A System-Side Variable Harmonic Impedance Estimation Method Based on Improved BP Neural Network

Abstract

The present invention discloses a method for estimating the variable harmonic impedance on the system side based on an improved BP neural network, comprising the following steps: S1: The system samples the harmonic voltage and harmonic current data at the common connection point, performs segmented processing on the harmonic voltage and harmonic current, and roughly estimates the harmonic impedance value of each segment by using binary linear regression; S2: The harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value are repeated and continuously arranged in the original order to generate an extended sequence, and filtering and normalization processing are performed; S3: A BP neural network model combined with an attention mechanism is built, and the loss function of the output layer of the BP neural network model is improved; S4: The processed harmonic voltage and harmonic current are used as inputs, and the processed harmonic impedance is used as an output to train and test the BP neural network model, and the estimated value of the variable harmonic impedance on the system side is obtained; The present invention uses an improved BP neural network to estimate the variable harmonic impedance on the system side, and still has a high estimation accuracy under the condition of large background harmonics.

Description

Technical Field

[0001] The present invention relates to the field of new power systems, and in particular to a method for estimating the variable harmonic impedance on the system side based on an improved BP neural network. Background Art

[0002] With the continuous increase in the proportion of wind power generation and photovoltaic power generation in China, as well as the access of various new energy devices and traditional electrical appliances, the harmonic impedance on the system side of the power grid will change slowly. Accurately measuring the harmonic impedance on the system side can ensure the stable operation of the power system and electrical equipment and improve the power quality. Currently, the harmonic impedance estimation methods are mainly divided into two types: "intervention method" and "non-intervention method". Among them, the intervention method calculates the harmonic impedance by artificially injecting appropriate harmonic currents into the power grid to generate interference, mainly including the switching operation of capacitor banks, etc., but this method may have a certain impact on the system; the "non-intervention method" refers to estimating the harmonic impedance on the system side by measuring the harmonic voltage and current at the common connection point of the system. The main methods include the regression method, the fluctuation method, the independent component method, etc. However, the above methods have low accuracy for the harmonic impedance on the system side in the case of high harmonic emission levels and slow changes in harmonic impedance. Therefore, the present invention proposes a method for estimating the variable harmonic impedance on the system side based on an improved BP neural network. Summary of the Invention

[0003] To achieve the above object, the technical solution provided by the present invention is as follows:

[0004] A method for estimating the variable harmonic impedance on the system side based on an improved BP neural network, comprising the following steps:

[0005] S1: The system samples the harmonic voltage and harmonic current data at the common connection point, performs segmented processing on the harmonic voltage and harmonic current, and roughly estimates the harmonic impedance value of each segment by using binary linear regression;

[0006] S2: Repeat and continuously arrange the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value in the original order to generate an extended sequence, and perform filtering and normalization processing;

[0007] S3: Build a BP neural network model combined with an attention mechanism, and improve the loss function of the output layer of the BP neural network model;

[0008] S4: Use the processed harmonic voltage and harmonic current as inputs, and the processed harmonic impedance as the output, train and test the BP neural network model to obtain the estimated value of the variable harmonic impedance on the system side;

[0009] Further, the specific steps of step S1 are as follows:

[0010] S1-1: The system samples the harmonic voltage and harmonic current at the common connection point and performs segmented processing:

[0011] The complex forms of harmonic voltage and harmonic current are sampled at the common connection point. A set of data includes a sampled harmonic voltage and the corresponding harmonic current. Depending on the number of sampling points, the number of data groups contained in each segment is different:

[0012]

[0013] In Equation (1), D is the number of segments, M is the number of sampling points, and m is the number of data groups contained in each segment;

[0014] S1-2: Roughly estimate the system-side harmonic impedance using binary linear regression:

[0015] V pcc =(I pcc +I s )×Z s =I pcc ×Z s +V s (2)

[0016] V pccr =I pccr ×Z sr -I pcci ×Z si +V sr (3)

[0017] V pcci =I pccr ×Z si +I pcci ×Z sr +V si (4)

