A gas generator loss-of-field identification method, system and electronic device

Abstract

This invention provides a method, system, and electronic device for identifying the loss of excitation of a gas generator, belonging to the field of power system relay protection. The method for identifying the loss of excitation of a gas generator includes: collecting the measured impedance trajectory of the gas generator within a set time period; determining the current feature vector based on the measured impedance trajectory; and determining whether the gas generator has lost excitation based on the current feature vector and a discriminant model. The discriminant model is obtained by training a multi-core twin support vector machine using a pre-trained sample set and based on the Ocean Predator algorithm, which integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning. Considering the dynamic process of the measured impedance of the gas generator, and using the Ocean Predator algorithm, which integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning, to train the multi-core twin support vector machine improves the accuracy and efficiency of the discriminant model for identifying loss of excitation, thereby simultaneously satisfying the speed and selectivity of loss of excitation protection actions.

Description

Technical Field

[0001] This invention relates to the field of power system relay protection, and in particular to a data-driven method, system, and electronic equipment for identifying the loss of excitation of a gas generator based on the temporal spatial variation characteristics of measured impedance. Background Technology

[0002] Loss of excitation, either partially or completely, in large gas turbine generators is a common and serious fault. Compared to steam turbine generators, gas turbine generators have lower asynchronous power and larger governor time lag. After loss of excitation, the rotor will overspeed more quickly, accompanied by unit vibration and other phenomena, posing a significant threat to the unit itself and the power grid. Therefore, faster operation of loss of excitation protection is required. During the "3.21" blackout in the Brazilian power grid in 2018, a large gas turbine unit in the northeastern power grid experienced a false tripping of its loss of excitation protection due to system oscillations. The disconnection of this unit exacerbated system instability, ultimately leading to its collapse.

[0003] Traditional loss-of-excitation protection, which primarily uses the static stability boundary of the turbine terminal impedance as the criterion, can only determine whether loss of excitation has occurred by measuring the final result of the impedance change at the turbine terminal. However, it only utilizes the final result of the impedance change measurement and ignores the dynamic process, failing to reflect the changes in the measured impedance under various disturbances in complex power grid environments. This makes it difficult to simultaneously meet the requirements of selectivity and speed. With the increasing complexity of power grid structures and changes in the operating environment, the reliability and speed of loss-of-excitation protection for large gas turbines face significant challenges. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, and electronic device for identifying the demagnetization of a gas generator, which can improve the accuracy and efficiency of identifying the demagnetization of a gas generator.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for identifying the loss of excitation in a gas generator includes:

[0007] The measurement impedance trajectory of the gas generator is collected; the measurement impedance trajectory includes the measurement impedance time sequence points at each moment within a set time period;

[0008] Based on the measured impedance trajectory, determine the current feature vector;

[0009] Based on the current feature vector, it is determined whether the gas generator has lost its magnetism using a discriminant model. The discriminant model is obtained by training a multi-core twin support vector machine using a pre-trained sample set and a marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning. The training sample set includes feature vectors of multiple demagnetized samples and feature vectors of multiple non-demagnetized samples.

[0010] Optionally, the gas generator demagnetization identification method further includes:

[0011] If the gas generator loses its magnetism, a demagnetization protection action will be performed on the gas generator.

[0012] Optionally, the current feature vector includes distance variance, minimum value of the derivative of the motion azimuth angle, mean value of the derivative of the motion azimuth angle, mean value of the derivative of the direction angle, maximum value of velocity, and velocity skewness;

[0013] The step of determining the current feature vector based on the measured impedance trajectory specifically includes:

[0014] Extract the timing points of the measured impedance trajectory within the set time window;

[0015] For any measurement impedance timing point, calculate the motion azimuth angle, direction angle, and velocity of the measurement impedance timing point based on the measurement impedance timing point;

[0016] Calculate the distance variance based on the timing points of each measured impedance;

[0017] Calculate the minimum and average values ​​of the derivatives of the motion azimuth angles based on the motion azimuth angles at each time point of the measured impedance.

[0018] Calculate the mean value of the derivative of the direction angle based on the direction angle at each measured impedance timing point;

[0019] Calculate the maximum velocity and velocity deviation based on the velocity at each measured impedance timing point.

[0020] Optionally, the distance variance of the i-th time window can be calculated using the following formula:

[0021]

[0022] in, Let N be the distance variance of the i-th time window. i Let be the number of timing points for measuring impedance in the i-th time window. For the k-th impedance measurement timing point, It is the reciprocal of the per-unit value of the gas turbine output power.

[0023] Optionally, the minimum and mean values ​​of the motion azimuth derivatives for the i-th time window are calculated using the following formulas:

[0024]

[0025]

[0026] Among them, D α.min,i u is the minimum value of the derivative of the motion azimuth angle in the i-th time window. Dα,iLet N be the mean of the derivatives of the motion azimuth angles of the i-th time window. i Let α be the number of timing points for measuring impedance in the i-th time window. k Let α be the azimuth angle of the k-th impedance measurement timing point. k+1 Δt is the motion azimuth angle of the (k+1)th impedance measurement timing point, and Δt is the calculation time interval.

