Power distribution system online intelligent fault detection method based on RRBF neural network

Abstract

The invention discloses a power distribution system online intelligent fault detection method based on an RRBF neural network. The method comprises the following steps: selecting a dynamic state estimation model of a power distribution system AC signal; the method comprises the following steps: in a three-phase power distribution system in a noise interference environment, calculating a phasor amplitude and a phase angle of synchronous phasor information, and preprocessing the calculated synchronous phasor information; performing classification label management on the preprocessed synchronous phasor measurement information; providing a regularization RRBF neural network model under the condition of supervising a multi-model residual error generation classification learning framework; the regularization RRBF neural network model is used to generate a residual error for the power distribution system area, and then online intelligent diagnosis is carried out on the fault area of the power distribution system according to the generated residual error. According to the method, online intelligent diagnosis is carried out on the fault area of the three-phase power distribution system in the noise environment, the generalization ability of the RRBF neural network is greatly improved, the accuracy of online fault diagnosis of the power distribution system is improved, and the response time is greatly shortened.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution system fault detection, and in particular to an online intelligent fault detection method for a power distribution system based on a RRBF neural network. Background Art

[0002] As a high-incidence area of ​​short-circuit faults, fault detection has always been the focus of ensuring the safe operation of the distribution system. Distribution system faults are emergencies, and their sample information is limited. Due to the lack of available labeled training samples and the complexity of the feature subset selection process, noise and random fluctuations in the feature vector space are prone to overfitting problems, and the generalization performance of radial basis neural networks in data feature learning is affected. However, in the existing technologies for fault detection of distribution systems based on neural networks, the generalization ability of neural networks has not been effectively improved, and these fault detection technologies also have a series of various defects.

[0003] According to the published patent documents, publication number: CN110649597A, a distribution network feeder automation control method based on radial basis function neural network is disclosed, including: distribution network fault characteristic analysis, discernible matrix simplification and k-means clustering method; the method determines the characteristic input layer and target output layer of the RBF neural network model to be built according to the distribution network fault characteristics, uses the discernible matrix simplification method to simplify the distribution network fault characteristic data, and uses the various types of distribution network fault data obtained after simplification as the training sample data of the RBF neural network model to construct a multi-input and multi-output distribution network fault RBF neural network model; the method inputs the line terminal signal, uses the radial basis function neural network to automatically determine the fault interval, automatically controls the switch separation and closing to realize fault isolation and realize automatic distributed feeder line fault diagnosis; this invention uses the low-dimensional training data set obtained by discernible matrix simplification to construct the RBF neural network model for the low-dimensional training set The fault characteristics of the distribution network are analyzed, but the nonlinear dynamic fault process of the actual distribution system on site is not considered; Publication No.: CN109494726A, discloses an online diagnosis method for power systems based on a DLRNN neural network, the detection method described in this invention collects voltage and current data of the power system, including steady-state data and fault data; the DLRNN neural network model structure is designed according to the system stability requirements and the DLRNN neural network is trained offline, and the offline trained neural network is used to monitor the power grid operation status in real time; the final analysis results show that the DLRNN neural network method uses artificial neural networks with good ability to handle complex unknown nonlinear systems, and uses DLRNN neural network parameters and real-time power grid operation data to meet online diagnosis and analysis requirements; however, no data preprocessing is performed on the synchronous phasor information collected by the on-site system, and the interference of on-site noise on the real-time power grid operation data in the actual system is not considered. Publication No.: CN113159345A, discloses a power grid fault identification method and system based on a fusion neural network model, including: obtaining power parameters of the power grid operation state as characteristic values ​​and preprocessing them, and inputting the preprocessed characteristic values ​​into the fusion neural network model to calculate the fault probability of the fault line; in view of the influence of signal transmission noise, the power grid fault identification method based on the fusion neural network model uses the advantages of artificial intelligence to mine the correlation between operation data and fault characteristics, and conducts comprehensive diagnosis of the power grid through multi-dimensional information, thereby reducing the influence of signal transmission noise or error on fault identification accuracy and improving the safety of power grid operation; however, noise and random fluctuations in the characteristic vector space will cause overfitting, and the generalization ability of the neural network needs to be improved. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0005] To this end, the purpose of the present invention is to propose an online intelligent fault detection method for a distribution system based on a RRBF neural network.

[0006] In order to achieve the above-mentioned purpose, the technical scheme of the present invention provides an online intelligent fault detection method for a distribution system based on an RRBF neural network, the method comprising: step S1: selecting a second-order model of a distribution system as a dynamic state estimation model of an AC signal of the distribution system; step S2: in a three-phase distribution system under a noise interference environment, using an extended Kalman filter method to calculate the phasor amplitude and phase angle of synchronized phasor information, and preprocessing the calculated synchronized phasor information to filter out a mixed signal of high-order harmonic components and white noise to form an initial state phasor information describing the state of the entire distribution system; step S3: classifying and labeling the preprocessed synchronized phasor measurement information; step S4: proposing a regularized RRBF neural network model under a supervised multi-model residual generation classification learning framework; wherein the regularized RRBF neural network model comprises an input layer, a hidden layer, a recursive layer, and an output layer that are sequentially connected to each other; step S5: using the regularized RRBF neural network model to generate residuals for a distribution system area, and then performing online intelligent diagnosis on the fault area of ​​the distribution system based on the generated residuals.

[0007] Preferably, the step S1 specifically includes: step S1.1: using a GPS receiving engine to receive a synchronous clock source for unified timing, and then concentrating the synchronized phasor information uploaded by PMUs at different geographical locations in a unified time coordinate system to obtain a mathematical expression of a dynamic state estimation model of the AC signal of the power distribution system; the mathematical expression of the dynamic state estimation model of the AC signal of the power distribution system is:

[0008] x(t)=Xcos(2πft+φ) (1)

[0009] In formula (1), x(t) is the AC signal of the distribution system; f is the rated frequency of the distribution system, t is the sampling time, X is the phase amplitude, and φ is the angle between the first sampling point of the AC signal of the distribution system and the peak value of the input signal.

[0010] Preferably, in step S2, the phasor amplitude and phase angle are extracted from the voltage phasor and the current phasor to obtain mathematical expressions of the phasor amplitude X and the phase angle φ;

[0011] The mathematical expressions of the phase amplitude X and the phase angle φ are respectively:

[0012]

[0013] φ=ac tan(X q / X i ) (3)

[0014] In formula (2) and (3), X i is the i-axis component of the phase X, X q is the q-axis component of the phase X;

[0015] Among them, X i and X q The mathematical expressions are:

[0016]

[0017] In equations (4) and (5), s is the complex variable in Laplace transform; ω is the resonant angular frequency, ω = 2πf; k i ,k q is the extended Kalman filter gain.