[0018] In Equations (2)-(4), V pcc and I pcc are the sampled harmonic voltage and harmonic current respectively, I s is the equivalent harmonic current source on the system side, Z s is the system-side harmonic impedance, V s is the equivalent harmonic voltage source on the system side, V pccr and V pcci are the real and imaginary parts of the harmonic voltage V pcc , V sr and V si are the real and imaginary parts of the equivalent harmonic voltage source V s on the system side, I pccr and I pcci are the real and imaginary parts of the harmonic current I pcc , Z sr and Z si are the real and imaginary parts of the system-side harmonic impedance Z s ;

[0019]

[0020] In Equation (5), T represents taking the transpose of a matrix, and the roughly estimated harmonic impedance value is obtained by solving Equation (5); X1, X2, Y1, and Y2 are solution matrices formed based on the real and imaginary parts of the harmonic voltage and the real and imaginary parts of the harmonic current. The specific expressions are as follows:

[0021]

[0022] Furthermore, the specific steps of step S2 are as follows:

[0023] S2-1: Repeating and continuously arranging the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value in the original order of the data segment to generate an extended sequence and performing filtering processing:

[0024] The harmonic voltage and harmonic current data obtained by sampling and the roughly estimated harmonic impedance value are repeated ten times in the original order, sorted in the front-to-back order, and then the moving average filtering processing is performed on each data sequence. The moving average filtering performs averaging processing through window data, and the filtering output is:

[0025]

[0026] In Equation (7), y is the sequence after the moving average filtering process, x is the original sequence, j is the sliding window position, and P is the sliding window size;

[0027] S2-2: Normalizing the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value obtained after repetition and filtering:

[0028]

[0029] In Equation (8), L i is the normalized value, X i is the i-th element in sequence X, X max is the maximum value of sequence X, X min is the minimum value of sequence X.

[0030] Furthermore, the specific steps of step S3 are as follows:

[0031] S3-1: Optimizing the estimation result at the actual impedance mutation point by improving the loss function:

[0032]

[0033] In Equation (9), mseloss is the mean square error, Y is the predicted output impedance value of the BP neural network model, N is the total number of samples, L is the actual harmonic impedance value, diff is the absolute error between the predicted impedance value and the actual impedance value, m is the set threshold, the values of the absolute error diff greater than the set threshold m are accumulated to obtain the total loss penalty, k is the weight for penalty, and loss is the improved loss function;

[0034] S3-2: Build a BP neural network model combined with an attention mechanism:

[0035] Build a BP neural network structure combined with an attention mechanism, including an input layer, an attention mechanism layer, two hidden layers, and an output layer with an improved loss function; the input layer corresponds to two feature inputs, the number of neurons in the two hidden layers is variable, an attention mechanism can strengthen feature selection, optimize weight allocation, and improve training efficiency, and the output layer corresponds to one feature output; the adam optimization algorithm is used, and the exponential moving averages of the first moment and the second moment of the gradient are calculated at the same time, and the calculation results are corrected for bias to avoid the gradient estimation in the initial stage of training from tending to zero. The implementation process of the adam optimization algorithm is as follows:

[0036]

[0037] In Equation (10), q is the number of iterations, g q is the gradient at the q-th iteration, m is the estimated value of the first moment of the gradient, v is the estimated value of the second moment of the gradient, θ is the weight, m q-1 、m q 、m q+1 and v q-1 、v q 、v q+1 and θ q-1 、θ q 、θ q+1 respectively represent the estimated values of the first moment of the gradient, the estimated values of the second moment of the gradient, and the weights of the previous iteration, the current iteration, and the next iteration, is the gradient of the weight θ q-1 at the (q - 1)-th iteration, β1 is the exponential decay rate controlling the estimated value of the first moment, β2 is the exponential decay rate controlling the estimated value of the second moment, represents the q-th power operation of the exponential decay rate β1 controlling the estimated value of the first moment and the exponential decay rate β2 controlling the estimated value of the second moment, α is the learning rate, μ is a very small constant to prevent the denominator from being zero, and ⊙ is the vector product operation.