[0027] Optionally, the method for establishing the discrimination model includes:

[0028] Obtain a measurement impedance trajectory sample set of the gas generator; the measurement impedance trajectory sample set includes multiple demagnetization samples and multiple non-demagnetization samples; each demagnetization sample and each non-demagnetization sample includes the measurement impedance time sequence points at each moment within a set time period;

[0029] The measurement impedance timing points of each demagnetized sample and each non-demagnetized sample within a set time window are extracted to obtain the measurement impedance trajectory sample set within the corresponding time window.

[0030] Based on the set of measured impedance trajectory samples within the set time window, calculate the feature vector of each demagnetization sample and the feature vector of each non-demagnetization sample.

[0031] Determine the demagnetization normalization matrix based on the feature vectors of each demagnetization sample;

[0032] Determine the non-demagnetization normalization matrix based on the feature vectors of each non-demagnetization sample;

[0033] Based on the demagnetization normalized matrix and the non-demagnetization normalized matrix, the multi-core twin support vector machine is iteratively trained using the marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning to obtain the discriminant model.

[0034] Optionally, the step of iteratively training a multi-core twin support vector machine based on the demagnetization normalized matrix and the non-demagnetization normalized matrix, using a marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimension learning, to obtain a discriminative model, specifically includes:

[0035] A chaotic opposition strategy is used for population initialization to determine the initial position and opposition position of each prey in the population, so as to obtain the initial population.

[0036] The initial positions of each prey in the initial population are iteratively updated using the marine predator algorithm to obtain the updated positions of each prey.

[0037] The adaptive t-distribution is determined based on the number of iterations of the marine predator algorithm;

[0038] Based on the adaptive t-distribution, the update position of each prey is mutated to determine the corresponding mutated position of each prey.

[0039] The fitness of each prey is determined based on the location of its mutations, and the prey is then divided into an elite group and a learning group based on their fitness.

[0040] Based on the mutation positions of the elite group and the learning group, the mutation positions of each prey in the learning group are cross-referenced to obtain the optimal position of each prey; the optimal position of each prey includes the first upper limit penalty coefficient, the second upper limit penalty coefficient, the kernel function weight, and the radial basis kernel hyperparameter.

[0041] Based on the optimal position of each prey, the hyperparameters of the multi-core twin support vector machine are trained and optimized according to the demagnetization normalization matrix and the non-demagnetization normalization matrix to obtain the discrimination model. The hyperparameters of the multi-core twin support vector machine include the normal vector of the demagnetized sample in the kernel space, the normal vector of the non-demagnetized sample in the kernel space, the bias coefficient of the demagnetized sample, and the bias coefficient of the non-demagnetized sample under the set time window.

[0042] Optionally, the discrimination model is:

[0043]

[0044] Where, f(X) num f(X) represents the discrimination result. num ) = 1 indicates that the corresponding gas generator has lost its magnetism, f(X) num ) = -1 indicates that the corresponding gas generator is not demagnetized, num is the total number of demagnetized and non-demagnetized samples, X num The characteristic matrix includes the demagnetization normalized matrix and the non-demagnetization normalized matrix, x j Let ω be the j-th eigenvector in the eigenmatrix, sgn() be the sign function, and ω be the eigenvector. 1,i Let ω be the normal vector of the demagnetization sample in the kernel space under the i-th time window. 2,i Let b be the normal vector of the non-demagnetized sample in the kernel space under the i-th time window. 1,i Let b be the bias coefficient of the demagnetization sample in the i-th time window. 2,i is the bias coefficient for the non-demagnetized sample in the i-th time window.

[0045] To achieve the above objectives, the present invention also provides the following solution:

[0046] A gas generator demagnetization identification system includes:

[0047] Impedance trajectory acquisition unit is used to acquire the measured impedance trajectory of the gas generator; the measured impedance trajectory includes the measured impedance time sequence points at each moment within a set time period;

[0048] The feature vector calculation unit, connected to the impedance trajectory acquisition unit, is used to determine the current feature vector based on the measured impedance trajectory.

[0049] The demagnetization identification unit, connected to the feature vector calculation unit, is used to determine whether the gas generator has lost its magnetism based on the current feature vector and a discrimination model. The discrimination model is obtained by training a multi-core twin support vector machine using a pre-trained sample set and a marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning. The training sample set includes feature vectors of multiple demagnetized samples and feature vectors of multiple non-demagnetized samples.

[0050] To achieve the above objectives, the present invention also provides the following solution:

[0051] An electronic device includes a memory and a processor, the memory storing a computer program, and the processor running the computer program to enable the electronic device to perform the above-described gas generator demagnetization identification method.

[0052] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The time series points of the measured impedance of a gas generator are collected at various times within a set time period to obtain the measured impedance trajectory. The current feature vector is determined based on the measured impedance trajectory. Based on the current feature vector, and using a discriminant model, it is determined whether the gas generator has lost excitation. The discriminant model is obtained by training a multi-core twin support vector machine using a pre-trained sample set and a marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning. Considering the dynamic process of the gas generator's measured impedance, and using the marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning to train the multi-core twin support vector machine, the accuracy and efficiency of the discriminant model in identifying loss of excitation are improved, thereby simultaneously satisfying the speed and selectivity of loss of excitation protection actions. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of the gas generator demagnetization identification method of the present invention;

[0055] Figure 2 This is a schematic diagram of an extended power grid system based on different power sources connected to the IEEE 39-node network.