[0018] Preferably, the step S3 specifically includes: step S3.1: uploading the synchronized phasor information according to the local node dispatching control center; step S3.2: using the phasor amplitude and phase angle calculated by the extended Kalman filter method to form a sample data set; step S3.3: randomly dividing the sample data set into a training set and a test set; wherein the training set is composed of random fault samples occurring on each branch line in the distribution system, and the test set is used to test the evaluation performance of the regularized RRBF neural network model; step S3.4: normalizing the preprocessed phasor amplitude and phase angle, and establishing the input variables of the regularized RRBF neural network model; step S3.5: establishing the output variables of the regularized RRBF neural network model according to the operating status of the branch in the distribution system, so as to realize the classification label management of the preprocessed synchronized phasor measurement information;

[0019] Wherein, the mathematical expression of the input vector is:

[0020]

[0021] In formula (6), U n The voltage phasor amplitude information measured for the nth node, arg(U n ) is the phase angle information of the voltage phasor measured by the nth node; I n The amplitude information of the current phasor measured for the nth node, arg(I n ) is the phase angle information of the current phasor measured by the nth node, and n is the number of nodes in the distribution system;

[0022] Wherein, the mathematical expression of the output vector is:

[0023]

[0024] In formula (7), q i∈[0,1], i=1,2,...n h , n h is the total number of distribution branches in the distribution system; the operating status of each branch is binary-coded, and the number of output vector components is equal to the number of categories, 1 represents a fault in the corresponding branch interval, and 0 represents no fault in the corresponding branch interval.

[0025] Preferably, the step S4 specifically comprises: step S4.1: describing and implementing a mapping f from a p-dimensional input space to a one-dimensional output space; wherein the mapping f is expressed as: R P →R 1 , this mapping relationship is described by the hidden layer neural network model, and this mapping relationship is defined by the activation function set φ on the neuron set x j and the corresponding weight set θ on the neuron ji The linear combination of ;

[0026] Step S4.2: Obtain a mathematical expression of the output vector of the regularized RRBF neural network model through the mapping f; the mathematical expression of the output vector of the regularized RRBF neural network model is:

[0027]

[0028] In formula (8), f i (x) is the output vector; x is the input vector, x=[x1,x2,…x m ], m is the input vector dimension; C j is the center of the hidden layer neuron, C j =[c 1,j ,c 2,j ,…,c m,j ], 1 <j<n;θ ji is the weight from the jth hidden layer neuron to the ith output neuron, 1 <i<n h , n h is the number of output layers; ||·|| is the Euclidean norm from the hidden layer nodes to the output layer nodes;

[0029] Step S4.3: Select the activation function φ of the regularized RRBF neural network model j ; The step S4.3 specifically includes: transmitting the input vector x to the radial basis function neuron layer through the input layer, and then calculating the input vector x and the neuron center C of the jth neuron in the hidden layer j The distance between them, and the distance as the corresponding radial basis excitation function φ j (x) input, then select the activation function φ j is a Gaussian function;

[0030] Among them, the radial basis excitation function φ j The mathematical expression of (x) is as follows:

[0031]

[0032] In formula (9), σ j is the width parameter and satisfies σ j >0;

[0033] Expected output d i (t) is:

[0034]

[0035] In formula (10), ε i (t) is the training set d i (t) and the output vector f i The error between (x(t)), error function ε i (t) is used to measure the difference between the network output and the expected output;

[0036] Step S4.4: Use the zero-order regularized error function to correct the error function ε of equation (10) i (t) is corrected to establish the regularized error term of the regularized RRBF neural network as:

[0037]

[0038] Step S4.5: During the training of the regularized RRBF neural network model, the penalty term λg T g constrains the orthogonal weight vector g in the regularized RRBF neural network model, and the regularized error function is obtained as follows:

[0039]

[0040] In formula (12), the regularization parameter λ is used to balance the fitting degree and complexity of the regularized RRBF neural network model; is the correlation term between the expected output variance and the regression operator. Each vector w j The contribution to the variance increment is

[0041] Step S4.6: The standard form of the regularized error function is obtained by equation (12):

[0042]

[0043] Step S4.7: Obtain the regularized error rate ρ of the regression matrix through equation (13) j for:

[0044]

[0045] Step S4.8: Regularized error rate ρ based on the regression matrix j , establish the accuracy constraint condition of the forward regression process of the regularized RRBF neural network; the mathematical expression of the accuracy constraint condition is:

[0046]

[0047] In formula (15), the regularized error rate ρ of the regression matrix j Used to control the network structure complexity of the regularized RRBF neural network and update the structure of the regularized RRBF neural network online; ξ is the accuracy threshold;

[0048] Step S4.9: In the process of training the regularized RRBF neural network model, the training accuracy threshold ξ is selected at each step. j The corresponding vector with the largest index value is used as a hidden layer node, and the vector of the previous step is orthogonalized until the accuracy threshold ξ satisfies the accuracy constraint condition in the forward regression process of the regularized RRBF neural network.

[0049] Preferably, the step S5 specifically comprises: step S5.1: dividing the power distribution system into a plurality of regions including overlapping tie lines, and extracting historical synchronized phasor information from each divided region; step S5.2: performing fault detection on each divided region using the regularized RRBF neural network model, and completing the fault detection task of each region in an independent parallel form; step S5.3: setting the output matrix of the RRBF fault region detection model to Q=[q1 q2…q j …q n ], the output matrix is ​​abbreviated as q * , the output residual is r = [q1-q * q2-q * …q n -q * ];

[0050] Step S5.4: The decision logic for establishing the RRBF fault area detection model is as follows:

[0051]

[0052] In formula (16), i is the node of the distribution system, is the absolute value of the residual vector corresponding to the i-th node, δ i (i=1, 2…n) is a threshold value associated with the residual output of the RRBF fault area detection model; Specifically refers to the node health status of the system at each moment; T s is the sampling time of the power distribution system; It is the time to detect and isolate the fault that occurred;

[0053] Step S5.5: Determine whether r1 is greater than δ i ; When the judgment result is yes, it is determined that the distribution line i is faulty, so as to detect the fault area of ​​the distribution system by using the synchronous phasor information of the distribution system under the supervised multi-model residual generation classification learning framework.

