Power transmission line positive sequence and zero sequence parameter identification method based on improved physical information neural network

Abstract

The invention relates to a power transmission line positive sequence and zero sequence parameter identification method based on an improved physical information neural network, and the method comprises the steps: carrying out the processing of double-end fault recording data of a line, and dividing the data into normal operation, fault transient state and fault steady state of the line; constructing a differential equation taking the line positive sequence parameter as a coefficient, and calculating the coefficient of the differential equation by using the RBF-PINN1 and the normal operation data of the line to obtain the line positive sequence parameter; and constructing a differential equation taking the line zero-sequence parameter and the fault distance as coefficients, and calculating the coefficient of the differential equation by using the RBF-PINN2 and the line fault steady-state data to obtain the line zero-sequence parameter and the fault distance. Neural network training is guided by using the power transmission line physical model, the requirement for large-scale data is avoided, and the data acquisition difficulty is remarkably reduced; voltage and current differentials are accurately calculated by using an automatic differential function of the neural network, so that the accuracy of line parameter identification is remarkably improved; and meanwhile, the anti-noise capability is remarkably improved, so that the high precision can still be kept in a noise environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system line parameter identification, and in particular relates to a method for identifying positive-sequence and zero-sequence parameters of a power transmission line based on an improved physical information neural network. Background Art

[0002] Transmission lines are an important part of the power system. The accuracy of their parameters directly affects the calculation results of key links in the power system, such as state estimation, power flow calculation, relay protection setting, and fault location, which in turn affects the safe and stable operation of the power system. Therefore, it is of great significance to study the identification methods of transmission line parameters.

[0003] Line parameter identification methods can be categorized into time-domain and power-frequency phasor methods, depending on the signal processing domain. The power-frequency phasor method primarily relies on power-frequency phasor data of electrical quantities such as voltage and current acquired from synchronized phasor measurement units (PMUs). Furthermore, fault recording devices can capture instantaneous voltage and current values ​​before and after a fault. Based on this data, their power-frequency phasors can also be extracted, enabling line parameter identification.

[0004] The time-domain method directly utilizes the instantaneous value data of voltage and current, and identifies the line parameters based on the differential equation of the transmission line. Compared with the power frequency phasor method, the time-domain method avoids the errors that may occur during the power frequency phasor extraction process. In addition, unlike the power frequency phasor method, which requires multiple sets of phasor data at different time periods, the time-domain method only requires a section of recorded data to solve the line parameters, reducing the difficulty of data acquisition. The time-domain method involves the differentiation of recorded data, and existing methods usually use numerical differentiation methods to obtain it. However, numerical differentiation methods will introduce certain approximation errors and are also susceptible to noise.

[0005] With the continuous development of artificial intelligence (AI) technology, AI algorithms have also been applied to transmission line parameter identification. However, existing methods often rely on large amounts of high-quality data. In the case of line parameter identification based on fault recording data, collecting large amounts of fault recording data is challenging due to the low probability of transmission line failures. In such small sample scenarios, most machine learning techniques lack robustness and even struggle to achieve stable convergence. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for identifying the positive-sequence and zero-sequence parameters of a transmission line based on an improved physical information neural network (PINN), aiming to achieve accurate identification of transmission line parameters in a small sample scenario, and the identification results are less affected by noise. The present invention first uses a density-based spatial clustering algorithm (DBSCAN) to design a segmentation algorithm for recorded data, thereby achieving efficient use of data. Next, the radial basis function neural network (RBF-NN) is combined with the physical information neural network (PINN) to propose an RBF-PINN algorithm, and the algorithm is applied to the identification of positive-sequence and zero-sequence parameters of transmission lines, thereby improving the accuracy of identification and noise resistance.

[0007] The present invention solves the technical problem by the following technical solutions:

[0008] A method for identifying positive-sequence and zero-sequence parameters of a transmission line based on an improved physical information neural network, the method comprising the following steps:

[0009] S1. Processing the line double-end fault recording data, dividing it into three sections: line normal operation data, fault transient data, and fault steady-state data;

[0010] S2. Based on the phase component model of the transmission line under normal operation, a differential equation with the line positive sequence parameter as the coefficient is constructed. The coefficient of the differential equation is calculated using RBF-PINN1 and the normal operation data of the line to obtain the line positive sequence parameter;

[0011] S3. Based on the phase component model of an asymmetric ground fault on a transmission line, a differential equation with the line zero-sequence parameters and fault distance as coefficients is constructed. The coefficients of the differential equation are calculated using RBF-PINN2 and line fault steady-state data to obtain the line zero-sequence parameters and fault distance.