[0056] Figure 3 Schematic diagram of circuit setup for excitation fault in gas generator;

[0057] Figure 4 The convergence curves are plotted at time windows of 0.5s and 1.5s.

[0058] Figure 5 A schematic diagram illustrating the overall process of the gas generator demagnetization identification method;

[0059] Figure 6 This is a schematic diagram of the demagnetization fault identification process with a time window length of 0.4-0.6s;

[0060] Figure 7 The demagnetization fault identification process has a time window length of 1.4-1.6s;

[0061] Figure 8 This is a schematic diagram of the gas generator demagnetization identification system of the present invention.

[0062] Symbol explanation:

[0063] Impedance trajectory acquisition unit-101, feature vector calculation unit-102, loss of magnetization identification unit-103. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] This invention addresses the fact that the entire process of measuring impedance changes at the generator terminals contains a wealth of system information. By utilizing the dynamic process of measuring impedance changes, it is possible to better identify loss of excitation faults. Drawing on the successful application of trajectory recognition and AI (Artificial Intelligence) technology in power systems, this invention provides a method, system, and electronic equipment for identifying loss of excitation in gas generators, based on the dynamic process of measuring impedance. This achieves the goal of simultaneously satisfying the speed and selectivity of loss of excitation protection actions.

[0066] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0067] Example 1

[0068] like Figure 1 As shown, the method for identifying the loss of magnetism in a gas generator according to the present invention includes:

[0069] S11: Acquire the measurement impedance trajectory of the gas generator. The measurement impedance trajectory includes the measurement impedance time sequence points at each moment within a set time period.

[0070] S12: Determine the current feature vector based on the measured impedance trajectory.

[0071] S13: Based on the current feature vector, determine whether the gas generator has lost its magnetism using a discriminant model. The discriminant model is obtained by training a multi-core twin support vector machine using a pre-set training sample set and based on the Ocean Predator algorithm, which integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning. The training sample set includes feature vectors from multiple demagnetized samples and feature vectors from multiple non-demagnetized samples.

[0072] S14: If the gas generator loses its magnetism, then the gas generator will be protected against loss of magnetism.

[0073] Furthermore, in step S12, the current feature vector includes the distance variance, the minimum value of the derivative of the motion azimuth angle, the mean value of the derivative of the motion azimuth angle, the mean value of the derivative of the direction angle, the maximum value of the velocity, and the velocity skewness.

[0074] Step S12 specifically includes:

[0075] (1) Extract the timing points of the measured impedance trajectory within a set time window. In this embodiment, a discrimination model for different time windows is trained. In practical applications, a discrimination model corresponding to the set time window is used for loss of magnetization identification.

[0076] (2) For any measurement impedance timing point, calculate the motion azimuth angle, direction angle and velocity of the measurement impedance timing point according to the measurement impedance timing point.

[0077] Specifically, the azimuth angle, direction angle, and velocity of the k-th impedance measurement timing point are calculated using the following formulas:

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] in, To obtain the voltage vector sequence using the full-wave Fourier algorithm, The current vector sequence is obtained by using the full-wave Fourier algorithm. Here, N represents the number of impedance measurement timing points within a fixed time window, and β represents the total number of impedance measurement timing points.k ζ is the first intermediate quantity. k As the second intermediate quantity, α is the reciprocal of the per-unit value of the gas turbine output power. k Let ρ be the azimuth angle of the k-th impedance measurement timing point. k Let v be the direction angle at the k-th impedance measurement timing point. k Δt represents the velocity at the k-th impedance measurement timing point, and Δt is the calculation time interval. In this embodiment, Δt = 25ms.

[0085] (3) Calculate the distance variance based on the time series points of each measured impedance.

[0086] Specifically, the distance variance of the i-th time window is calculated using the following formula:

[0087]

[0088] in, Let N be the distance variance of the i-th time window. i Let be the number of timing points for measuring impedance in the i-th time window. For the k-th impedance measurement timing point, It is the reciprocal of the per-unit value of the gas turbine output power.

[0089] (4) Calculate the minimum value of the derivative of the azimuth angle and the mean value of the derivative of the azimuth angle based on the azimuth angle of each measured impedance timing point.

[0090] Specifically, the minimum value and the mean value of the derivative of the motion azimuth angle for the i-th time window are calculated using the following formula:

[0091]

[0092]

[0093] Among them, D α.min,i u is the minimum value of the derivative of the motion azimuth angle in the i-th time window. Dα,i Let N be the mean of the derivatives of the motion azimuth angles of the i-th time window. i Let α be the number of timing points for measuring impedance in the i-th time window. k Let α be the azimuth angle of the k-th impedance measurement timing point. k+1 Δt is the motion azimuth angle of the (k+1)th impedance measurement timing point, and Δt is the calculation time interval.

[0094] (5) Calculate the mean value of the derivative of the direction angle based on the direction angle of each measured impedance timing point.