[0054] Beneficial effects of the present invention:

[0055] (1) In the online intelligent fault detection method for a power distribution system based on an RRBF neural network of the present invention, an extended Kalman filter method is used to filter out mixed signals such as harmonic components and white noise in the PMU information processing link, thereby improving the information measurement accuracy.

[0056] (2) The online intelligent fault detection method for the distribution system based on the RRBF neural network of the present invention extracts the amplitude and phase angle information of the synchronous phasor measurement information, performs classification label management to form a training data set, and improves the accuracy of the training set data.

[0057] (3) The online intelligent fault detection method for a power distribution system based on an RRBF neural network of the present invention divides the power distribution system into a number of areas containing overlapping interconnection lines, establishes a local area fault diagnosis model, and reduces the computational complexity of the method.

[0058] (4) The online intelligent fault detection method for the power distribution system based on the RRBF neural network of the present invention uses a zero-order regularized error function to correct the error function and constrain the complexity of the RRBF neural network model.

[0059] (5) The online intelligent fault detection method for distribution system based on RRBF neural network of the present invention uses regularization parameters to balance the model fitting degree and model complexity, thereby reducing the risk of overfitting of the neural network.

[0060] (6) The online intelligent fault detection method for distribution system based on RRBF neural network of the present invention adopts regularized error rate accuracy constraint condition to select hidden center vector and update network structure, thereby reducing network scale and improving the generalization ability of RRBF neural network.

[0061] (7) The online intelligent fault detection method for a power distribution system based on an RRBF neural network of the present invention establishes the decision logic of the RRBF fault detection model for common faults in the power distribution system, thereby improving the accuracy of state estimation.

[0062] (8) The online intelligent fault detection method for the power distribution system based on the RRBF neural network of the present invention uses an independent parallel form to complete the detection tasks of each area to improve the fault detection efficiency, without the need to store a large amount of historical data, and taking into account both calculation speed and accuracy.

[0063] Additional aspects and advantages of the invention will become apparent from the following description, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic flow chart of an online intelligent fault detection method for a power distribution system based on a RRBF neural network according to an embodiment of the present invention is shown;

[0065] Figure 2 A schematic flow chart of an online intelligent fault detection method for a power distribution system based on a RRBF neural network according to another embodiment of the present invention is shown;

[0066] Figure 3 A structural diagram showing a node topology of an IEEE 13-node power system according to an embodiment of the present invention;

[0067] Figure 4 A flowchart showing a supervised multi-model residual generation and classification learning framework according to an embodiment of the present invention is shown;

[0068] Figure 5 The output diagram of the simulation experiment for intelligent diagnosis of a single-phase grounding fault in branch 680-671 according to an embodiment of the present invention is shown;

[0069] Figure 6 The output diagram of the simulation experiment of online intelligent fault diagnosis of branch 634-633 according to one embodiment of the present invention is shown;

[0070] Figure 7 The output diagram of the simulation experiment of online intelligent fault diagnosis performed by the node 634 according to one embodiment of the present invention is shown;

[0071] Figure 8 A simulation experiment output diagram of a common short circuit fault diagnosis according to an embodiment of the present invention is shown;

[0072] Fig. 9 The figure shows the simulation experiment output diagram of the RRBF diagnosis model under 20dB and 30dB noise levels according to one embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0074] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0075] Figure 1 FIG. 1 is a schematic flow chart showing an online intelligent fault detection method for a power distribution system based on an RRBF neural network according to an embodiment of the present invention. Figure 1 As shown, the online intelligent fault detection method of the distribution system based on the RRBF neural network includes:

[0076] Step S1: Selecting a second-order model of a power distribution system as a dynamic state estimation model of an AC signal of the power distribution system;

[0077] Step S2: In a three-phase power distribution system under a noise interference environment, the phasor amplitude and phase angle of the synchronized phasor information are calculated using the extended Kalman filter method, and the calculated synchronized phasor information is preprocessed to filter out the mixed signal of the high-order harmonic component and the white noise to form the initial state phasor information describing the state of the entire power distribution system;

[0078] Step S3: classify and label the preprocessed synchronized phasor measurement information;

[0079] Step S4: Under the supervised multi-model residual generation classification learning framework, a regularized RRBF neural network model is proposed; wherein the regularized RRBF neural network model includes an input layer, a hidden layer, a recursive layer, and an output layer which are sequentially connected to each other;

[0080] Step S5: Generate residuals for the distribution system area using the regularized RRBF neural network model, and then perform online intelligent diagnosis on the fault area of ​​the distribution system based on the generated residuals.

[0081] In this embodiment, under the supervised multi-model residual generation classification learning framework, in order to solve the overfitting problem in the online fault detection process of the distribution system, the present invention proposes an online intelligent fault detection method for the distribution system based on RRBF neural network by utilizing artificial intelligence principles, neural network models and network optimization theory. The method can effectively constrain the complexity of the neural network model, improve the generalization ability of the radial basis function neural network, intelligently identify the fault area and fault phase under different types of short-circuit faults, and improve the accuracy of online fault diagnosis of the distribution system, which has great practical significance for improving the stability of industrial operation and the quality of power supply to users.

[0082] In one embodiment of the present invention, the step S1 specifically includes: step S1.1: using the GPS receiving engine to receive the synchronous clock source for unified timing, and then concentrating the synchronous phasor information uploaded by the PMUs in different geographical locations in a unified time coordinate system to obtain the mathematical expression of the dynamic state estimation model of the AC signal of the distribution system; the mathematical expression of the dynamic state estimation model of the AC signal of the distribution system is:

[0083] x(t)=Xcos(2πft+φ) (1)

[0084] In formula (1), x(t) is the AC signal of the distribution system; f is the rated frequency of the distribution system, t is the sampling time, X is the phase amplitude, and φ is the angle between the first sampling point of the AC signal of the distribution system and the peak value of the input signal.

[0085] In this embodiment, under the supervised multi-model residual generation classification learning framework, the master station uses the GPS receiving engine to receive the synchronous clock source for unified timing, and then concentrates the synchronous phasor information uploaded by PMUs in different geographical locations in a unified time coordinate system to achieve the purpose of establishing the global initial state phasor information of the distribution system.

