CN110210152B - Ultrahigh-order harmonic source modeling method - Google Patents
Ultrahigh-order harmonic source modeling method Download PDFInfo
- Publication number
- CN110210152B CN110210152B CN201910494902.8A CN201910494902A CN110210152B CN 110210152 B CN110210152 B CN 110210152B CN 201910494902 A CN201910494902 A CN 201910494902A CN 110210152 B CN110210152 B CN 110210152B
- Authority
- CN
- China
- Prior art keywords
- output
- ultra
- value
- layer
- neuron
- 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
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/30—Circuit design
- G06F30/36—Circuit design at the analogue level
- G06F30/367—Design verification, e.g. using simulation, simulation program with integrated circuit emphasis [SPICE], direct methods or relaxation methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Computer Hardware Design (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Geometry (AREA)
- Supply And Distribution Of Alternating Current (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to an ultrahigh-order harmonic source modeling method, which comprises the following steps of S1: providing ultra-high harmonic current time series data under different powers in any time period; step S2: training the data of the step S1 by using a neural network algorithm; step S3: generating prediction data of each ultrahigh harmonic current amplitude under different powers by using a neural network algorithm; step S4: error calculation is performed on the predicted data and the measured value generated in step S3, and a performance evaluation coefficient R with the smallest error is selected according to the error calculation result2Selecting R at a given calculation accuracy epsilon2Value fitting curve is used as fitting result; step S5: and (5) analyzing and judging the final fitting result of the step (S4), and if the error between the obtained predicted value and the actually measured value is within the range, obtaining a trained neural network model. The invention can predict the output ultra-high harmonic current characteristic of any one ultra-high harmonic source so as to take necessary treatment measures in advance or filter by a more appropriate filter.
Description
Technical Field
The invention relates to the field of power electronics, in particular to an ultrahigh harmonic source modeling method.
Background
With the rapid development of modern industries, transportation, finance, information and other industries and the implementation of energy saving, consumption reduction and environmental protection policies, clean and efficient energy utilization becomes the mainstream. The high density access of a large amount of distributed energy sources and the wide application of flexible power transmission technology enable the power distribution network to show a 'source-network-load' tight coupling characteristic. When the power system supplies power to the nonlinear equipment and the loads, the nonlinear equipment and the loads transmit, convert and absorb fundamental wave energy supplied by the system generator, convert part of the fundamental wave energy into harmonic wave energy, and return the harmonic wave energy to the power system to become a main ultrahigh-order harmonic wave source of the power grid.
At present, with the permeability of electrical equipment such as wind power converters, photovoltaic inverters, electric vehicle charging piles and the like in a power grid becoming higher and higher, the generated ultra-high harmonics cause more and more power quality problems, such as intermittent operation or function failure of the equipment, incapability of operating or damage of the equipment, power carrier communication failure, noise emission of the equipment or devices and the like, and the damage becomes more and more serious undoubtedly in the future. Therefore, the propagation characteristic of the ultra-high harmonic in the power grid is accurately estimated, the actual condition of the ultra-high harmonic in the power grid is accurately mastered, and the method is necessary for preventing the harm of the ultra-high harmonic and maintaining the safe operation of the power grid.
The most common harmonic source modeling methods at present mainly include the following methods: (1) constant current source model and norton equivalent model: although the model is easy to process, the model is too simple and rough and is easy to cause large errors. And the atypical running condition is not considered, so that the accuracy and the application range of the model are reduced. (2) Simplified model based on cross-frequency admittance matrix: although the model considers the mechanism of harmonic generation more, no theoretical basis proves that the characteristics of the harmonic source can be linearized at a certain operating point, and the model only corresponds to a certain operating point and has different harmonic source models for different operating points. (3) A harmonic current expression model is established for a topological circuit structure of a common power electronic device, and the amplitude and the phase angle of each harmonic current generated on an alternating current side are obtained through Fourier analysis. The premise of the above-mentioned modeling is that, assuming that the structure of the harmonic source is unchanged, when the structure is slightly changed, the relationship changes, and the analytical modeling must be performed again.
Disclosure of Invention
In view of this, the present invention provides an ultrahigh-order harmonic source modeling method, which can predict the output ultrahigh-order harmonic current characteristic of any one ultrahigh-order harmonic source, so as to take necessary treatment measures in advance or select a more appropriate filter for filtering.