[0012] Moreover, the S1 is specifically:

[0013] (1) Input recording data sampling time t and voltage u a1 , and calculate the fourth-order voltage difference Δ 4 u a1 ;

[0014] (2) t and Δ4 u a1 Combine to form the input data set (t, Δ 4 u a1 ), and standardize them;

[0015] (3) Set the initial radius value ε = 0.01 and its increment Δε = 0.001. Since the data points in the transient phase of the fault are far away from other points and cannot be connected to any cluster, DBSCAN will regard them as noise points. Therefore, the effective target cluster number is set to 2;

[0016] (4) Use DBSCAN to perform clustering and obtain class labels;

[0017] (5) Calculate the number of valid clusters excluding noise points. If the number of valid clusters is equal to 2, proceed to the next step; otherwise, adjust ε and return to step (4);

[0018] (6) The maximum moment in the first cluster is taken as the starting moment of the transient stage, and the minimum moment in the second cluster is taken as the ending moment of the transient stage. Based on this, the recorded data is divided into three sections: normal operation, fault transient and steady state.

[0019] The advantages and beneficial effects of the present invention are:

[0020] 1. The present invention inherits the advantages of the traditional time domain method, such as not needing to extract power frequency phasors and requiring less data.

[0021] 2. The present invention utilizes the physical model of the transmission line to guide the neural network training, avoiding the need for large-scale data and significantly reducing the difficulty of data acquisition.

[0022] 3. The present invention uses the automatic differentiation function of the neural network to accurately calculate the voltage and current differentials, significantly improving the accuracy of line parameter identification.

[0023] 4. The addition of RBF-NN significantly improves the noise resistance of the present invention, so that it can still maintain high accuracy in a noisy environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of the positive-sequence and zero-sequence parameter identification method for transmission lines based on the improved PINN;

[0025] Figure 2 The voltage recording and first to fourth order differential waveforms of the line when a two-phase short circuit grounding fault occurs;

[0026] Figure 3 This is the phase component model diagram of the transmission line under normal operating conditions;

[0027] Figure 4Schematic diagram of RBF-PINN1 for solving the positive sequence parameters of transmission lines;

[0028] Figure 5 This is the phase component model diagram of the transmission line when an asymmetric ground fault occurs;

[0029] Figure 6 Schematic diagram of RBF-PINN2 for solving zero-sequence parameters and fault distance of transmission lines;

[0030] Figure 7 This is a transmission line simulation model diagram;

[0031] Figure 8 This is a schematic diagram of the segmentation results of the recorded data. DETAILED DESCRIPTION

[0032] The present invention will be further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the scope of protection of the present invention cannot be limited thereto.

[0033] like Figure 1 As shown, a method for identifying positive-sequence and zero-sequence parameters of a transmission line based on an improved physical information neural network is innovative in that the steps of the method are as follows:

[0034] S1. When a two-phase short circuit grounding fault occurs on the line, the voltage waveform of phase a at one end of the line and its first to fourth order differential waveforms are as follows: Figure 2 As shown in the figure, it can be seen that the differential operation amplifies the high-frequency components of the transient stage of the fault, making its differential value significantly higher than that of the other two stages, and this gap becomes more significant as the order increases. And although the differential operation also has an amplifying effect on noise, the high-order differential value of the transient stage is still significantly larger than that of the other two stages. Therefore, the recorded data can be segmented according to the fourth-order difference of voltage or current. DBSCAN is a density-based clustering algorithm that can effectively distinguish clusters from noise. The algorithm only needs to set two parameters: radius and minimum number of sample points, without specifying the number of clusters. However, the algorithm is particularly sensitive to the choice of radius. In the task of segmenting recorded data, it can be clearly divided into three segments.

[0035] Therefore, the present invention improves the DBSCAN algorithm by dynamically adjusting the radius by specifying the number of clusters, thereby achieving a more accurate segmentation effect.