[0086] In one embodiment of the present invention, in step S2, the phasor amplitude and phase angle are extracted from the voltage phasor and the current phasor to obtain mathematical expressions of the phasor amplitude X and the phase angle φ;

[0087] The mathematical expressions of the phase amplitude X and the phase angle φ are respectively:

[0088]

[0089] φ=ac tan(X q / X i ) (3)

[0090] In formula (2) and (3), X i is the i-axis component of the phase X, X q is the q-axis component of the phase X;

[0091] Among them, X i and X q The mathematical expressions are:

[0092]

[0093] In equations (4) and (5), s is the complex variable in Laplace transform; ω is the resonant angular frequency, ω = 2πf; k i ,k q is the extended Kalman filter gain.

[0094] In one embodiment of the present invention, the step S3 specifically includes: step S3.1: uploading the synchronized phasor information according to the local node dispatching control center; step S3.2: using the phasor amplitude and phase angle calculated by the extended Kalman filter method to form a sample data set; step S3.3: randomly dividing the sample data set into a training set and a test set; wherein the training set is composed of random fault samples occurring on each branch line in the distribution system, and the test set is used to test the evaluation performance of the regularized RRBF neural network model; step S3.4: normalizing the preprocessed phasor amplitude and phase angle, and establishing the input variables of the regularized RRBF neural network model; step S3.5: establishing the output variables of the regularized RRBF neural network model according to the operating status of the branch in the distribution system, so as to realize the classification label management of the preprocessed synchronized phasor measurement information;

[0095] Wherein, the mathematical expression of the input vector is:

[0096]

[0097] In formula (6), U n The voltage phasor amplitude information measured for the nth node, arg(U n ) is the phase angle information of the voltage phasor measured by the nth node; I n The amplitude information of the current phasor measured for the nth node, arg(I n ) is the phase angle information of the current phasor measured by the nth node, and n is the number of nodes in the distribution system;

[0098] Wherein, the mathematical expression of the output vector is:

[0099]

[0100] In formula (7), q i ∈[0,1], i=1,2,...n h , n his the total number of distribution branches in the distribution system; the operating status of each branch is binary-coded, and the number of output vector components is equal to the number of categories, 1 represents a fault in the corresponding branch interval, and 0 represents no fault in the corresponding branch interval.

[0101] In one embodiment of the present invention, the step S4 specifically includes: step S4.1: describing and implementing a mapping f from a p-dimensional input space to a one-dimensional output space; wherein the mapping f is expressed as: R P →R 1 , this mapping relationship is described by the hidden layer neural network model, and this mapping relationship is defined by the activation function set φ on the neuron set x j and the corresponding weight set θ on the neuron ji The linear combination of ;

[0102] Step S4.2: Obtain a mathematical expression of the output vector of the regularized RRBF neural network model through the mapping f; the mathematical expression of the output vector of the regularized RRBF neural network model is:

[0103]

[0104] In formula (8), f i (x) is the output vector; x is the input vector, x=[x1,x2,…x m ], m is the input vector dimension; C j is the center of the hidden layer neuron, C j =[c 1,j ,c 2,j ,…,c m,j ], 1 <j<n;θ ji is the weight from the jth hidden layer neuron to the ith output neuron, 1 <i<n h , n h is the number of output layers; ||·|| is the Euclidean norm from the hidden layer nodes to the output layer nodes;

[0105] Step S4.3: Select the activation function φ of the regularized RRBF neural network model j ; The step S4.3 specifically includes: transmitting the input vector x to the radial basis function neuron layer through the input layer, and then calculating the input vector x and the neuron center C of the jth neuron in the hidden layer j The distance between them, and the distance as the corresponding radial basis excitation function φ j (x) input, then select the activation function φ j is a Gaussian function;

[0106] Among them, the radial basis excitation function φ jThe mathematical expression of (x) is as follows:

[0107]

[0108] In formula (9), σ j is the width parameter and satisfies σ j >0;

[0109] Expected output d i (t) is:

[0110]

[0111] In formula (10), ε i (t) is the training set d i (t) and the output vector f i The error between (x(t)), error function ε i (t) is used to measure the difference between the network output and the expected output;

[0112] Step S4.4: Use the zero-order regularized error function to correct the error function ε of equation (10) i (t) is corrected to establish the regularized error term of the regularized RRBF neural network as:

[0113]

[0114] Step S4.5: During the training of the regularized RRBF neural network model, the penalty term λg T g constrains the orthogonal weight vector g in the regularized RRBF neural network model, and the regularized error function is obtained as follows:

[0115]

[0116] In formula (12), the regularization parameter λ is used to balance the fitting degree and complexity of the regularized RRBF neural network model; is the correlation term between the expected output variance and the regression operator. Each vector w j The contribution to the variance increment is

[0117] Step S4.6: The standard form of the regularized error function is obtained by equation (12):

[0118]

[0119] Step S4.7: Obtain the regularized error rate ρ of the regression matrix through equation (13) j for:

[0120]

[0121] Step S4.8: Regularized error rate ρ based on the regression matrix j , establish the accuracy constraint condition of the forward regression process of the regularized RRBF neural network; the mathematical expression of the accuracy constraint condition is:

[0122]

[0123] In formula (15), the regularized error rate ρ of the regression matrix j Used to control the network structure complexity of the regularized RRBF neural network and update the structure of the regularized RRBF neural network online; ξ is the accuracy threshold;

[0124] Step S4.9: In the process of training the regularized RRBF neural network model, the training accuracy threshold ξ is selected at each step. j The corresponding vector with the largest index value is used as a hidden layer node, and the vector of the previous step is orthogonalized until the accuracy threshold ξ satisfies the accuracy constraint condition in the forward regression process of the regularized RRBF neural network.

[0125] In one embodiment of the present invention, the step S5 specifically includes: step S5.1: dividing the power distribution system into a plurality of regions including overlapping tie lines, and extracting historical synchronized phasor information from each divided region; step S5.2: performing fault detection on each divided region using the regularized RRBF neural network model, and completing the fault detection task of each region in an independent parallel form; step S5.3: setting the output matrix of the RRBF fault region detection model to Q=[q1 q2…q j …q n ], the output matrix is ​​abbreviated as q * , the output residual is r = [q1-q * q2-q * …q n -q * ];

[0126] Step S5.4: The decision logic for establishing the RRBF fault area detection model is as follows:

[0127]

[0128] In formula (16), i is the node of the distribution system, is the absolute value of the residual vector corresponding to the i-th node, δ i (i=1, 2…n) is a threshold value associated with the residual output of the RRBF fault area detection model; Specifically refers to the node health status of the system at each moment; T sis the sampling time of the power distribution system; It is the time to detect and isolate the fault that occurred;

[0129] Step S5.5: Determine whether r1 is greater than δ i ; When the judgment result is yes, it is determined that the distribution line i is faulty, so as to detect the fault area of ​​the distribution system by using the synchronous phasor information of the distribution system under the supervised multi-model residual generation classification learning framework.