[0095] Specifically, the mean value of the direction angle derivative of the i-th time window is calculated using the following formula:

[0096]

[0097] Among them, u ρ,i Let be the mean of the direction angle derivatives for the i-th time window.

[0098] (6) Calculate the maximum speed and speed deviation based on the speed at each measured impedance timing point.

[0099] Specifically, the maximum velocity and velocity skewness of the i-th time window are calculated using the following formula:

[0100]

[0101]

[0102] Among them, v max,i v is the maximum velocity value in the i-th time window. S,i S represents the velocity skewness of the i-th time window. v,i Let be the standard deviation of the velocity sequence for the i-th time window.

[0103] As one specific implementation method, the method for establishing the discrimination model includes:

[0104] S21: Obtain the measurement impedance trajectory sample set of the gas generator. The measurement impedance trajectory sample set includes multiple demagnetization samples and multiple non-demagnetization samples. Each demagnetization sample and each non-demagnetization sample includes the measurement impedance time sequence points at each moment within a set time period.

[0105] In this embodiment, the measured impedance trajectories under various operating conditions of the gas generator being studied and connected to the power system are determined. Specifically, a sample set of measured impedance trajectories is established by obtaining simulation models or historical waveform data based on actual parameters.

[0106] The simulation platform is based on PSCAD / EMTDC, and the topology of the target power system is as follows: Figure 2 As shown in the figure, 1 to 39 represent nodes. Considering the integration of power systems with varying strengths and a hybrid wind-solar power system, loss-of-excitation faults and other abnormal operating states were simulated under various operating conditions. Verification of the loss-of-excitation protection scheme was conducted on one of the three parallel-operating gas generators, G1. G5 is a hybrid wind-solar power source. Different types of loss-of-excitation faults were set at each of the 39 nodes, including open-circuit loss-of-excitation, short-circuit loss-of-excitation, and partial loss-of-excitation. Non-loss-of-excitation disturbance conditions included short circuit, open circuit, ultra-static stability, load shedding, unit tripping, governor failure, increased commissioning of nearby units, normal load fluctuations, power flow control, transformer tap adjustment, and power fluctuations. Measurement impedance trajectories under the above conditions were collected to establish a trajectory sample library, totaling 700 samples, and divided into a measurement impedance trajectory sample set and a test set at a ratio of 4:1.

[0107] Among them, the excitation fault setting circuit of the gas generator is as follows Figure 3 As shown in the diagram: UL is the excitation voltage, IL is the excitation current, Rm is the demagnetizing resistor, MK1 is the first switch, and MK2 is the second switch. During normal operation, the first switch MK1 is closed and the second switch MK2 is open. In the event of a total demagnetization fault, the first switch MK1 is open and the second switch MK2 is closed. The demagnetizing resistor Rm, depending on its value, achieves short-circuit demagnetization (Rm = 0), open-circuit demagnetization (Rm = ∞), and demagnetization via different demagnetizing resistors (Rm ∈ (0, ∞)). Furthermore, partial demagnetization can be achieved by directly adjusting the magnitude of the DC current source.

[0108] S22: Extract the measurement impedance timing points of each demagnetized sample and each non-demagnetized sample within the set time window to obtain the measurement impedance trajectory sample set within the corresponding time window.

[0109] S23: Calculate the feature vector of each demagnetized sample and the feature vector of each non-demagnetized sample based on the measured impedance trajectory sample set within the set time window.

[0110] Specifically, the calculation process of the feature vector of the demagnetized sample and the feature vector of the non-demagnetized sample is the same as the calculation process of the current feature vector in step S12. The corresponding feature vector is calculated by extracting different time windows, which will not be repeated here.

[0111] S24: Determine the demagnetization normalization matrix A based on the feature vectors of each demagnetization sample. i .

[0112] S25: Determine the non-demagnetization normalization matrix B based on the eigenvectors of each non-demagnetization sample. i .

[0113] Specifically, the demagnetization normalization matrix A i for Non-demagnetization normalized matrix B i for m is the total number of demagnetized samples, and n is the total number of non-demagnetized samples.

[0114] in, F = A, B, where subscript i represents the i-th time window and subscript j represents the j-th sample (either a demagnetized or non-demagnetized sample). This represents the feature vector of the j-th sample within the i-th time window. This represents the distance variance of the j-th sample within the i-th time window. This represents the minimum value of the derivative of the motion azimuth angle of the j-th sample under the i-th time window. This represents the mean derivative of the motion azimuth angle of the j-th sample within the i-th time window. This represents the mean of the derivatives of the orientation angles of the j-th sample within the i-th time window. This represents the maximum velocity of the j-th sample within the i-th time window. This represents the velocity skewness of the j-th sample within the i-th time window.

[0115] S26: Based on the demagnetization normalized matrix and the non-demagnetization normalized matrix, the MKL-TSVM (Multi-Kernel Learning Twin-SVM) is iteratively trained using the marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning to obtain the discriminant model.

[0116] Specifically, (1) a chaotic opposition strategy is used to initialize the population, determine the initial position and opposition position of each prey in the population, and obtain the initial population.