[0036] The specific process of the proposed algorithm is as follows:

[0037] Step 1: Input the sampling time t and voltage u of the recorded data a1 , and calculate the fourth-order voltage difference Δ 4 u a1 ;

[0038] Step 2: t and Δ 4 u a1 Combine to form the input data set (t, Δ 4 u a1 ), and standardize them;

[0039] Step 3. Set the initial radius value ε = 0.01 and its increment Δε = 0.001. Since the data points in the transient phase of the fault are far away from other points and cannot be connected to any cluster, DBSCAN will regard them as noise points. Therefore, set the effective target cluster number to 2;

[0040] Step 4: Use DBSCAN to perform clustering and obtain class labels;

[0041] Step 5: Calculate the number of valid clusters excluding noise points. If the number of valid clusters is equal to 2, proceed to step 6; otherwise, adjust ε and return to step 4.

[0042] Step 6: The maximum moment in the first cluster is taken as the start moment of the transient stage, and the minimum moment in the second cluster is taken as the end moment of the transient stage. Based on this, the recorded data is divided into three sections: normal operation state, fault transient state and steady state.

[0043] S2, the phase component model of the transmission line under normal operating conditions is as follows Figure 3 As shown in the figure, the length of the line is l, R s 、L s is the self-resistance and self-inductance of each phase, R m 、L m is the mutual resistance and mutual inductance between phases, C p 、C g is the phase capacitance and grounding capacitance. aj 、u bj 、u cj (j=1, 2) is the voltage across the line, i aj 、i bj 、i cj (j=1, 2) is the current at both ends of the line, i aj ′、i bj ′、i cj '(j=1, 2) is the current at both ends of the line after passing through the grounding capacitance and phase capacitance. Figure 3 We can get equations (1) and (3).

[0044]

[0045] Among them, C1 is the positive sequence capacitance of the line, C0 is the zero sequence capacitance of the line, and it satisfies

[0046]

[0047]

[0048] in,

[0049]

[0050] According to formula (1) and formula (3), construct RBF-PINN1 as Figure 4 As shown in the figure, the input of RBF-NN1 is the sampling time t of the recorded data under normal operation of the line. zc The output is used to fit the seven dependent variables in the above equations (1) and (3), and is given corresponding physical meanings, respectively. In terms of structure, RBF-NN1 is a three-layer network. The first layer is the input layer, which is responsible for receiving only the recorded data in the normal state and does not process it in any way. The second layer is the RBF layer, which has 60 neurons / center points. At the same time, the Gaussian function is used as the radial basis function, as shown in Equation (5). The third layer is the output layer, which maps the output of the RBF layer to a 7-dimensional output space through linear transformation. The output results of the RBF layer and the output layer are shown in Equations (6) and (7).

[0051]

[0052] Among them, c i and β i is a trainable parameter.

[0053] o (1) =[φ1(x;c1,σ1),φ2(x;c2,σ2),…,φ 60 (x;c 60 , σ 60 )] (6)

[0054]

[0055] Among them, W (2) is the weight matrix of the output layer, b (2) is the bias of the output layer, W (2) and b (2) are all trainable parameters.

[0056] In order to make RBF-NN1 fit the recorded data as much as possible, the

[0057]

[0058] Where N is the number of sampling points of the recorded data.

[0059] In order to make the prediction results of RBF-NN1 satisfy the physical laws, first, according to the first equation of formula (1), the physical information function is constructed as

[0060]

[0061] Where i represents the i-th data point in the recording sequence.

[0062] At the same time, according to the first equation in formula (3), the physical information function is constructed as

[0063]

[0064] Then, construct the physical information loss function

[0065]

[0066] In order to make RBF-NN1 fit the recorded data while also meeting the physical laws, a loss function is constructed.

[0067] Loss1=Loss net +Loss phy1 (13)

[0068] Loss2=Loss net +Loss phy2 (14)

[0069] The training process of RBF-PINN1 is as follows: first, the Adam optimizer is used to minimize the loss function Loss1 to obtain C1; on this basis, the Adam optimizer is used to minimize the loss function Loss2 to obtain R1 and L1.