[0130] The technical solution of the present invention will be demonstrated with a specific embodiment below. Figure 2 As shown, the online intelligent fault detection method for distribution system based on RRBF neural network of the present invention is mainly implemented in two steps: first, offline training model, that is, offline training of the established regularized RRBF neural network model, when the accuracy threshold ξ satisfies the accuracy constraint condition in the positive regression process of the regularized RRBF neural network, so as to achieve the purpose of establishing the RRBF fault diagnosis model; second, online fault diagnosis, that is, using the regularized RRBF neural network model to generate residuals for the distribution system area, and then performing online intelligent diagnosis on the fault area of ​​the distribution system according to the generated residuals.

[0131] 1. The specific steps for offline training model are as follows:

[0132] (1) Step S1: Selecting a second-order model of the power distribution system as a dynamic state estimation model of the AC signal of the power distribution system;

[0133] The step S1 specifically includes:

[0134] Step S1.1: Use the GPS receiving engine to receive the synchronous clock source for unified timing, and then centralize the synchronized phasor information uploaded by PMUs in different geographical locations in a unified time coordinate system to obtain the mathematical expression of the dynamic state estimation model of the AC signal of the distribution system; the mathematical expression of the dynamic state estimation model of the AC signal of the distribution system is:

[0135] x(t)=Xcos(2πft+φ) (1)

[0136] In formula (1), x(t) is the AC signal of the distribution system; f is the rated frequency of the distribution system, t is the sampling time, X is the phase amplitude, and φ is the angle between the first sampling point of the AC signal of the distribution system and the peak value of the input signal.

[0137] (2) Step S2: In a three-phase power distribution system under a noise interference environment, the phasor amplitude and phase angle of the synchronized phasor information are calculated using an extended Kalman filter method, and the calculated synchronized phasor information is preprocessed to filter out a mixed signal of high-order harmonic components and white noise to form initial state phasor information describing the state of the entire power distribution system;

[0138] In the step S2, the phasor amplitude and phase angle are extracted from the voltage phasor and the current phasor to obtain mathematical expressions of the phasor amplitude X and the phase angle φ;

[0139] The mathematical expressions of the phase amplitude X and the phase angle φ are respectively:

[0140]

[0141] φ=ac tan(X q / X i ) (3)

[0142] In formula (2) and (3), X i is the i-axis component of the phase X, X q is the q-axis component of the phase X;

[0143] Among them, X i and X q The mathematical expressions are:

[0144]

[0145] In equations (4) and (5), s is the complex variable in Laplace transform; ω is the resonant angular frequency, ω = 2πf; k i ,k q is the extended Kalman filter gain.

[0146] (3) Step S3: Classify and label the preprocessed synchronized phasor measurement information;

[0147] The step S3 specifically includes:

[0148] Step S3.1: dispatching the control center according to the local node to upload the synchronized phasor information;

[0149] Step S3.2: The phasor amplitude and phase angle calculated by the extended Kalman filter method form a sample data set;

[0150] Step S3.3: randomly dividing the sample data set into a training set and a test set; wherein the training set is composed of random fault samples occurring on each branch line in the power distribution system, and the test set is used to test the evaluation performance of the regularized RRBF neural network model;

[0151] Step S3.4: normalizing the preprocessed phasor amplitude and phase angle, and establishing input variables of the regularized RRBF neural network model;

[0152] Step S3.5: establishing the output variable of the regularized RRBF neural network model according to the operating state of the branch in the power distribution system, so as to realize the classification label management of the pre-processed synchronized phasor measurement information;

[0153] Wherein, the mathematical expression of the input vector is:

[0154]

[0155] In formula (6), U n The voltage phasor amplitude information measured for the nth node, arg(U n ) is the phase angle information of the voltage phasor measured by the nth node; I n The amplitude information of the current phasor measured for the nth node, arg(I n ) is the phase angle information of the current phasor measured by the nth node, and n is the number of nodes in the distribution system;

[0156] Wherein, the mathematical expression of the output vector is:

[0157]

[0158] In formula (7), q i ∈[0,1], i=1,2,...n h , n h is the total number of distribution branches in the distribution system; the operating status of each branch is binary-coded, and the number of output vector components is equal to the number of categories, 1 represents a fault in the corresponding branch interval, and 0 represents no fault in the corresponding branch interval.

[0159] (4) Step S4: Under the supervised multi-model residual generation classification learning framework, a regularized RRBF neural network model is proposed; wherein the regularized RRBF neural network model includes an input layer, a hidden layer, a recursive layer, and an output layer that are sequentially connected to each other;

[0160] The step S4 specifically includes:

[0161] Step S4.1: Describe and implement the mapping f from the p-dimensional input space to the one-dimensional output space; wherein the mapping f is expressed as: R P →R 1 , this mapping relationship is described by the hidden layer neural network model, and this mapping relationship is defined by the activation function set φ on the neuron set x j and the corresponding weight set θ on the neuron jiThe linear combination of ;

[0162] Step S4.2: Obtain a mathematical expression of the output vector of the regularized RRBF neural network model through the mapping f; the mathematical expression of the output vector of the regularized RRBF neural network model is:

[0163]

[0164] In formula (8), f i (x) is the output vector; x is the input vector, x=[x1,x2,…x m ], m is the input vector dimension; C j is the center of the hidden layer neuron, C j =[c 1,j ,c 2,j ,…,c m,j ], 1 <j<n;θ ji is the weight from the jth hidden layer neuron to the ith output neuron, 1 <i<n h , nh is the number of output layers; ||·|| is the Euclidean norm from the hidden layer nodes to the output layer nodes;

[0165] Step S4.3: Select the activation function φ of the regularized RRBF neural network model j ;

[0166] The step S4.3 specifically includes: transmitting the input vector x to the radial basis function neuron layer through the input layer, and then calculating the input vector x and the neuron center C of the jth neuron in the hidden layer. j The distance between them, and the distance as the corresponding radial basis excitation function φ j (x) input, then select the activation function φ j is a Gaussian function;

[0167] Among them, the radial basis excitation function φ j The mathematical expression of (x) is as follows:

[0168]

[0169] In formula (9), σ j is the width parameter and satisfies σ j >0;