[0117] In the model parameter optimization problem of demagnetization identification, the initial position X of the p-th prey in the population (the position of each prey represents the value of a set of optimization variables) is... p It can be represented as:

[0118] X p =[C 1,p C 2,p γ p ε p ].

[0119] Among them, C 1,p C is the first upper limit penalty coefficient for the p-th prey. 2,p γ is the second upper limit penalty coefficient for the p-th prey. p Let ε be the kernel function weight for the p-th prey. p Let be the radial basis kernel hyperparameter of the p-th prey.

[0120] The chaotic opposition strategy generates the opposite position of the prey's initial position, and the generation method is as follows:

[0121]

[0122] in, Let lb be the q-th variable opposite to the p-th prey position. q μb is the upper bound of the q-th variable. q Let η be the lower bound of the q-th variable. p For Tent chaotic mapping, The multiplication symbol is X. p,q Let C be the q-th variable representing the initial position of the p-th prey (the first variable is C). 1,p The second variable is C. 2,p The third variable is γ p The fourth variable is ε pThe fitness of the randomly generated initial prey positions and their opposite positions is sorted to obtain the top r prey positions with the best fitness as the initial population.

[0123] (2) The initial position of each prey in the initial population is updated iteratively using the marine predator algorithm to obtain the updated position of each prey.

[0124] (3) Determine the adaptive t-distribution based on the number of iterations of the marine predator algorithm.

[0125] (4) Based on the adaptive t-distribution, the update position of each prey is mutated to determine the mutated position of each prey.

[0126] Specifically, after completing the three-stage MPA (Marine Predators Algorithm) update, an adaptive t-distribution operator is set to automatically adjust the population mutation probability based on the number of optimization iterations. The mutation method is as follows:

[0127] X p ′=X p +X p ·t(iter);

[0128] In the formula, X p Let ′ be the mutation position of the p-th prey, and t(iter) represent the t-distribution with the iteration number iter as the degree of freedom.

[0129] (5) Determine the fitness of each prey based on the variation location of each prey, and divide the prey into elite group and learning group based on fitness.

[0130] (6) Based on the mutation positions of the elite group and the learning group, the mutation positions of each prey in the learning group are cross-referenced to obtain the optimal position of each prey. The optimal position of each prey includes the first upper bound penalty coefficient, the second upper bound penalty coefficient, the kernel function weight, and the radial basis kernel hyperparameter.

[0131] Specifically, the prey influenced by FADs are ranked by fitness and divided into an elite group with high fitness and a learning group with low fitness. The average dimensionality of the learning group with low fitness is subtracted from that of the elite group, and the top H groups with the largest differences are cross-referenced. The final cross-reference decision is based on whether the fitness of the learning group improves after cross-reference. The cross-reference process can be represented as follows:

[0132]

[0133]

[0134]

[0135] Among them, Xav,q X represents the average dimension of the q-th dimension (the q-th variable at the mutation position) of the elite group. p，q Let ' be the q-th variable in the mutation position of the p-th prey in the elite group, ΔX p,q X represents the difference in the average dimensionality between the learning group and the elite group. p "This is the optimal position for the p-th prey." It is the p-th prey after the q-th dimension crosses.

[0136] (7) Based on the optimal position of each prey, the hyperparameters of the multi-core twin support vector machine are trained and optimized according to the demagnetization normalization matrix and the non-demagnetization normalization matrix to obtain the discrimination model. The hyperparameters of the multi-core twin support vector machine include the normal vector of the demagnetized sample in the kernel space, the normal vector of the non-demagnetized sample in the kernel space, the bias coefficient of the demagnetized sample, and the bias coefficient of the non-demagnetized sample under the set time window.

[0137] It has

[0138] S = [K(A) i ,[A i B i ] T )e1];

[0139] R = [K(B i ,[A i B i ] T )e2];

[0140] Where S is the first intermediate matrix, R is the second intermediate matrix, e1 is an m-dimensional unit column vector, e2 is an n-dimensional unit column vector, and A i B is the demagnetization normalization matrix. i Let K be the non-demagnetization normalized matrix, and K(·) be the kernel function. As shown in the following equation:

[0141] K(x,y)=εK Poly (x,y)+(1-ε)K RBF (x,y);

[0142] In the formula, ε is used to adjust the proportion of different kernel functions, and its value ranges from [0,1]. ε represents the radial basis kernel hyperparameter in the optimal prey position obtained from the improved marine predator algorithm. poly (x,y) is the Poly kernel function, K RBF (x, y) is the RBF kernel function, and the two are calculated by the following formula:

[0143] K Poly (x,y)=((x,y)+1) d ;

[0144] KRBF (x,y)=exp(-γ||xy|| 2 );

[0145] Where d is the order of the positive integer polynomial, typically taking the value of 2 or 3, and γ represents the reciprocal of the influence radius of the sample selected by the model as a support vector, which is the kernel function weight in the optimal prey position obtained by the improved marine predator algorithm.

[0146] The training process of the hyperplane for classifying demagnetized samples in MKL-TSVM is as follows:

[0147]

[0148] Solving the standard quadratic programming problem above, we get:

[0149] [ω 1,i b 1,i ] T =-(S T S) -1 R T α.