The invention is realized by adopting the following scheme: an ultrahigh-order harmonic source modeling method comprises the following steps:
step S1: providing ultra-high harmonic current time series data under different powers in any time period;
step S2: training the ultra-high harmonic current time series data under different powers by using a neural network algorithm;
step S3: generating prediction data of each ultrahigh harmonic current amplitude under different powers by using a neural network algorithm;
step S4: performing error calculation on the prediction data and the actual measurement value generated in step S3, namely the super-harmonic current time-series data under different powers mentioned in step S1, and selecting a performance evaluation coefficient R with the minimum error according to the error calculation result2Selecting R at a given calculation accuracy epsilon2The value fitting curve is used as a fitting result to obtain the output characteristic of the ultrahigh harmonic source;
step S5: and analyzing and judging the final fitting result of the step S4, if the error between the obtained predicted value and the actually measured value is within 0.0001-0.000001, obtaining a trained neural network model, inputting the power value and the frequency of the ultrahigh harmonic source into the trained neural network model, and outputting the predicted value of the output current amplitude of the ultrahigh harmonic source with the required frequency under the power by the model.
Further, the step S2 specifically includes the following steps:
step S21: providing ultrahigh harmonic current value data corresponding to different frequencies under different powers, comprising: input vector of input layer, hidden layer input vectorHidden layer output vectorOutput layer input vectorAnd an output vector of the output layer;
wherein the input vector of the input layer includes the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd the amplitude of the current actually measured by the ultra-high order harmonic source
The output vector of the output layer includes the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd ultrahigh harmonic source predicted current amplitude
Establishing a neural network model, wherein an input layer has 2 nodes, an output layer has 1 node, and an empirical formula is utilized: s is 2n +1, wherein n is the number of nodes of the input layer, and the number of nodes of the hidden layer is calculated to be 5; the number of the nodes is gradually increased on the hidden layer of 5 nodes, when the number of the nodes is increased to 10, the network error of the node number is increased and is not reduced, and the node selection number is optimal.
Step S22: connection weight omega for input layer and hidden layerihThe connection weight omega of the hidden layer and the output layerhoThreshold b of each neuron in the hidden layerhAnd threshold b of each neuron of output layeroRespectively make one [ -11 ]]Random number inside, and let the error function be:setting the range of the calculation precision value epsilon to be 0.0001-0.000001 as a training termination condition, and setting the range of the maximum learning times M to be 1000-; the learning rate is selected in the range of 0.01-0.8;
step S23: normalizing the ultrahigh harmonic current value data corresponding to different frequencies under different powers in the step S21 to [01] by adopting an S-shaped activation function;
step S24: randomly selecting a group of input samples, i.e. the data x (k) of the ultra-high harmonic current values corresponding to different frequencies at different powers in step S21 and the corresponding expected outputs d (k);
wherein the ultra high harmonic current value data x (k) includes power of an ultra high harmonic sourceUltra-high order harmonic frequencyAnd the amplitude of the current actually measured by the ultra-high order harmonic source
Step S25: computing input hi for each neuron in the hidden layerh(k) Then using the input and the laser
Live function calculation of output ho of each neuron of hidden layerh(k):
In the formula, ωihWeights, x, for input layer neurons pointing to hidden layer neuronsi(k) For data on the ith neuron in the kth set of data, bhA threshold value for each neuron of the hidden layer;
input hi of neurons in the hidden layerh(k) Passing through Sigmoid type functions:
hoh(k)=f(hih(k))
Step S26: calculating input yi of each neuron of output layero(k) Then using input and activation function to calculate output yo of each neuron in output layero(k):
In the formula, ωhoWeights, ho, for neurons of the hidden layer to neurons of the output layerh(k) For the output of neurons of the hidden layer, boIs the threshold value of each neuron in the output layer;
input yi of each neuron of output layero(k) Passing through Sigmoid type functions:the mapping of (2) obtains the output yo of each neuron of the output layero(k)
yoo(k)=f(yio(k))
Step S27: calculating partial derivative delta of error function to each neuron of output layero(k):
δo(k)=(do(k)-yoo(k))yoo(k)(1-yoo(k))
Step S28: using the connection weight omega from the hidden layer to the output layerho(k) Partial derivative delta of error function to each neuron of output layero(k) And output ho of hidden layerh(k) Calculating partial derivative delta of error function to each neuron of hidden layerh(k):
Step S29: using error functions for neurons of the output layerPartial derivative deltao(k) And output ho of neurons of the hidden layerh(k) To correct the connection weight omegaho(k) And a threshold value bo(k):
In order to adjust the connection weight value after the adjustment,in order to adjust the connection weight before the adjustment,in order to achieve the adjusted threshold value, the threshold value is,the threshold before adjustment is adopted, eta is the learning rate and takes a value between (0, 1);
step S210: partial derivative delta of each neuron in hidden layer by using error functionh(k) And input x of each neuron of input layeri(k) Modifying the connection weight and the threshold value;
in order to adjust the connection weight value after the adjustment,in order to adjust the connection weight before the adjustment,in order to achieve the adjusted threshold value, the threshold value is,the threshold before adjustment is adopted, eta is the learning rate and takes a value between (0, 1);
step S211: according to the output vector of the output layer, i.e. including the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd ultrahigh harmonic source predicted current amplitudeCorresponding desired output do(k) Calculating a global error E:
step S212: judging whether the network error meets the requirement, and finishing the algorithm when E is less than epsilon or the learning times is more than the set maximum times M; otherwise, the next learning sample and the corresponding expected output are randomly selected, and the process returns to step S24 to enter the next learning process.