[0170] Expected output d i (t) is:

[0171]

[0172] In formula (10), ε i (t) is the training set d i(t) and the output vector f i The error between (x(t)), error function ε i (t) is used to measure the difference between the network output and the expected output;

[0173] Step S4.4: Use the zero-order regularized error function to correct the error function ε of equation (10) i (t) is corrected to establish the regularized error term of the regularized RRBF neural network as:

[0174]

[0175] Step S4.5: During the training of the regularized RRBF neural network model, the penalty term λg T g constrains the orthogonal weight vector g in the regularized RRBF neural network model, and the regularized error function is obtained as follows:

[0176]

[0177] In formula (12), the regularization parameter λ is used to balance the fitting degree and complexity of the regularized RRBF neural network model; is the correlation term between the expected output variance and the regression operator. Each vector w j The contribution to the variance increment is

[0178] Step S4.6: The standard form of the regularized error function is obtained by equation (12):

[0179]

[0180] Step S4.7: Obtain the regularized error rate ρ of the regression matrix through equation (13) j for:

[0181]

[0182] Step S4.8: Regularized error rate ρ based on the regression matrix j , establish the accuracy constraint condition of the forward regression process of the regularized RRBF neural network; the mathematical expression of the accuracy constraint condition is:

[0183]

[0184] In formula (15), the regularized error rate ρ of the regression matrix j Used to control the network structure complexity of the regularized RRBF neural network and update the structure of the regularized RRBF neural network online; ξ is the accuracy threshold;

[0185] Step S4.9: In the process of training the regularized RRBF neural network model, the training accuracy threshold ξ is selected at each step. j The corresponding vector with the largest index value is used as a hidden layer node, and the vector of the previous step is orthogonalized until the accuracy threshold ξ satisfies the accuracy constraint condition in the forward regression process of the regularized RRBF neural network.

[0186] The above-mentioned accuracy constraint condition is used to select the center vector of the hidden layer. In the regression process, each vector corresponds to a group of samples in the sample set. When the regression process contains enough feature information, the selection of the hidden layer node center is stopped.

[0187] Furthermore, we need to emphasize that when conducting online fault diagnosis, data must also be obtained in real time. Specifically, in a three-phase distribution system under noise interference, the extended Kalman filter method is used to calculate the phasor amplitude and phase angle of the synchronized phasor information, and the calculated synchronized phasor information is preprocessed to achieve the purpose of extracting the phasor amplitude and phase angle of the synchronized phasor information. Furthermore, we need to explain the next specific implementation steps of online fault diagnosis.

[0188] 2. The specific implementation steps of online fault diagnosis are as follows:

[0189] Step S5: using the regularized RRBF neural network model to generate residuals for the distribution system area, and then performing online intelligent diagnosis on the fault area of ​​the distribution system based on the generated residuals.

[0190] The step S5 specifically includes:

[0191] Step S5.1: Divide the distribution system into a number of regions containing overlapping tie lines, and extract historical synchronized phasor information from each of the divided regions;

[0192] Step S5.2: using the regularized RRBF neural network model to perform fault detection on each divided area, and completing the fault detection task of each area in an independent parallel manner;

[0193] Step S5.3: Set the output matrix of the RRBF fault area detection model to Q = [q1 q2…q j …q n ], the output matrix is ​​abbreviated as q * , the output residual is r = [q1-q * q2-q * …q n -q * ];

[0194] Step S5.4: The decision logic for establishing the RRBF fault area detection model is as follows:

[0195]

[0196] In formula (16), i is the node of the distribution system, is the absolute value of the residual vector corresponding to the i-th node, δ i (i=1, 2…n) is a threshold value associated with the residual output of the RRBF fault area detection model; Specifically refers to the node health status of the system at each moment; T s is the sampling time of the power distribution system; It is the time to detect and isolate the fault that occurred;

[0197] Step S5.5: Determine whether r1 is greater than δ i ; When the judgment result is yes, it is determined that the distribution line i has a fault, so as to realize the online intelligent fault detection method of the distribution system based on the improved RRBF neural network model (i.e., the regularized RRBF neural network model) under the supervised multi-model residual generation classification learning framework, and use the synchronous phasor information of the distribution system to detect the fault area of ​​the distribution system.

[0198] In this specific embodiment, the online intelligent fault detection method for distribution system based on RRBF neural network provided by the present invention, in the dynamic operation process of nonlinear distribution system in noise interference environment, improves the effective complexity of neural network model by using zero-order regularized error function, and can effectively select the hidden layer center vector; establishes RRBF neural network model, can improve the generalization ability of neural network model; classifies and labels the synchronous phasor measurement information in the distribution system, and proposes a fault detection method based on RRBF neural network according to the supervised multi-model residual generation classification learning rule, which can diagnose common short-circuit faults occurring on branches online, improve the fault diagnosis rate, and shorten the response time; improve the operation reliability of distribution system, improve the online intelligent diagnosis performance of distribution system, accurately diagnose faults of distribution lines and electrical components, and achieve second-level response to common faults of distribution network.

[0199] like Figure 3As shown in the figure, the structure diagram of the node topology of the IEEE 13-bus power system is shown; in order to analyze the effectiveness of the online intelligent fault detection method based on the RRBF neural network, the IEEE 13-bus system is used as a test platform to perform online intelligent diagnosis on the common fault states of the IEEE 13-bus system, which contains 1 generator, 2 transformers and 10 AC lines, where the generator represents the equivalent machine of the Canadian power network connected to the system; the fault types include 11 common short-circuit faults including single-phase grounding fault, phase-to-phase short-circuit fault, and three-phase fault. The fault locations are located at the fault points of 0% to 90% of the AC line with a step increase of 10%, the fault time is 1T (1s) to 100T (3s), the simulation time is 10s, and the simulation frequency is 50Hz. In the IEEE 13-bus system, the simulation generates 9000 samples, of which there are 7613 stable samples and 1387 unstable samples. 800 groups of sample points are extracted from the stable samples, and these data sets are randomly organized into 460 training sets and 340 test sets.

[0200] In order to make the system state fully observable, a synchronized phasor measurement device is installed at each node to measure the synchronized phasor information of voltage and current. The sample set of the training process covers the fault conditions under the typical operating conditions of the distribution system, including voltage and current phasor information at different load levels, different fault locations and different fault types.