[0070] S3. Taking the a-phase grounding short circuit at the fault point f as an example, the phase component model of the transmission line when an asymmetric grounding fault occurs is as follows: Figure 5 As shown in the figure, α represents the ratio of the length of f from the left end of the line to the entire line length, that is, the fault distance is αl. Assuming that the impedance parameters of the transmission line are uniform everywhere, the self-resistance and self-inductance of each phase of the line to the left of f are αR s ,αL s , the mutual resistance and mutual inductance between phases are aR m , αL m , the self-resistance and self-inductance of each phase of the line on the right side of f are (1-α)R s 、(1-α)L s , the mutual resistance and mutual inductance between phases are (1-α)R m 、(1-α)L m To simplify the calculation, the phase-to-phase capacitance and grounding capacitance of the line are concentrated and equivalent at both ends of the line, which are C p and C g .Depend on Figure 5We can obtain equations (15), (16) and (19).

[0071]

[0072]

[0073] in,

[0074]

[0075] According to equations (15), (16) and (19), construct RBF-PINN2 as Figure 6 As shown in Figure 2, the input of RBF-NN2 is the sampling time t of the recorded data when an asymmetric ground fault occurs on the line. gz The output is used to fit the 12 dependent variables in equations (15), (16) and (19), and is given corresponding physical meanings, respectively. The structure of RBF-NN2 is similar to that of RBF-NN1, with the main differences being that the RBF layer has 65 neurons and the output dimension of the output layer is 12.

[0076] In order to make RBF-NN2 fit the recorded data as much as possible, the

[0077]

[0078] In order to make the prediction results of RBF-NN2 satisfy the physical laws, first, according to formula (15), the physical information function is constructed as

[0079]

[0080] According to the first equation in (16), the physical information function is constructed

[0081]

[0082] According to formula (19), construct the physical information function

[0083]

[0084] Then, construct the physical information loss function

[0085]

[0086] In order to make RBF-NN2 satisfy the physical laws while fitting the recorded data, a loss function is constructed.

[0087] Loss1=Loss net +Loss phy1 (27)

[0088] Loss2=Loss net +Loss phy2 (28)

[0089] Loss3=Loss net +Loss phy3 (29)

[0090] The training process of RBF-PINN2 is as follows: first, use the Adam optimizer to minimize the loss function Loss1 to obtain C0; on this basis, use the Adam optimizer to minimize the loss function Loss2 to obtain α; finally, use the Adam optimizer to minimize the loss function Loss3 to obtain R0 and L0.

[0091] Build a double-terminal transmission line simulation model in the MATLAB / simulink simulation environment, such as Figure 7 As shown in the figure, the π-type line model is adopted, and its parameters are shown in Table 1; The fault point f is located at a distance al from the left end of the line.

[0092] Table 1 Line parameters and their actual values

[0093] Line parameters True value <![CDATA[C1 / μF]]> 0.28665 R1 / Ω 0.57285 <![CDATA[L1 / H]]> 0.04202 <![CDATA[C0 / μF]]> 0.17440 R0 / Ω 17.38800 <![CDATA[L0 / H]]> 0.18569

[0094] The first set of experimental results of this embodiment is used to verify whether the present invention can effectively divide the recorded data into three sections. When a single-phase ground short circuit or a two-phase ground short circuit occurs in the line, the segmented results of the recorded data of the present invention are as follows: Figure 8 As shown in the figure, the voltage u of phase a at the left end of the line is a1 As an example, the algorithm's segmentation results are intuitively demonstrated. It can be seen that the present invention can effectively divide the recorded data into three segments: normal operating state, fault transient state, and fault steady state.

[0095] The second set of experimental results in this embodiment is used to verify whether the present invention accurately identifies the positive-sequence and zero-sequence parameters of the transmission line. When a single-phase ground fault or a two-phase ground fault occurs in the transmission line, the identification results and absolute errors of the line parameters and the fault distance proportional coefficient of the present invention, the PINN-based method, and the Prony algorithm-based method are shown in Table 2. As can be seen from Table 2, compared with the PINN-based method, the present invention has higher accuracy; compared with the Prony algorithm-based method, although the identification accuracy of the positive-sequence parameters of the present invention is lower, it has significantly higher accuracy in the identification of the zero-sequence parameters. In addition, when identifying parameters other than the zero-sequence capacitance, the absolute error of the present invention is within 0.5%, and when identifying the zero-sequence capacitance, the absolute error is also within 3%.