Further, the step S4 specifically includes the following steps:
step S41: if the algorithm is finished, performing inverse normalization on output data of the neural network algorithm result to obtain a real value;
step S42: evaluating the predicted value of the current amplitude of the ultrahigh harmonic source according to the performance evaluation coefficient R of the neural network2:
The current amplitude prediction value of the ultra-high order harmonic wave of the ith data is obtained; y isiThe current amplitude real value of the ultrahigh harmonic wave of the ith data; n is the number of data to select a final fitting result curve so as to obtain the output characteristic of the ultrahigh harmonic source; evaluation of coefficient R for neural network Performance2Of selected value of R2The closer the value is to 1, the better the training effect is, the better the model fitting is, and the well-trained neural network model can be obtained; so that R under a given calculation accuracy epsilon is selected2And taking a fitting curve of the values as a fitting result, and further obtaining the output characteristic of the ultrahigh harmonic source.
Further, the step S5 specifically includes the following steps:
step S51: the power of the ultra-high harmonic source to the input data of the neural network model trained in step S4And ultra-high harmonic frequency
Step S52, the trained neural network model outputs dataThat is, the predicted current amplitude of the specific frequency input in step S51 at the power specified in step S51 by the ultra-high harmonic source;
step S53: and obtaining the output characteristic of the predicted ultra-high harmonic source.
Compared with the prior art, the invention has the following beneficial effects:
(1) the invention can predict the output ultra-high harmonic current characteristic of any one ultra-high harmonic source so as to take necessary treatment measures in advance or select a more appropriate filter for filtering.
(2) The method realizes the analysis of harmonic sources with complex topological structures and variable operating conditions, has the advantages of short training time, high precision and dynamic modeling, and is an effective method for modeling the harmonic sources.
Drawings
Fig. 1 is a frequency spectrum diagram of ultrahigh harmonic current output by an electric vehicle charging pile under different powers according to an embodiment of the invention.
Fig. 2 is a frequency spectrum diagram of an ultra-high harmonic current output by an electric vehicle charging pile under a certain power according to an embodiment of the invention.
FIG. 3 is a flowchart illustrating ultra-high harmonic source characteristic estimation according to an embodiment of the invention.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
As shown in fig. 3, the present embodiment provides an ultrahigh-order harmonic source modeling method, including the following steps:
step S1: providing ultra-high harmonic current time sequence data under different powers which can represent the operating conditions of the ultra-high harmonic source within any time period;
step S2: training the ultra-high harmonic current time series data under different powers by using a neural network algorithm;
step S3: generating prediction data of each ultrahigh harmonic current amplitude under different powers by using a neural network algorithm;
step S4: performing error calculation on the predicted data and the measured value generated in step S3, i.e. the super-harmonic current time-series data (or simulated value) under different powers mentioned in step S1, and selecting the performance evaluation coefficient R with the minimum error according to the error calculation result2Selecting R at a given calculation accuracy epsilon2The value fitting curve is used as a fitting result to obtain the output characteristic of the ultrahigh harmonic source; coefficient of performance evaluation R2Value R2The closer the value is to 1, the more variables in the equation in step S42 are indicated(prediction value of ultrahigh harmonic current amplitude of ith data) to yiThe stronger the interpretation ability of the (true value of the ultrahigh harmonic current amplitude of the ith data), the better the training effect, and the better the model fits the data;
step S5: and analyzing and judging the final fitting result of the step S4, if the error between the obtained predicted value and the actually measured value is within 0.0001-0.000001, obtaining a trained neural network model, inputting the power value and the frequency of the ultrahigh harmonic source into the trained neural network model, and outputting the predicted value of the output current amplitude of the ultrahigh harmonic source with the required frequency under the power by the model.