[0038] Furthermore, the specific steps of the step S4 are as follows:

[0039] Taking the real part / imaginary part of the processed harmonic voltage and harmonic current as input features, and the real part / imaginary part of the roughly estimated harmonic impedance as output features, use 90% of the data for BP neural network training, and the remaining 10% for testing the real part / imaginary part of the harmonic impedance on the system side.

[0040] Compared with the prior art, the principle and advantages of this solution are as follows:

[0041] The present invention discloses a method for estimating variable harmonic impedance on the system side based on an improved BP neural network, including the following steps: First, the system samples the harmonic voltage and harmonic current data at the common connection point, performs segmented processing on the harmonic voltage and harmonic current, and roughly estimates the harmonic impedance value of each segment using binary linear regression; Second, repeat and continuously arrange the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value in the original order to generate an extended sequence, and perform filtering and normalization processing; Then, build a BP neural network model combined with an attention mechanism, and improve the loss function of the output layer of the BP neural network model; Finally, use the processed harmonic voltage and harmonic current as input, and the processed harmonic impedance as output, train and test the BP neural network model to obtain the estimated value of the variable harmonic impedance on the system side; The present invention uses an improved BP neural network to estimate the variable harmonic impedance on the system side, and still has high estimation accuracy in the case of large background harmonics. Brief Description of the Drawings

[0042] Figure 1 It is a flowchart of the method for estimating variable harmonic impedance on the system side based on an improved BP neural network in the embodiment of the present invention;

[0043] Figure 2 It is a Norton equivalent circuit diagram in the embodiment of the present invention;

[0044] Figure 3 It is the estimated result of the real part of the variable harmonic impedance on the system side when the harmonic emission level k = 0.1 in the embodiment of the present invention;

[0045] Figure 4 It is the estimated result of the imaginary part of the variable harmonic impedance on the system side when the harmonic emission level k = 0.1 in the embodiment of the present invention;

[0046] Figure 5 It is the root mean square error RMSE of the estimated real part of the variable harmonic impedance on the system side between the method of the present invention and the binary linear regression method at different harmonic emission levels k in the embodiment of the present invention;

[0047] Figure 6 It is the root mean square error RMSE of the estimated imaginary part of the variable harmonic impedance on the system side between the method of the present invention and the binary linear regression method at different harmonic emission levels k in the embodiment of the present invention. Detailed Embodiment

[0048] The present invention will be further described below in conjunction with specific embodiments:

[0049] Figure 1 The figure shows a flowchart of a system-side variable harmonic impedance estimation method based on an improved BP neural network, including the following steps:

[0050] S1: The system samples the harmonic voltage and harmonic current data at the common connection point, performs segmented processing on the harmonic voltage and harmonic current, and roughly estimates the harmonic impedance value of each segment using binary linear regression;

[0051] The specific steps of step S1 are as follows:

[0052] S1-1: The system samples the harmonic voltage and harmonic current at the common connection point and performs segmented processing:

[0053] The complex forms of the harmonic voltage and harmonic current are sampled at the common connection point. A set of data includes a sampled harmonic voltage and the corresponding harmonic current. According to the number of sampling points, the number of data groups included in each segment is different:

[0054]

[0055] In formula (11), D is the number of segments, M is the number of sampling points, and m is the number of data groups included in each segment;

[0056] S1-2: Roughly estimate the system-side harmonic impedance using binary linear regression:

[0057] V pcc =(I pcc +I s )×Z s =I pcc ×Z s +V s (12)

[0058] V pccr =I pccr ×Z sr -I pcci ×Z si +V sr (13)

[0059] V pcci =I pccr ×Z si +I pcci ×Z sr +V si (14)

[0060] In formulas (12)-(14), V pcc and I pcc are the sampled harmonic voltage and harmonic current respectively, and Is is the equivalent harmonic current source on the system side, Z s is the harmonic impedance on the system side, V s is the equivalent harmonic voltage source on the system side, V pccr and V pcci are the real and imaginary parts of the harmonic voltage V pcc V sr and V si are the real and imaginary parts of the equivalent harmonic voltage source V on the system side, I s I pccr and I pcci are the real and imaginary parts of the harmonic current I pcc I sr and Z si are the real and imaginary parts of the harmonic impedance Z on the system side s ;