[0201] Under the framework of supervised multi-model residual generation and classification learning, the online intelligent fault detection method based on RRBF neural network makes full use of measurement information and empirical knowledge, and uses the RRBF neural network model as a classifier to identify the residual vector features corresponding to the fault and health modes of the distribution system. Each fault model is a dynamic neural network, which correspondingly identifies a certain category of system behavior; the RRBF neural network model diagnoses the possibility of faults or abnormal features in system measurements and gives fault diagnosis results to describe the health status of the system, generating fault detection information to conduct quantitative and qualitative analysis of the operating status of the distribution system.

[0202] like Figure 4 As shown, the supervised multi-model residual generation and classification learning framework consists of two stages, multi-model residual generation and isolation decision stage, which learns all categories of system behavior and diagnoses the possibility of faults or abnormal features in system measurements.

[0203] like Figure 5As shown in the figure, a simulation experiment is carried out, and the RRBF neural network is used to perform intelligent diagnosis on the single-phase grounding fault of branch 680-671 in the IEEE 13 power system. The influence of different training sets on the RRBF diagnosis model is compared. It can be found that the RRBF neural network algorithm trained with phase angle information and amplitude information has higher accuracy than the network trained with amplitude information only. The fault diagnosis accuracy range is 96.67% to 98.07% with an average estimation error.

[0204] like Figure 6 As shown, a simulation experiment was conducted to perform online intelligent fault diagnosis on branch 634-633, and the fault diagnosis result of node 634 was displayed. 0.13s after the fault occurred, the diagnosis result of the RRBF neural network increased from 0.106 to 0.967, indicating that node 634 was in a fault state.

[0205] like Figure 7 As shown in the figure, a simulation experiment is carried out to perform online intelligent fault diagnosis on node 634. The output qC of the RRBF fault diagnosis method increases from 1.02s to 0.975. According to the statistical results, the response time of the RRBF fault phase detection model trained by amplitude and phase angle is 0.016s, and the response time of the RRBF detection model trained by amplitude is 0.1s. The RRBF neural network trained by amplitude and phase angle can perform fault diagnosis quickly and accurately. The RRBF neural network fault diagnosis model performs well on the training data, and can also accurately and quickly diagnose the fault area in the actual fault process, and has good generalization ability.

[0206] like Figure 8 As shown in the figure, a simulation experiment was carried out to diagnose common short-circuit faults in the power system based on the RRBF fault phase detection model; in 140 groups of verification samples, the error range of RRBF neural network fault diagnosis using synchronized phasor amplitude and phase angle data sets was 0.172-0.264, which can meet the actual engineering standards.

[0207] like Fig. 9 As shown in the figure, simulation experiments are carried out. Under the noise levels of 20dB and 30dB, the output vector of the RRBF diagnosis model oscillates in the range of [0.836, 0.987], and the response time is 0.22s. The RRBF neural network model trained by amplitude and phase angle enhances the fault diagnosis accuracy and generalization ability in a noisy environment, and can accurately diagnose the fault area online and achieve a fault diagnosis response in seconds.

[0208] Table 1 shows the results of online intelligent diagnosis of common operating conditions using the online intelligent fault detection method for distribution system based on RRBF neural network of the present invention.

[0209] Table 1 Online intelligent diagnosis results of common operating conditions in power systems

[0210]

[0211] As shown in Table 1, online intelligent diagnosis is performed on 12 common operating conditions in the power system, including normal system operation, single-phase grounding fault and three-phase short circuit fault. The fault diagnosis results show that the actual output of the RRBF fault diagnosis method is basically consistent with the expected value, and the fault location can be accurately detected, with a fault diagnosis accuracy of 97.14%.

[0212] In the framework of supervised multi-model residual generation classification learning, this method establishes an RRBF neural network linear regression model, uses the zero-order regularized error function to constrain the effective complexity of the neural network model under the regularization criterion, and uses the regularized error rate to select the hidden layer center vector to improve the generalization ability of the RRBF neural network model; then, classification label management is performed on the synchronized phasor information, and a fault detection method based on the RRBF neural network is proposed according to the supervised multi-model residual generation classification learning rule to identify the fault area in the distribution system; there is no need to analyze the network topology of the power system, and both calculation speed and accuracy are taken into account to realize online intelligent fault detection of the distribution system in a noisy environment.

[0213] According to the simulation experiment results and the comparative experiment results, we can see that the online intelligent fault detection method of the distribution system based on the RRBF neural network of the present invention performs online intelligent fault detection on the distribution system in a noisy environment, performs data preprocessing on the synchronous phasor information in a low signal-to-noise ratio environment and removes bad data; under the regularization criterion, the zero-order regularized error function is used to constrain the model complexity, and the regularized error rate is used to select the hidden layer center vector, improve the generalization ability of the RRBF neural network, and solve the overfitting problem; according to the supervised multi-model residual generation classification learning rule, a fault detection method based on the RRBF neural network is proposed to identify the fault area in the distribution system; the accuracy range of the fault diagnosis method based on the RRBF neural network is 96.57-98.07%, and the response time is 0.21-0.43s, which realizes the second-level response of the online intelligent diagnosis of the power system fault, and solves the overfitting and accuracy problems of the intelligent fault diagnosis in the distribution system in the noisy environment.

[0214] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An online intelligent fault detection method for a power distribution system based on RRBF neural network, characterized in that: include: Step S1: Selecting a second-order model of a power distribution system as a dynamic state estimation model of an AC signal of the power distribution system; Step S2: In a three-phase power distribution system under a noise interference environment, the phasor amplitude and phase angle of the synchronized phasor information are calculated using the extended Kalman filter method, and the calculated synchronized phasor information is preprocessed to filter out the mixed signal of the high-order harmonic component and the white noise to form the initial state phasor information describing the state of the entire power distribution system; Step S3: classify and label the preprocessed synchronized phasor measurement information; Step S4: Under the supervised multi-model residual generation classification learning framework, a regularized RRBF neural network model is proposed; wherein the regularized RRBF neural network model includes an input layer, a hidden layer, a recursive layer, and an output layer which are sequentially connected to each other; Step S5: using the regularized RRBF neural network model to generate residuals for the distribution system area, and then performing online intelligent diagnosis on the fault area of ​​the distribution system based on the generated residuals.