In this embodiment, the step S2 specifically includes the following steps:
step S21: providing ultrahigh harmonic current value data corresponding to different frequencies under different powers, comprising: input vector of input layer, hidden layer input vectorHidden layer output vectorOutput layer input vectorAnd an output vector of the output layer;
wherein the input vector of the input layer includes the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd the amplitude of the current actually measured by the ultra-high order harmonic source
The output vector of the output layer includes the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd ultrahigh harmonic source predicted current amplitude
Establishing a neural network model, wherein an input layer has 2 nodes, an output layer has 1 node, and an empirical formula is utilized: s is 2n +1, wherein n is the number of nodes of the input layer, and the number of nodes of the hidden layer is calculated to be 5; the number of the nodes is gradually increased on the hidden layer of 5 nodes, when the number of the nodes is increased to 10, the network error of the node number is increased, the network error is not obviously reduced, and the node selection number is optimal.
Step S22: connection weight omega for input layer and hidden layerihThe connection weight omega of the hidden layer and the output layerhoThreshold b of each neuron in the hidden layerhAnd threshold b of each neuron of output layeroRespectively make one [ -11 ]]Random number inside, and let the error function be:setting the range of the calculation precision value epsilon to be 0.0001-0.000001 as a training termination condition, and setting the range of the maximum learning times M to be 1000-; the learning rate is selected in the range of 0.01-0.8;
step S23: normalizing the ultrahigh harmonic current value data corresponding to different frequencies under different powers in the step S21 to [01] by adopting an S-shaped activation function;
step S24: randomly selecting a group of input samples, i.e. the data x (k) of the ultra-high harmonic current values corresponding to different frequencies at different powers in step S21 and the corresponding expected outputs d (k);
wherein the ultra high harmonic current value data x (k) includes power of an ultra high harmonic sourceUltra-high order harmonic frequencyAnd the amplitude of the current actually measured by the ultra-high order harmonic source
Step S25: computing input hi for each neuron in the hidden layerh(k) Then, the output ho of each neuron in the hidden layer is calculated by using the input and the activation functionh(k):
In the formula, ωihWeights, x, for input layer neurons pointing to hidden layer neuronsi(k) For data on the ith neuron in the kth set of data, bhA threshold value for each neuron of the hidden layer;
input hi of neurons in the hidden layerh(k) Passing through Sigmoid type functions:
hoh(k)=f(hih(k))
Step S26: calculating input yi of each neuron of output layero(k) Then using input and activation function to calculate output yo of each neuron in output layero(k):
In the formula, ωhoWeights, ho, for neurons of the hidden layer to neurons of the output layerh(k) For the output of neurons of the hidden layer, boIs the threshold value of each neuron in the output layer;
input of each neuron of output layeryio(k) Passing through Sigmoid type functions:the mapping of (2) obtains the output yo of each neuron of the output layero(k)
yoo(k)=f(yio(k))
Step S27: calculating partial derivative delta of error function to each neuron of output layero(k):
δo(k)=(do(k)-yoo(k))yoo(k)(1-yoo(k))
Step S28: using the connection weight omega from the hidden layer to the output layerho(k) Partial derivative delta of error function to each neuron of output layero(k) And output ho of hidden layerh(k) Calculating partial derivative delta of error function to each neuron of hidden layerh(k):
Step S29: partial derivative delta of each neuron of output layer by using error functiono(k) And output ho of neurons of the hidden layerh(k) To correct the connection weight omegaho(k) And a threshold value bo(k):
In order to adjust the connection weight value after the adjustment,in order to adjust the connection weight before the adjustment,in order to achieve the adjusted threshold value, the threshold value is,for the threshold before adjustment, η is the learning rate and is taken between (0, 1).
Step S210: partial derivatives delta of neurons in the hidden layer using error functionsh(k) And input x of each neuron of input layeri(k) Modifying the connection weight and the threshold value;
in order to adjust the connection weight value after the adjustment,in order to adjust the connection weight before the adjustment,in order to achieve the adjusted threshold value, the threshold value is,for the threshold before adjustment, η is the learning rate and is taken between (0, 1).