[0150] The training process of the non-demagnetization sample classification hyperplane of MKL-TSVM is as follows:

[0151]

[0152] Solving the standard quadratic programming problem above, we obtain...

[0153] [ω 2,i b 2,i ] T =(R T R) -1 S T λ.

[0154] Where α is α j The column vector formed by λ is λ j The column vector formed by C1 is α j The upper limit of C2 is λ. j The upper limit, α j and λ j Let ω be the Lagrange multiplier. 1,i Let ω be the normal vector of the demagnetization sample in the kernel space under the i-th time window. 2,i Let b be the normal vector of the non-demagnetized sample in the kernel space under the i-th time window. 1,i Let b be the bias coefficient of the demagnetization sample in the i-th time window. 2,i is the bias coefficient for the non-demagnetized sample in the i-th time window.

[0155] In this embodiment, the population of IMPA (Improved Marine Predator Algorithm) is set to 40, the maximum number of iterations is 100, the range of C1 and C2 is [0, 1000], and the range of γ is [0, 100]. Solving the above equation yields the discriminant model:

[0156]

[0157] Where, f(X) num f(X) represents the discrimination result. num ) = 1 indicates that the corresponding gas generator has lost its magnetism, f(X) num ) = -1 indicates that the corresponding gas generator is not demagnetized, num is the total number of demagnetized and non-demagnetized samples, X num The eigenma is a 7×num matrix, including the demagnetization normalized matrix and the non-demagnetization normalized matrix, x j Let ω be the j-th eigenvector in the eigenmatrix, sgn() be the sign function, and ω be the eigenvector. 1,i Let ω be the normal vector of the demagnetization sample in the kernel space under the i-th time window. 2,i Let b be the normal vector of the non-demagnetized sample in the kernel space under the i-th time window. 1,i Let b be the bias coefficient of the demagnetization sample in the i-th time window. 2,i is the bias coefficient for the non-demagnetized sample in the i-th time window.

[0158] After training MKL-TSVM, retain the feature attributes and model parameters, including: parameters of the model for different time windows, and ω obtained after training. 1,i ω 2,i and b 1,i b 2,i Hyperparameters of different time window models, Lagrange coefficients α i , λ i Upper limit penalty coefficient C 1,i C 2,i The hyperparameter d of the poly kernel function i RBF kernel hyperparameter γ i Kernel function weights ε i .

[0159] Furthermore, after MKL-TSVM training is completed, evaluation metrics are selected based on the validation set to assess the performance of the discriminant model. In addition to accuracy (ACC), precision (PRE), recall (REC), and F1 score are chosen as evaluation metrics, and their calculation formulas are as follows:

[0160]

[0161]

[0162]

[0163]

[0164] Wherein, TP is the number of samples that were actually demagnetized but were classified as demagnetized, FN is the number of samples that were actually demagnetized but were classified as not demagnetized, TN is the number of samples that were actually not demagnetized but were classified as not demagnetized, and FP is the number of samples that were actually not demagnetized but were classified as demagnetized. The mean accuracy under different time windows is shown in Table 1. Figure 4 The convergence curves are shown for time windows of 0.5s and 1.5s.

[0165] Table 1. Mean accuracy under different time windows

[0166] Time window <![CDATA[ACC v.av (%)]]> <![CDATA[ACC t (%)]]> <![CDATA[PRE t (%)]]> <![CDATA[REC t (%)]]> <![CDATA[F 1.t ]]> 0.5s 99.88 99.38 98.00 100 98.99 1.5s 100 100 100 100 100

[0167] Correspondingly, the hyperparameters of MKL-TSVM in the 0.5s and 1.5s time windows are as follows:

[0168] ε 0.5s =0.1787164, C 1,0.5s =112.9767, C 2,0.5s =385.6, γ 0.5s =58.76644, ε 1.5s =0.2178961, C 1,1.5s =122.0703, C 2,1.5s =235.0703, γ 1.5s =87.52403.

[0169] In actual operation of gas generator demagnetization identification, real-time acquisition of generator terminal impedance change information is performed; when protection is activated, motion timing features are extracted from the measured impedance time-series spatial change trajectory; t is calculated. i The feature vector within a ±0.1s time window is input into the discrimination model, and an adaptive time-window discrimination strategy based on kernel space classification distance is used to complete the demagnetization identification. The overall process of the gas generator demagnetization identification method is as follows: Figure 5 As shown.

[0170] Specifically, taking the training model's time window length as the center, two time windows are taken with a 0.1s interval before and after, respectively. The selected features are obtained, and the kernel space classification average distance is calculated as follows:

[0171]

[0172] If |L|>0.5, the judgment result is output directly;

[0173] If |L|<0.5, then the model trained with a longer time window is used to take two time windows with a 0.1s interval before and after, respectively, to obtain the selected features and re-discriminate.

[0174] In real-time applications, after each action decision is completed, the result is verified and fed back to the sample set to achieve regular offline learning, thereby improving the performance of the discrimination model.

[0175] It should be noted that this invention uses 1 / 5 of all samples as test samples to test the effect of the real-time application. The recognition performance of the adaptive recognition model is shown in Table 2.

[0176] Table 2 shows the adaptive time window recognition performance on the test set.