[0061]

[0062] In Equation (15), T represents taking the transpose of a matrix, and the roughly estimated harmonic impedance value is obtained by solving Equation (15); X1, X2, Y1, and Y2 are solution matrices formed based on the real and imaginary parts of the harmonic voltage and the real and imaginary parts of the harmonic current. The specific expressions are as follows:

[0063]

[0064] Next, in step S2, the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value are repeated and continuously arranged in the original order to generate an extended sequence, followed by filtering and normalization processing. The specific process is as follows:

[0065] S2-1: Repeating and continuously arranging the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value in the original order of the data segment to generate an extended sequence and filtering processing:

[0066] The sampled harmonic voltage and current data and the roughly estimated harmonic impedance value are repeated ten times in the original order, sorted in the front-to-back order, and then a moving average filtering process is performed on each data sequence. The moving average filtering averages the data through a window. The filtering output is:

[0067]

[0068] In Equation (17), y is the sequence after the moving average filtering process, x is the original sequence, j is the sliding window position, and P is the sliding window size;

[0069] S2-2: Normalization processing of the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value obtained after repetition and filtering:

[0070]

[0071] In Equation (18), L i is the normalized value, X i is the i-th element in sequence X, X max is the maximum value of sequence X, X min is the minimum value of sequence X.

[0072] The following step S3 is to build a BP neural network model combined with an attention mechanism and improve the loss function of the output layer of the BP neural network model. The specific process is as follows:

[0073] S3-1: Optimize the estimation result of the actual impedance mutation point by improving the loss function:

[0074]

[0075] In Equation (19), mseloss is the mean square error, Y is the predicted output impedance value of the BP neural network model, N is the total number of samples, L is the actual harmonic impedance value, diff is the absolute error between the predicted impedance value and the actual impedance value, m is the set threshold, the values of the absolute error diff greater than the set threshold m are accumulated to obtain the total loss penalty, k is the weight for penalty, and loss is the improved loss function;

[0076] S3-2: Build a BP neural network model combined with an attention mechanism:

[0077] Build a BP neural network structure combined with an attention mechanism, including an input layer, an attention mechanism layer, two hidden layers, and an output layer with an improved loss function; the input layer corresponds to two feature inputs, the number of neurons in the two hidden layers is variable, an attention mechanism can strengthen feature selection, optimize weight distribution and improve training efficiency, and the output layer corresponds to one feature output; adopt the adam optimization algorithm, calculate the exponential moving average of the first moment and the second moment of the gradient at the same time, and perform bias correction on the calculation result to avoid the gradient estimation in the initial stage of training from tending to zero. The implementation process of the adam optimization algorithm is:

[0078]

[0079] In Equation (20), q is the number of iterations, g q is the gradient at the q-th iteration, m is the estimated value of the first moment of the gradient, v is the estimated value of the second moment of the gradient, θ is the weight, m q-1 、m q 、m q+1 and v q-1 、v q 、v q+1 versus θ q-1 、θ q 、θq+1 respectively represent the first-order moment estimate of the gradient, the second-order moment estimate of the gradient, and the weight for the last iteration, the current iteration, and the next iteration is the weight θ when the iteration number is q - 1 q-1 is the gradient of, β1 is the exponential decay rate controlling the first-order moment estimate, and β2 is the exponential decay rate controlling the second-order moment estimate represents the q-th power operation of the exponential decay rate β1 controlling the first-order moment estimate and the exponential decay rate β2 controlling the second-order moment estimate, α is the learning rate, μ is a very small constant to prevent the denominator from being zero, and ⊙ is the vector product operation.

[0080] Next, step S4 takes the processed harmonic voltage and harmonic current as inputs, and the processed harmonic impedance as the output, trains and tests the BP neural network model, and obtains the estimated value of the system-side variable harmonic impedance. The specific process is as follows:

[0081] Take the real part / imaginary part of the processed harmonic voltage and harmonic current as input features, and the real part / imaginary part of the roughly estimated harmonic impedance as output features. Use 90% of the data for BP neural network training, and the remaining 10% for testing the real part / imaginary part of the system-side harmonic impedance.

