeling the PQD samples acquired in the step 1 according to disturbance types contained in the signals;
step 3, converting the PQD sample marked in the step 2 into a sequence form;
step 4, dividing the samples serialized in the step 3 into a training set and a testing set, wherein the training set accounts for 70% of the total samples, and the testing set data accounts for 30% of the total samples;
step 5, constructing a bidirectional long-short term memory neural network model, which comprises an input layer part, a hidden layer part and an output layer part which are sequentially connected from bottom to top;
step 6, training the bidirectional long and short term memory neural network model constructed in the step 5, traversing each training data in the training set each time, wherein each traversal is called a generation, and performing a plurality of generations of training on the neural network model, namely obtaining the trained bidirectional long and short term memory neural network model after a plurality of generations;
step 7, overfitting judgment
Using the two-way long and short term memory neural network model trained in the step 6, testing by using 80% of data in the test set to obtain data accuracy, then testing by using the rest 20% of test set data, if the accuracy is greatly reduced, generating an overfitting phenomenon, adjusting the hyper-parameters of the two-way long and short term memory neural network model, then retraining the two-way long and short term memory neural network model, executing overfitting judgment again after training, and repeating the steps until overfitting does not occur, thereby obtaining the two-way long and short term memory neural network model with good generalization;
and 8, performing PQD judgment by using the two-way long and short term neural network model trained again in the step 7, wherein input data are signal sequence data, and output data are the electric energy types corresponding to each data in the sequence.
2. The method of claim 1 wherein the complex PQD in step 1 comprises all combinations of two or more basic PQDs.
3. The method for analyzing the disturbance of the quality of the complex electric energy based on the two-way long-short term memory as claimed in claim 1, wherein the formula (1) adopted in the step 1 is used for the complex PQD combination according to the following principle: a. mutation in two different ways cannot occur simultaneously with the same parameter; b. different parameters may mutate simultaneously; c. the presence of additive perturbations is not limited by parameter variations.
4. The method for analyzing the quality disturbance of the complex electric energy based on the two-way long and short term memory as claimed in claim 1, wherein the sequence in the step 3 comprises two parts, the first part is a signal sequence, and the sampling interval is set according to the sampling time; the second part is a label sequence, the type of each element is marked by the label sequence, and the label sequence corresponds to the elements in the sampling sequence one by one.
5. The method for analyzing the quality disturbance of the complex electric energy based on the two-way long and short term memory according to claim 1, wherein the input layer part in the step 5 only comprises an input layer, the hidden layer part comprises a plurality of hidden layers which are sequentially connected, the hidden layers comprise a two-way long and short term memory layer, a full connection layer and a discarding layer, and the output layer part is a Soft-Max layer.
6. The method according to claim 5, wherein the input layer input data format is [ samples, sequence-length, dim ], where samples is the input data sample size and is the same as the number of sequences, sequence-length is the input sequence length, and dim is the data dimension.
7. The method for analyzing the disturbance of the quality of the complex electric energy based on the two-way long and short term memory as claimed in claim 5, wherein the neural network layers except the first input layer are connected with the neural network layer of the previous layer through activation functions.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810198781.8A CN108562811B (en) | 2018-03-12 | 2018-03-12 | Bidirectional long-short term memory-based complex power quality disturbance analysis method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810198781.8A CN108562811B (en) | 2018-03-12 | 2018-03-12 | Bidirectional long-short term memory-based complex power quality disturbance analysis method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108562811A CN108562811A (en) | 2018-09-21 |
| CN108562811B true CN108562811B (en) | 2020-05-22 |
Family
ID=63531608
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810198781.8A Active CN108562811B (en) | 2018-03-12 | 2018-03-12 | Bidirectional long-short term memory-based complex power quality disturbance analysis method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108562811B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109214107B (en) * | 2018-09-26 | 2020-05-05 | 大连海事大学 | Ship navigation behavior online prediction method |
| CN109508811A (en) * | 2018-09-30 | 2019-03-22 | 中冶华天工程技术有限公司 | Parameter prediction method is discharged based on principal component analysis and the sewage treatment of shot and long term memory network |
| CN110222953A (en) * | 2018-12-29 | 2019-09-10 | 北京理工大学 | A kind of power quality hybrid perturbation analysis method based on deep learning |
| CN110728195B (en) * | 2019-09-18 | 2022-04-01 | 武汉大学 | Power quality disturbance detection method based on YOLO algorithm |