[0177] Identification process ACC (%) REC(%) PRE(%) F1 score first step 98.57 96.15 100 98.03 Step 2 100 100 100 100

[0178] Figure 6 and Figure 7 For the identification process on the test set (num=3), the feature calculation time windows for the two steps are 0.4-0.6s and 1.4-1.6s, respectively, with a time window interval of 0.1s. The accuracy of the first-step training model is less than 100% due to large trajectory differences. In this case, an adaptive time-window demagnetization discrimination strategy based on the average classification distance in kernel space can be used to improve the model's reliability. In the first step, 4 demagnetization samples and 18 perturbation / oscillation samples are located within the critical region, and 2 demagnetization samples are missed. When the samples are input into the long-time-window model for the second discrimination, some samples that were far from the class boundaries in kernel space during the first step can now be moved closer to the boundaries, and the missed demagnetization samples can also be corrected.

[0179] To better demonstrate the superiority of the proposed loss-of-excitation scheme, a comparative test was conducted on the verification set using a traditional loss-of-excitation protection system with apple-shaped impedance operation characteristics as the main criterion. The protection system adopts the widely used AND gate output logic. The operation and timing of the traditional protection and the proposed protection are shown in Table 3.

[0180] Table 3. Actions and time taken for traditional protection and protection by this invention.

[0181]

[0182] As can be seen, the average accuracy can reach 100% when using a dual-time-window discrimination strategy based on the distance of the classification function, that is, the loss of excitation fault can be detected within 1.6s. Compared with the operation of traditional loss of excitation protection, the speed and reliability of the loss of excitation protection proposed in this invention are greatly improved.

[0183] This invention breaks through the bottleneck of traditional loss of excitation protection, resolves the contradiction between speed and selectivity, and has high reliability and excellent generalization ability. It still has excellent adaptability when facing complex changes in the power grid.

[0184] Example 2

[0185] In order to implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, a gas generator demagnetization identification system is provided below.

[0186] like Figure 8 As shown, the gas generator demagnetization identification system provided in this embodiment includes: an impedance trajectory acquisition unit 101, a feature vector calculation unit 102, and a demagnetization identification unit 103.

[0187] The impedance trajectory acquisition unit 101 is used to acquire the measured impedance trajectory of the gas generator. The measured impedance trajectory includes the measured impedance time sequence points at each moment within a set time period.

[0188] The feature vector calculation unit 102 is connected to the impedance trajectory acquisition unit 101. The feature vector calculation unit 102 is used to determine the current feature vector based on the measured impedance trajectory.

[0189] The demagnetization identification unit 103 is connected to the feature vector calculation unit 102. The demagnetization identification unit 103 is used to determine whether the gas generator has lost its magnetism based on the current feature vector and a discriminant model. The discriminant model is obtained by training a multi-core twin support vector machine using a pre-set training sample set and based on the ocean predator algorithm, which integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning. The training sample set includes feature vectors from multiple demagnetization samples and feature vectors from multiple non-demagnetization samples.

[0190] Example 3

[0191] This embodiment provides an electronic device, including a memory and a processor. The memory is used to store computer programs, and the processor runs the computer programs to enable the electronic device to execute the gas generator demagnetization identification method of Embodiment 1.

[0192] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0193] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying the loss of excitation in a gas generator, characterized in that, The method for identifying the demagnetization of a gas generator includes: The measurement impedance trajectory of the gas generator is collected; the measurement impedance trajectory includes the measurement impedance time sequence points at each moment within a set time period; Based on the measured impedance trajectory, determine the current feature vector; Based on the current feature vector, it is determined whether the gas generator has lost its magnetism using a discriminant model. The discriminant model is obtained by training a multi-core twin support vector machine using a pre-trained sample set and a marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning. The training sample set includes feature vectors of multiple demagnetized samples and feature vectors of multiple non-demagnetized samples. The method for establishing the discriminant model includes: Obtain a measurement impedance trajectory sample set of the gas generator; the measurement impedance trajectory sample set includes multiple demagnetization samples and multiple non-demagnetization samples; each demagnetization sample and each non-demagnetization sample includes the measurement impedance time sequence points at each moment within a set time period; The measurement impedance timing points of each demagnetized sample and each non-demagnetized sample within a set time window are extracted to obtain the measurement impedance trajectory sample set within the corresponding time window. Based on the set of measured impedance trajectory samples within the set time window, calculate the feature vector of each demagnetization sample and the feature vector of each non-demagnetization sample. Determine the demagnetization normalization matrix based on the feature vectors of each demagnetization sample; Determine the non-demagnetization normalization matrix based on the feature vectors of each non-demagnetization sample; Based on the demagnetization normalized matrix and the non-demagnetization normalized matrix, the multi-core twin support vector machine is iteratively trained using the marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimension learning to obtain the discriminant model. Specifically, based on the demagnetization normalized matrix and the non-demagnetization normalized matrix, and using a marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning, the multi-core twin support vector machine is iteratively trained to obtain a discriminative model, including: A chaotic opposition strategy is used for population initialization to determine the initial position and opposition position of each prey in the population, so as to obtain the initial population. The initial positions of each prey in the initial population are iteratively updated using the marine predator algorithm to obtain the updated positions of each prey. The adaptive t-distribution is determined based on the number of iterations of the marine predator algorithm; Based on the adaptive t-distribution, the update position of each prey is mutated to determine the corresponding mutated position of each prey. The fitness of each prey is determined based on the location of its mutations, and the prey is then divided into an elite group and a learning group based on their fitness. Based on the mutation positions of the elite group and the learning group, the mutation positions of each prey in the learning group are cross-referenced to obtain the optimal position of each prey; the optimal position of each prey includes the first upper bound penalty coefficient, the second upper bound penalty coefficient, the kernel function weight, and the radial basis kernel hyperparameter. Based on the optimal position of each prey, the hyperparameters of the multi-core twin support vector machine are trained and optimized according to the demagnetization normalization matrix and the non-demagnetization normalization matrix to obtain the discrimination model. The hyperparameters of the multi-core twin support vector machine include the normal vector of the demagnetized sample in the kernel space, the normal vector of the non-demagnetized sample in the kernel space, the bias coefficient of the demagnetized sample, and the bias coefficient of the non-demagnetized sample under the set time window.