Step S211: according to the output vector of the output layer, i.e. including the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd ultrahigh harmonic source predicted current amplitudeCorresponding desired output do(k) Calculating a global error E:
step S212: judging whether the network error meets the requirement, and finishing the algorithm when E is less than epsilon or the learning times is more than the set maximum times M; otherwise, the next learning sample and the corresponding expected output are randomly selected, and the process returns to step S24 to enter the next learning process.
In this embodiment, the step S4 specifically includes the following steps:
step S41: if the algorithm is finished, performing inverse normalization on output data of the neural network algorithm result to obtain a real value;
step S42: evaluating the predicted value of the current amplitude of the ultrahigh harmonic source according to the performance evaluation coefficient R of the neural network2:
The current amplitude prediction value of the ultra-high order harmonic wave of the ith data is obtained; y isiThe current amplitude real value of the ultrahigh harmonic wave of the ith data; n is the number of data to select a final fitting result curve so as to obtain the output characteristic of the ultrahigh harmonic source; evaluation of coefficient R for neural network Performance2Of selected value of R2The closer the value is to 1, the better the training effect is, the better the model fitting is, and the well-trained neural network model can be obtained; so that R under a given calculation accuracy epsilon is selected2And taking a fitting curve of the values as a fitting result, and further obtaining the output characteristic of the ultrahigh harmonic source.
In this embodiment, the step S5 specifically includes the following steps:
step S51: the power of the ultra-high harmonic source to the input data of the neural network model trained in step S4And ultra-high harmonic frequency
Step S52, the trained neural network model outputs dataThat is, the predicted current amplitude of the specific frequency input in step S51 at the power specified in step S51 by the ultra-high harmonic source;
step S53: and obtaining the output characteristic of the predicted ultra-high harmonic source.
Preferably, the implementation process of this embodiment is as follows: the method mainly comprises the following steps:
(1) acquiring ultrahigh harmonic current data or current data obtained by simulation within a period of time, which is acquired by an electric energy quality tester, from an ultrahigh harmonic source output end, and forming time series data;
(2) training the super-high harmonic current time series data under different powers by using a neural network algorithm;
(3) generating prediction data of each sub-ultrahigh harmonic current amplitude under a series of different powers by using a neural network algorithm;
(4) error calculation is carried out on the generated prediction data and the measured value (or the simulated value), and a performance evaluation coefficient R with the minimum error is selected according to the error calculation result2(coefficient of Performance evaluation R2Value R2The closer the value is to 1, the better the training effect, the better the model fitting), the appropriate R is selected2The fitting curve of the values is used as a fitting result, and further the output characteristic of the ultrahigh harmonic source is obtained;
(5) and inputting the power value and the frequency of the ultrahigh harmonic source to the trained neural network model, and outputting the predicted value of the output current amplitude of the ultrahigh harmonic source with the required frequency under the power by the model.
As shown in fig. 1 below, the output ultrahigh harmonic current value of the electric vehicle charging pile under different powers is obtained. FIG. 2 shows the value of the super-high harmonic current at a certain power, which is compared with the predicted value of the harmonic current of the super-high harmonic source at the same power obtained by the neural network algorithm, and R is selected2And analyzing the fitting curve when the voltage is 0.95433 as a fitting result, and further obtaining the output characteristic of the ultrahigh harmonic current of the electric automobile charging pile.
In the embodiment, the model is established by training actual measurement data or data obtained by simulation of the ultra-high harmonic source under different powers. The ultra-high harmonic current output frequency spectrum characteristic of the ultra-high harmonic source under a certain power condition is obtained.
In the embodiment, the ultrahigh harmonic current data (which can be from a large number of actual measurements or simulations) output by the ultrahigh harmonic source is trained, the output data of the network is compared with the actual data, and the weight is changed according to the learning rule until the difference between the output data and the actual data of the network reaches the required error range, so that the ultrahigh harmonic model is established.
The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.