[0082] To verify the effectiveness of the system-side variable harmonic impedance estimation method based on the improved BP neural network, a Figure 2 Norton equivalent circuit is built in MATLAB, and the simulation parameters are set as follows: The user-side harmonic current source I c = 40 + j60 A, with a 10% random interference added to the real part and a 15% random interference added to the imaginary part; the user-side harmonic impedance Z c = 50 + j100 Ω, with a 10% random interference added to the real part and a 5% random interference added to the imaginary part; the system-side harmonic current source I s The amplitude is k times the amplitude of the user-side harmonic current source I c A 10% random interference is added to the amplitude of the harmonic current source I s The phase angle of the system-side harmonic current source I s is 56°, with a 5% random interference added; the system-side harmonic impedance Z s is adjusted within different sampling intervals to simulate the mutation and continuous change of the impedance value. Within the first 400 sampling points, set the system-side harmonic impedance Z s = 20 + j15 Ω, and within the sampling points from 401 to 1200, set the system-side harmonic impedance Z s = Z c / m, where m ranges from 2.981 to 8.944. Within the sampling points from 1201 to 1600, set the system-side harmonic impedance Z s = 10 + j7.5 Ω. Throughout the process, for the system-side harmonic impedance Z sAdd 2.5% random interference to the real part and 5% random interference to the imaginary part, and sample to obtain the total number of samples M = 1600.

[0083] Figure 3 and Figure 4 are the estimation results of the real part and the imaginary part of the system-side variable harmonic impedance when the harmonic emission level k = 0.1. Through calculation, the impedance estimation error of the real part is about 7.19%, and the impedance estimation error of the imaginary part is about 1.76%. It can be seen that the harmonic impedance estimation result is relatively accurate and the overall fitting degree is relatively high.

[0084] In order to further verify that the method of the present invention has a high estimation accuracy, the estimation results of the method of the present invention are compared with the estimation results of the binary linear regression method at different harmonic emission levels k, as Figure 5 and Figure 6 shown. As the harmonic emission level increases, when using the binary linear regression method, the root mean square error RMSE of the estimation results of the real part and the imaginary part of the harmonic impedance both increase significantly; while the root mean square error RMSE of the real part of the harmonic impedance estimated by the method of the present invention is always less than 1.5, and the root mean square error RMSE of the imaginary part is always less than 0.9, which is much smaller than the root mean square error RMSE of the binary linear regression method under the same harmonic emission level. This shows that the method of the present invention still has a high estimation accuracy under the conditions of a high harmonic impedance emission level and changes in the system-side harmonic impedance.