2. The method for identifying the demagnetization of a gas generator according to claim 1, characterized in that, The method for identifying the demagnetization of a gas generator also includes: If the gas generator loses its magnetism, a demagnetization protection action will be performed on the gas generator.

3. The method for identifying the loss of magnetism of a gas generator according to claim 1, characterized in that, The current feature vector includes the distance variance, the minimum value of the derivative of the motion azimuth angle, the mean value of the derivative of the motion azimuth angle, the mean value of the derivative of the direction angle, the maximum value of the velocity, and the velocity skewness; The step of determining the current feature vector based on the measured impedance trajectory specifically includes: Extract the timing points of the measured impedance trajectory within the set time window; For any measurement impedance timing point, calculate the motion azimuth angle, direction angle, and velocity of the measurement impedance timing point based on the measurement impedance timing point; Calculate the distance variance based on the timing points of each measured impedance; Calculate the minimum and average values ​​of the derivatives of the azimuth angles of motion based on the azimuth angles of motion at each time point of the measured impedance. Calculate the mean value of the derivative of the direction angle based on the direction angle at each measured impedance timing point; Calculate the maximum velocity and velocity deviation based on the velocity at each measured impedance timing point.

4. The method for identifying the demagnetization of a gas generator according to claim 3, characterized in that, The following formula is used to calculate the first... i Distance variance of each time window: ； in, For the first i The distance variance of each time window N i For the first i The number of timing points for measuring impedance in each time window. For the first k One measurement impedance timing point, It is the reciprocal of the per-unit value of the gas turbine output power.

5. The method for identifying the demagnetization of a gas generator according to claim 3, characterized in that, The following formula is used to calculate the first... i Minimum and mean values ​​of the derivative of the azimuth angle of motion for each time window: ； ； in, For the first i Minimum value of the derivative of the motion azimuth angle of each time window For the first i Mean of the derivative of the azimuth angle of motion for each time window N i For the first i The number of timing points for measuring impedance in each time window. For the first k The motion azimuth angle of each impedance timing point For the first k +1 motion azimuth angle of the impedance measurement timing point This is for calculating the time interval.

6. The method for identifying the demagnetization of a gas generator according to claim 1, characterized in that, The discriminant model is as follows: ； in, f ( X num The result is the judgment result. f ( X num )=1 indicates that the corresponding gas generator has lost its magnetism. f ( X num )=-1 indicates that the corresponding gas generator is not demagnetized. num This represents the total number of demagnetized and non-demagnetized samples. X num The characteristic matrix includes the demagnetization normalized matrix and the non-demagnetization normalized matrix. x j The first in the characteristic matrix j There are eigenvectors, and sgn() is the sign function. ω 1,i For the first i The normal vector of the demagnetized sample in kernel space under each time window ω 2,i For the first i The normal vector of the non-demagnetized sample in kernel space under each time window b 1,i For the first i Bias coefficient of demagnetized samples under each time window, b 2,i For the first i Bias coefficients of non-demagnetized samples under each time window.

7. A gas generator demagnetization identification system, used to implement the gas generator demagnetization identification method according to any one of claims 1-6, characterized in that, The gas generator demagnetization identification system includes: Impedance trajectory acquisition unit is used to acquire the measured impedance trajectory of the gas generator; the measured impedance trajectory includes the measured impedance time sequence points at each moment within a set time period; The feature vector calculation unit, connected to the impedance trajectory acquisition unit, is used to determine the current feature vector based on the measured impedance trajectory. The demagnetization identification unit, connected to the feature vector calculation unit, is used to determine whether the gas generator has lost its magnetism based on the current feature vector and a discrimination model. The discrimination model is obtained by training a multi-core twin support vector machine using a pre-trained sample set and a marine predator algorithm that integrates chaotic opposition, adaptive t-distribution, and grouped dimensionality learning. The training sample set includes feature vectors of multiple demagnetized samples and feature vectors of multiple non-demagnetized samples.

8. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the gas generator demagnetization identification method according to any one of claims 1 to 6.

Images