[0096] Table 2 Parameter identification results and absolute errors of the present invention

[0097]

[0098] The third set of experimental results in this embodiment verifies whether the accuracy of the present invention's line parameter identification is affected by the size of the transition resistance. Table 3 shows the identification results and absolute errors for line parameters and fault distance proportionality coefficients when the transmission line experiences a single-phase or two-phase ground fault through transition resistances of 0.001Ω, 10Ω, 20Ω, and 50Ω, respectively. As Table 3 demonstrates, the present invention maintains high accuracy even when grounding through a relatively large transition resistance.

[0099] Table 3 Parameter identification results and absolute errors of the present invention under different transition resistances

[0100]

[0101] The fourth set of experimental results in this embodiment verifies whether the accuracy of line parameter identification by the present invention is affected by fault distance. When a single-phase or two-phase grounding fault occurs on the line, and the fault distance proportionality coefficients are 0.1, 0.3, and 0.5, respectively, the identification results and absolute errors for the line parameters and fault distance proportionality coefficients are shown in Table 4. As shown in Table 4, the present invention achieves high accuracy at various fault distances.

[0102] Table 4 Parameter identification results and absolute errors of the present invention at different fault distances

[0103]

[0104] The fifth set of experimental results of this embodiment is used to verify whether the accuracy of the line parameter identification of the present invention is affected by noise. Gaussian noise with a mean of 0 and a standard deviation of 0.05%, 0.1%, 0.15% and 0.2% respectively is introduced into the voltage and current data at both ends of the line. When a single-phase ground short circuit or a two-phase ground short circuit fault occurs in the transmission line, the identification results and absolute errors of the line parameters and fault distance proportional coefficients of the present invention, the PINN-based method and the Prony algorithm-based method are shown in Table 5. It can be seen from Table 5 that compared with the other two methods, the present invention can still maintain a relatively low identification error under higher noise intensity, indicating that it has strong anti-noise ability. Although the identification results of zero-sequence capacitance are easily affected by noise, the identification accuracy of the present invention is still significantly better than the other two algorithms.

[0105] Table 5 Parameter identification results and absolute errors of different methods under different noise intensities

[0106]

[0107]

[0108] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for identifying positive-sequence and zero-sequence parameters of a transmission line based on an improved physical information neural network, characterized by: The steps of the method are: S1. Processing the line double-end fault recording data, dividing it into three sections: line normal operation data, fault transient data, and fault steady-state data; S2. Based on the phase component model of the transmission line under normal operation, a differential equation with the line positive sequence parameter as the coefficient is constructed. The coefficient of the differential equation is calculated using RBF-PINN1 and the normal operation data of the line to obtain the line positive sequence parameter; S3. Based on the phase component model of an asymmetric ground fault on a transmission line, a differential equation with the line zero-sequence parameters and fault distance as coefficients is constructed. The coefficients of the differential equation are calculated using RBF-PINN2 and line fault steady-state data to obtain the line zero-sequence parameters and fault distance.

2. The method for identifying positive-sequence and zero-sequence parameters of a power transmission line based on an improved physical information neural network according to claim 1, characterized in that: The S1 is specifically: (1) Input recording data sampling time t and voltage u a1 , and calculate the fourth-order voltage difference Δ 4 u a1 ; (2) t and Δ 4 u a1 Combine to form the input data set (t, Δ 4 u a1 ), and standardize them; (3) Set the initial radius value ε = 0.01 and its increment Δε = 0.

001. Since the data points in the transient phase of the fault are far away from other points and cannot be connected to any cluster, DBSCAN will regard them as noise points. Therefore, the effective target cluster number is set to 2; (4) Use DBSCAN to perform clustering and obtain class labels; (5) Calculate the number of valid clusters excluding noise points. If the number of valid clusters is equal to 2, proceed to the next step; otherwise, adjust ε and return to step (4); (6) The maximum moment in the first cluster is taken as the starting moment of the transient stage, and the minimum moment in the second cluster is taken as the ending moment of the transient stage. Based on this, the recorded data is divided into three sections: normal operation, fault transient and steady state.

Images