[0085] The above-described embodiments are only the preferred embodiments of the present invention, and do not limit the scope of implementation of the present invention. Therefore, all changes made according to the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for estimating the side - variable harmonic impedance of a system based on an improved BP neural network, characterized in that, It includes the following steps: S1: The system samples to obtain the harmonic voltage and harmonic current data at the common connection point, processes the harmonic voltage and harmonic current in segments, and roughly estimates the harmonic impedance value of each segment using binary linear regression: S1-1: The system samples to obtain the harmonic voltage and harmonic current at the common connection point and performs segmented processing: The complex forms of the harmonic voltage and harmonic current are sampled at the common connection point. A set of data includes a sampled harmonic voltage and the corresponding harmonic current. Depending on the number of sampling points, the number of data groups included in each segment is different: In Equation (1), D is the number of segments, M is the number of sampling points, and m is the number of data groups included in each segment; S1-2: Roughly estimate the system-side harmonic impedance using binary linear regression: V pcc = (I pcc + I s ) × Z s = I pcc × Z s + V s (2) V pccr = I pccr × Z sr - I pcci × Z si + V sr (3) V pcci = I pccr × Z si + I pcci × Z sr + V si (4) In formulas (2)-(4), V pcc and I pcc are the harmonic voltage and harmonic current obtained by sampling, I s is the equivalent harmonic current source on the system side, Z s is the harmonic impedance on the system side, V s is the equivalent harmonic voltage source on the system side, V pccr and V pcci are the real and imaginary parts of the harmonic voltage V pcc , V sr and V si are the real and imaginary parts of the equivalent harmonic voltage source V s on the system side, I pccr and I pcci are the real and imaginary parts of the harmonic current I pcc , Z sr and Z si are the real and imaginary parts of the harmonic impedance Z s on the system side; In Equation (5), T represents taking the transpose of the matrix. The roughly estimated harmonic impedance value is obtained by solving Equation (5); X1, X2, Y1, and Y2 are the solution matrices composed of the real and imaginary parts of the harmonic voltage and the real and imaginary parts of the harmonic current. The specific expressions are: S2: Repeat and continuously arrange the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value in the original order to generate an extended sequence, and perform filtering and normalization processing: S2-1: Repeat and continuously arrange the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value in the original order of the data segments to generate an extended sequence and perform filtering processing: Repeat the sampled harmonic voltage and harmonic current data and the roughly estimated harmonic impedance value ten times in the original order, sort them in the front-back order, and then perform moving average filtering processing on each data sequence. The moving average filtering performs averaging processing through window data, and the filtering output is: In Equation (7), y is the sequence after moving average filtering processing, x is the original sequence, j is the sliding window position, and P is the sliding window size; S2-2: Perform normalization processing on the harmonic voltage, harmonic current, and the roughly estimated harmonic impedance value obtained after repetition and filtering: In formula (8), L i is the normalized value, X i is the i-th element in sequence X, X max is the maximum value of sequence X, X min is the minimum value of sequence X; S3: Build a BP neural network model combined with an attention mechanism and improve the loss function of the output layer of the BP neural network model: S3-1: Optimize the estimation result at the actual impedance mutation point by improving the loss function: In Equation (9), mseloss is the mean square error, Y is the predicted output impedance value of the BP neural network model, N is the total number of samples, L is the actual harmonic impedance value, diff is the absolute error between the predicted impedance value and the actual impedance value, m is the set threshold, the values of the absolute error diff greater than the set threshold m are accumulated to obtain the total loss penalty, k is the weight for penalty, and loss is the improved loss function; S3-2: Build a BP neural network model combined with an attention mechanism: Build a BP neural network structure combined with an attention mechanism, including an input layer, an attention mechanism layer, two hidden layers, and an output layer with an improved loss function; the input layer corresponds to two feature inputs, the number of neurons in the two hidden layers is variable, an attention mechanism can strengthen feature selection, optimize weight allocation and improve training efficiency, and the output layer corresponds to one feature output; adopt the adam optimization algorithm, calculate the exponential moving average of the first-order moment and the second-order moment of the gradient at the same time, and perform bias correction on the calculation results to avoid the gradient estimation at the initial stage of training from deviating to zero. The implementation process of the adam optimization algorithm is as follows: In Equation (10), q is the number of iterations, and g q is the gradient at the q-th iteration, m is the first-order moment estimate of the gradient, v is the second-order moment estimate of the gradient, θ is the weight, and m q-1 , m q , m q+1 and v q-1 , v q , v q+1 corresponding to θ q-1 , θ q , θ q+1 represent the first-order moment estimate of the gradient, the second-order moment estimate of the gradient, and the weight at the last iteration, the current iteration, and the next iteration, respectively. is the gradient of the weight θ q-1 at the (q - 1)-th iteration, β1 is the exponential decay rate controlling the first-order moment estimate, β2 is the exponential decay rate controlling the second-order moment estimate, represents the q-th power operation of the exponential decay rate β1 controlling the first-order moment estimate and the exponential decay rate β2 controlling the second-order moment estimate. α is the learning rate, μ is a very small constant to prevent the denominator from being zero, and ⊙ is the vector product operation; S4: Use the processed harmonic voltage and harmonic current as inputs, and the processed harmonic impedance as the output to train and test the BP neural network model to obtain the estimated value of the system-side variable harmonic impedance: Use the real part / imaginary part of the processed harmonic voltage and harmonic current as input features, and the real part / imaginary part of the roughly estimated harmonic impedance as output features. Use 90% of the data for BP neural network training, and the remaining 10% is used to test the real part / imaginary part of the system-side harmonic impedance.

Images