Claims (2)
1. An ultrahigh-order harmonic source modeling method is characterized in that: the method comprises the following steps:
step S1: providing ultra-high harmonic current time series data under different powers in any time period;
step S2: training the ultra-high harmonic current time series data under different powers by using a neural network algorithm;
step S3: generating prediction data of each ultrahigh harmonic current amplitude under different powers by using a neural network algorithm;
step S4: the predicted data and actual measurement value generated in step S3The ultra-high harmonic current time-series data under different powers mentioned in the step S1 are subjected to error calculation, and according to the error calculation result, a performance evaluation coefficient R with the minimum error is selected2Selecting R at a given calculation accuracy epsilon2The value fitting curve is used as a fitting result to obtain the output characteristic of the ultrahigh harmonic source;
step S5: analyzing and judging the final fitting result of the step S4, if the error between the obtained predicted value and the measured value is within 0.0001-0.000001, obtaining a trained neural network model, inputting the power value and the frequency of the ultra-high harmonic source to the trained neural network model, and outputting the predicted value of the output current amplitude of the ultra-high harmonic source with the required frequency under the power by the model;
wherein, the step S4 specifically includes the following steps:
step S41: if the algorithm is finished, performing inverse normalization on output data of the neural network algorithm result to obtain a real value;
step S42: evaluating the predicted value of the current amplitude of the ultrahigh harmonic source according to the performance evaluation coefficient R of the neural network2:
The current amplitude prediction value of the ultra-high order harmonic wave of the ith data is obtained; y isiThe current amplitude real value of the ultrahigh harmonic wave of the ith data; n is the number of data to select a final fitting result curve so as to obtain the output characteristic of the ultrahigh harmonic source; evaluation of coefficient R for neural network Performance2Of selected value of R2The closer the value is to 1, the better the training effect is, the better the model fitting is, and the well-trained neural network model can be obtained; so that R under a given calculation accuracy epsilon is selected2The fitting curve of the values is used as a fitting result, and further the output characteristic of the ultrahigh harmonic source is obtained;
wherein, the step S5 specifically includes the following steps:
step S51: the power of the ultra-high harmonic source to the input data of the neural network model trained in step S4And ultra-high harmonic frequency
Step S52, the trained neural network model outputs dataThat is, the predicted current amplitude of the specific frequency input in step S51 at the power specified in step S51 by the ultra-high harmonic source;
step S53: and obtaining the output characteristic of the predicted ultra-high harmonic source.
2. The modeling method for an ultra-high order harmonic source of claim 1, wherein: the step S2 specifically includes the following steps:
step S21: providing ultrahigh harmonic current value data corresponding to different frequencies under different powers, comprising: input vector of input layer, hidden layer input vectorHidden layer output vectorOutput layer input vectorAnd an output vector of the output layer;
wherein the input vector of the input layer includes the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd the amplitude of the current actually measured by the ultra-high order harmonic source
The output vector of the output layer includes the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd ultrahigh harmonic source predicted current amplitude
Establishing a neural network model, wherein an input layer has 2 nodes, an output layer has 1 node, and an empirical formula is utilized: s is 2n +1, wherein n is the number of nodes of the input layer, and the number of nodes of the hidden layer is calculated to be 5; gradually increasing the number of nodes on the hidden layer by 5 nodes, increasing the number of nodes when the number of nodes is increased to 10, and preventing the network error from being reduced, wherein the number of node selection is optimal;
step S22: connection weight omega for input layer and hidden layerihThe connection weight omega of the hidden layer and the output layerhoThreshold b of each neuron in the hidden layerhAnd threshold b of each neuron of output layeroRespectively make one [ -1, 1 [ -1 [ ]]Random number inside, and let the error function be:setting the range of the calculation precision value epsilon to be 0.0001-0.000001 as a training termination condition, and setting the range of the maximum learning times M to be 1000-; the learning rate is selected in the range of 0.01-0.8;
step S23: normalizing the ultrahigh harmonic current value data corresponding to different frequencies under different powers in the step S21 to [0, 1] by adopting an S-shaped activation function;