| CN110829417B (en) * | 2019-11-14 | 2023-03-31 | 电子科技大学 | Electric power system transient stability prediction method based on LSTM double-structure model |
| WO2021129668A1 (en) * | 2019-12-24 | 2021-07-01 | 华为技术有限公司 | Neural network training method and device |
| CN111781244B (en) * | 2020-07-15 | 2021-10-26 | 中国科学院自动化研究所 | Infrared thermal imaging type coating detection method based on long-term and short-term memory network |
| CN113536655B (en) * | 2021-04-07 | 2022-07-22 | 北京聚树核科技有限公司 | Heliostat artificial intelligence deviation rectifying method, heliostat artificial intelligence deviation rectifying device, electronic equipment and storage medium |
| CN114992138B (en) * | 2022-06-13 | 2024-03-26 | 湘潭大学 | GAN-CNN-BiLSTM-based industrial circulating water system water supply pump fault diagnosis method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101034389A (en) * | 2007-03-19 | 2007-09-12 | 江西省电力科学研究院 | Electrical energy power quality disturbance automatic identification method and system based on information fusion |
| CN105205560A (en) * | 2015-09-25 | 2015-12-30 | 国网甘肃省电力公司电力科学研究院 | Photovoltaic power supply power prediction method based on positive and negative error variable weights |
| CN106778923A (en) * | 2017-03-23 | 2017-05-31 | 广东电网有限责任公司珠海供电局 | A kind of Power Quality Disturbance sorting technique and device |
| CN107123987A (en) * | 2017-05-11 | 2017-09-01 | 华东交通大学 | Electrical energy power quality disturbance recognition methods based on on-line training weighed SVM |
| CN107679531A (en) * | 2017-06-23 | 2018-02-09 | 平安科技（深圳）有限公司 | Licence plate recognition method, device, equipment and storage medium based on deep learning |
-
2018
- 2018-03-12 CN CN201810198781.8A patent/CN108562811B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101034389A (en) * | 2007-03-19 | 2007-09-12 | 江西省电力科学研究院 | Electrical energy power quality disturbance automatic identification method and system based on information fusion |
| CN105205560A (en) * | 2015-09-25 | 2015-12-30 | 国网甘肃省电力公司电力科学研究院 | Photovoltaic power supply power prediction method based on positive and negative error variable weights |
| CN106778923A (en) * | 2017-03-23 | 2017-05-31 | 广东电网有限责任公司珠海供电局 | A kind of Power Quality Disturbance sorting technique and device |
| CN107123987A (en) * | 2017-05-11 | 2017-09-01 | 华东交通大学 | Electrical energy power quality disturbance recognition methods based on on-line training weighed SVM |
| CN107679531A (en) * | 2017-06-23 | 2018-02-09 | 平安科技（深圳）有限公司 | Licence plate recognition method, device, equipment and storage medium based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 基于改进型神经网络的电能质量扰动识别;金燕等;《电力系统及其自动化》;20121231;第34卷(第4期);第59-61页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108562811A (en) | 2018-09-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108562811B (en) | Bidirectional long-short term memory-based complex power quality disturbance analysis method | |
| CN109946389B (en) | Structural damage identification method based on ensemble empirical mode decomposition and convolutional neural network | |
| CN110261109B (en) | Rolling bearing fault diagnosis method based on bidirectional memory cyclic neural network | |
| CN108921285B (en) | Bidirectional gate control cyclic neural network-based classification method for power quality disturbance | |
| CN109245099A (en) | A kind of electric load discrimination method, device, equipment and readable storage medium storing program for executing | |
| CN109360581A (en) | Sound enhancement method, readable storage medium storing program for executing and terminal device neural network based | |
| CN110070102B (en) | Method for establishing sequence-to-sequence model for identifying power quality disturbance type | |
| CN111914883A (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| CN111368892A (en) | Generalized S transformation and SVM electric energy quality disturbance efficient identification method | |
| CN112329914B (en) | Fault diagnosis method and device for buried transformer substation and electronic equipment | |
| CN113642484B (en) | Magnetotelluric signal noise suppression method and system based on BP neural network | |
| CN111291918B (en) | Rotating machine degradation trend prediction method based on stationary subspace exogenous vector autoregression | |
| CN105787219A (en) | Method for building conducted interference coupling channel multiple linear regression model by near frequency point sampling | |
| Xu et al. | An adaptive spectrum segmentation method to optimize empirical wavelet transform for rolling bearings fault diagnosis | |
| CN111983569B (en) | Radar interference suppression method based on neural network | |
| CN111665050B (en) | Rolling bearing fault diagnosis method based on clustering K-SVD algorithm | |
| CN113177558A (en) | Radiation source individual identification method based on feature fusion of small samples | |
| Zhu et al. | Complex disturbances identification: A novel PQDs decomposition and modeling method | |
| CN113884844A (en) | Transformer partial discharge type identification method and system | |
| CN114700587A (en) | Missing welding defect real-time detection method and system based on fuzzy reasoning and edge calculation | |
| CN110222390B (en) | Gear crack identification method based on wavelet neural network | |
| CN107894837A (en) | Dynamic sentiment analysis model sample processing method and processing device | |
| CN116881712A (en) | Electromagnetic pulse signal identification method for movable cracks of concrete dam | |
| CN116399592A (en) | Bearing fault diagnosis method based on channel attention dual-path feature extraction | |
| Wang et al. | Research on fault diagnosis method of wind turbine bearing based on deep belief 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 |