step S24: randomly selecting a group of input samples, i.e. the data x (k) of the ultra-high harmonic current values corresponding to different frequencies at different powers in step S21 and the corresponding expected outputs d (k);
wherein the ultra high harmonic current value data x (k) includes power of an ultra high harmonic sourceUltra-high order harmonic frequencyAnd the amplitude of the current actually measured by the ultra-high order harmonic source
Step S25: computing input hi for each neuron in the hidden layerh(k) Then, the output ho of each neuron in the hidden layer is calculated by using the input and the activation functionh(k):
In the formula, ωihWeights, x, for input layer neurons pointing to hidden layer neuronsi(k) For data on the ith neuron in the kth set of data, bhA threshold value for each neuron of the hidden layer;
input hi of neurons in the hidden layerh(k) Passing through Sigmoid type functions:
hoh(k)=f(hih(k))
Step S26: calculating input yi of each neuron of output layero(k),Then, the output yo of each neuron in the output layer is calculated by using the input and the activation functiono(k):
In the formula, ωhoWeights, ho, for neurons of the hidden layer to neurons of the output layerh(k) For the output of neurons of the hidden layer, boIs the threshold value of each neuron in the output layer;
input yi of each neuron of output layero(k) Passing through Sigmoid type functions:the mapping of (2) obtains the output yo of each neuron of the output layero(k)
yoo(k)=f(yio(k))
Step S27: calculating partial derivative delta of error function to each neuron of output layero(k):
δo(k)=(do(k)-yoo(k))yoo(k)(1-yoo(k))
Step S28: using the connection weight omega from the hidden layer to the output layerho(k) Partial derivative delta of error function to each neuron of output layero(k) And output ho of hidden layerh(k) Calculating partial derivative delta of error function to each neuron of hidden layerh(k):
Step S29: partial derivative delta of each neuron of output layer by using error functiono(k) And output ho of neurons of the hidden layerh(k) To correct the connection weight omegaho(k) And a threshold value bo(k):
In order to adjust the connection weight value after the adjustment,in order to adjust the connection weight before the adjustment,in order to achieve the adjusted threshold value, the threshold value is,the threshold before adjustment is adopted, eta is the learning rate and takes a value between (0, 1);
step S210: partial derivative delta of each neuron in hidden layer by using error functionh(k) And input x of each neuron of input layeri(k) Modifying the connection weight and the threshold value;
in order to adjust the connection weight value after the adjustment,in order to adjust the connection weight before the adjustment,in order to achieve the adjusted threshold value, the threshold value is,the threshold before adjustment is adopted, eta is the learning rate and takes a value between (0, 1);
step S211: according to the output vector of the output layer, i.e. including the power of the ultra-high harmonic sourceUltra-high order harmonic frequencyAnd ultrahigh harmonic source predicted current amplitudeCorresponding desired output do(k) Calculating a global error E:
step S212: judging whether the network error meets the requirement, and finishing the algorithm when E is less than epsilon or the learning times is more than the set maximum times M; otherwise, the next learning sample and the corresponding expected output are randomly selected, and the process returns to step S24 to enter the next learning process.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910494902.8A CN110210152B (en) | 2019-06-06 | 2019-06-06 | Ultrahigh-order harmonic source modeling method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910494902.8A CN110210152B (en) | 2019-06-06 | 2019-06-06 | Ultrahigh-order harmonic source modeling method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110210152A CN110210152A (en) | 2019-09-06 |
CN110210152B true CN110210152B (en) | 2021-01-29 |
Family
ID=67791519
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910494902.8A Active CN110210152B (en) | 2019-06-06 | 2019-06-06 | Ultrahigh-order harmonic source modeling method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110210152B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110021929B (en) * | 2018-12-29 | 2023-01-13 | 国网内蒙古东部电力有限公司经济技术研究院 | Electromagnetic transient time domain simulation modeling method for fast switch type fault current limiter |
CN112531710B (en) * | 2020-11-18 | 2022-10-04 | 福州大学 | Method for predicting and evaluating harmonic source access |
CN113484605B (en) * | 2021-07-05 | 2022-04-22 | 华北电力大学 | Charger super-high harmonic emission evaluation method and device and storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101242101A (en) * | 2007-02-09 | 2008-08-13 | 上海磁浮交通工程技术研究中心 | Establishment method and simulation method for load harmony wave stabalized status model of power system |
CN101441231A (en) * | 2008-09-27 | 2009-05-27 | 武汉大学 | Harmonic electric energy metering error analytical apparatus |
CN103424621A (en) * | 2013-08-20 | 2013-12-04 | 江苏大学 | Artificial neural network detecting method of harmonic current |
CN105550450A (en) * | 2015-12-17 | 2016-05-04 | 国网智能电网研究院 | Modeling method of electric energy quality interference source characteristic harmonic wave |
Family Cites Families (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8340945B2 (en) * | 2009-08-24 | 2012-12-25 | International Business Machines Corporation | Method for joint modeling of mean and dispersion |
US9747214B2 (en) * | 2013-05-22 | 2017-08-29 | Schwegman Lundberg & Woessner, P.A. | Forecast modeling cache prefetch system and method |
US9379546B2 (en) * | 2013-06-07 | 2016-06-28 | The Board Of Trustees Of The University Of Alabama | Vector control of grid-connected power electronic converter using artificial neural networks |
CN103745065A (en) * | 2014-01-17 | 2014-04-23 | 国家电网公司 | Method for determining current after electrified railway multi-harmonic source superposition |
CN104951836A (en) * | 2014-03-25 | 2015-09-30 | 上海市玻森数据科技有限公司 | Posting predication system based on nerual network technique |
CN104049159A (en) * | 2014-05-16 | 2014-09-17 | 北京京东方能源科技有限公司 | Fault detection method and device of inverter |
CN106849705A (en) * | 2017-03-24 | 2017-06-13 | 三峡大学 | A kind of current predictive control method based on single-phase Vienna rectifiers |
CN107147142A (en) * | 2017-05-12 | 2017-09-08 | 天津大学 | It is a kind of count and probabilistic distributed power source as harmonic source modeling method |
CN107769259B (en) * | 2017-12-01 | 2019-08-23 | 北京航空航天大学 | A kind of inverter current forecast Control Algorithm based on discrete averaging model |
CN109800520B (en) * | 2019-01-25 | 2023-03-24 | 国网浙江省电力有限公司湖州供电公司 | Electric vehicle charging station harmonic modeling method based on neural network |
-
2019
- 2019-06-06 CN CN201910494902.8A patent/CN110210152B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101242101A (en) * | 2007-02-09 | 2008-08-13 | 上海磁浮交通工程技术研究中心 | Establishment method and simulation method for load harmony wave stabalized status model of power system |
CN101441231A (en) * | 2008-09-27 | 2009-05-27 | 武汉大学 | Harmonic electric energy metering error analytical apparatus |
CN103424621A (en) * | 2013-08-20 | 2013-12-04 | 江苏大学 | Artificial neural network detecting method of harmonic current |
CN105550450A (en) * | 2015-12-17 | 2016-05-04 | 国网智能电网研究院 | Modeling method of electric energy quality interference source characteristic harmonic wave |
Also Published As
Publication number | Publication date |
---|---|
CN110210152A (en) | 2019-09-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Fu et al. | A hybrid approach for multi-step wind speed forecasting based on two-layer decomposition, improved hybrid DE-HHO optimization and KELM | |
CN110210152B (en) | Ultrahigh-order harmonic source modeling method | |
Niu et al. | Uncertainty modeling for chaotic time series based on optimal multi-input multi-output architecture: Application to offshore wind speed | |
Ma et al. | A hybrid dynamic equivalent using ANN-based boundary matching technique | |
CN112651290A (en) | Hydroelectric generating set vibration trend prediction method and system | |
CN115062668B (en) | Harmonic parameter detection method and system based on RAdam optimized width learning | |
Kaplan et al. | A novel method based on Weibull distribution for short-term wind speed prediction | |
CN111553510A (en) | Short-term wind speed prediction method | |
Mehmood et al. | Design of backtracking search heuristics for parameter estimation of power signals | |
Xu et al. | Application of deep learning for high-speed differential via TDR impedance fast prediction | |
Zhang et al. | Short-term PV output power forecasting based on CEEMDAN-AE-GRU | |
CN112018755A (en) | Photovoltaic power distribution network reactive voltage prediction method and system based on cyclic neural network | |
CN114611676A (en) | New energy power generation system impedance model identification method and system based on neural network | |
CN113466710B (en) | SOC and SOH collaborative estimation method for energy storage battery in receiving-end power grid containing new energy | |
CN113987831A (en) | Identification method for low-frequency oscillation characteristics of multiple virtual synchronous generators | |
Zilio et al. | Black-box large-signal average modeling of DC-DC converters using NARX-ANNs | |
Darabian et al. | A UPFC‐based robust damping controller for optimal use of renewable energy sources in modern renewable integrated power systems | |
Dong et al. | Short-term load forecasting based on EEMD-approximate entropy and ELM | |
CN116826735A (en) | Broadband oscillation identification method and device for new energy station | |
CN108663570B (en) | Current harmonic analysis method based on trigonometric function neural network | |
Su et al. | A combined model based on secondary decomposition technique and grey wolf optimizer for short-term wind power forecasting | |
Taher et al. | Wavelet and signal analyzer based high-frequency ripple extraction in the context of MPPT algorithm in solar PV systems | |
Shivakumar et al. | Power system stability enhancement using bio inspired genetic and PSO algorithm implementation | |
Vittal et al. | A hybrid dynamic equivalent using ann-based boundary matching technique | |
Liu et al. | Intelligent voltage prediction of active distribution network with high proportion of distributed photovoltaics |
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 |