2. The online intelligent fault detection method for power distribution system based on RRBF neural network according to claim 1 is characterized in that: The step S1 specifically includes: Step S1.1: Use the GPS receiving engine to receive the synchronous clock source for unified timing, and then centralize the synchronized phasor information uploaded by PMUs in different geographical locations in a unified time coordinate system to obtain the mathematical expression of the dynamic state estimation model of the AC signal of the distribution system; the mathematical expression of the dynamic state estimation model of the AC signal of the distribution system is: x(t)=Xcos(2πft+φ) (1) In formula (1), x(t) is the AC signal of the distribution system; f is the rated frequency of the distribution system, t is the sampling time, X is the phase amplitude, and φ is the angle between the first sampling point of the AC signal of the distribution system and the peak value of the input signal.

3. The online intelligent fault detection method for power distribution system based on RRBF neural network according to claim 2 is characterized in that: In the step S2, the phasor amplitude and phase angle are extracted from the voltage phasor and the current phasor to obtain mathematical expressions of the phasor amplitude X and the phase angle φ; The mathematical expressions of the phase amplitude X and the phase angle φ are respectively: φ=actan(X q / X i ) (3) In formula (2) and (3), X i is the i-axis component of the phase X, X q is the q-axis component of the phase X; Among them, X i and X q The mathematical expressions are: In equations (4) and (5), s is the complex variable in Laplace transform; ω is the resonant angular frequency, ω = 2πf; k i ,k q is the extended Kalman filter gain.

4. The online intelligent fault detection method for power distribution system based on RRBF neural network according to claim 3 is characterized in that: The step S3 specifically includes: Step S3.1: dispatching the control center according to the local node to upload the synchronized phasor information; Step S3.2: The phasor amplitude and phase angle calculated by the extended Kalman filter method form a sample data set; Step S3.3: randomly dividing the sample data set into a training set and a test set; wherein the training set is composed of random fault samples occurring on each branch line in the power distribution system, and the test set is used to test the evaluation performance of the regularized RRBF neural network model; Step S3.4: normalizing the preprocessed phasor amplitude and phase angle, and establishing input variables of the regularized RRBF neural network model; Step S3.5: establishing the output variable of the regularized RRBF neural network model according to the operating state of the branch in the power distribution system, so as to achieve classification label management of the pre-processed synchronized phasor measurement information; Wherein, the mathematical expression of the input vector is: In formula (6), U n The voltage phasor amplitude information measured for the nth node, arg(U n ) is the phase angle information of the voltage phasor measured by the nth node; I n The amplitude information of the current phasor measured for the nth node, arg(I n ) is the phase angle information of the current phasor measured by the nth node, and n is the number of nodes in the distribution system; Wherein, the mathematical expression of the output vector is: In formula (7), q i ∈[0,1], i=1,2,...n h , n h is the total number of distribution branches in the distribution system; the operating status of each branch is binary-coded, and the number of output vector components is equal to the number of categories, 1 represents a fault in the corresponding branch interval, and 0 represents no fault in the corresponding branch interval.

5. The online intelligent fault detection method for power distribution system based on RRBF neural network according to claim 4 is characterized in that: The step S4 specifically includes: Step S4.1: Describe and implement the mapping f from the p-dimensional input space to the one-dimensional output space; wherein the mapping f is expressed as: R P →R 1 , this mapping relationship is described by the hidden layer neural network model, and this mapping relationship is defined by the activation function set φ on the neuron set x j and the corresponding weight set θ on the neuron ji The linear combination of ; Step S4.2: Obtain a mathematical expression of the output vector of the regularized RRBF neural network model through the mapping f; the mathematical expression of the output vector of the regularized RRBF neural network model is: In formula (8), f i (x) is the output vector; x is the input vector, x=[x1,x2,…x m ], m is the input vector dimension; C j is the center of the hidden layer neuron, C j =[c 1,j ,c 2,j ,…,c m,j ], 1 <j<n;θ ji is the weight from the jth hidden layer neuron to the ith output neuron, 1 <i<n h , n h is the number of output layers; ||·|| is the Euclidean norm from the hidden layer nodes to the output layer nodes; Step S4.3: Select the activation function φ of the regularized RRBF neural network model j ; The step S4.3 specifically includes: transmitting the input vector x to the radial basis function neuron layer through the input layer, and then calculating the input vector x and the neuron center C of the jth neuron in the hidden layer. j The distance between them, and the distance as the corresponding radial basis excitation function φ j (x) input, then select the activation function φ j is a Gaussian function; Among them, the radial basis excitation function φ j The mathematical expression of (x) is as follows: In formula (9), σ j is the width parameter and satisfies σ j >0; Expected output d i (t) is: In formula (10), ε i (t) is the training set d i (t) and the output vector f i The error between (x(t)), error function ε i (t) is used to measure the difference between the network output and the expected output; Step S4.4: Use the zero-order regularized error function to correct the error function ε of equation (10) i (t) is corrected to establish the regularized error term of the regularized RRBF neural network as: Step S4.5: During the training of the regularized RRBF neural network model, the penalty term λg T g constrains the orthogonal weight vector g in the regularized RRBF neural network model, and the regularized error function is obtained as follows: In formula (12), the regularization parameter λ is used to balance the fitting degree and complexity of the regularized RRBF neural network model; is the correlation term between the expected output variance and the regression operator. Each vector w j The contribution to the variance increment is Step S4.6: The standard form of the regularized error function is obtained by equation (12): Step S4.7: Obtain the regularized error rate ρ of the regression matrix through equation (13) j for: Step S4.8: Regularized error rate ρ based on the regression matrix j , establish the accuracy constraint condition of the forward regression process of the regularized RRBF neural network; the mathematical expression of the accuracy constraint condition is: In formula (15), the regularized error rate ρ of the regression matrix j Used to control the network structure complexity of the regularized RRBF neural network and update the structure of the regularized RRBF neural network online; ξ is the accuracy threshold; Step S4.9: In the process of training the regularized RRBF neural network model, the training accuracy threshold ξ is selected at each step. j The corresponding vector with the largest index value is used as a hidden layer node, and the vector of the previous step is orthogonalized until the accuracy threshold ξ satisfies the accuracy constraint condition in the forward regression process of the regularized RRBF neural network.

6. The online intelligent fault detection method for power distribution system based on RRBF neural network according to claim 5 is characterized in that: The step S5 specifically includes: Step S5.1: Divide the distribution system into a number of regions containing overlapping tie lines, and extract historical synchronized phasor information from each of the divided regions; Step S5.2: using the regularized RRBF neural network model to perform fault detection on each divided area, and completing the fault detection task of each area in an independent parallel manner; Step S5.3: Set the output matrix of the RRBF fault area detection model to Q = [q1 q2…q j …q n ], the output matrix is ​​abbreviated as q * , the output residual is r = [q1-q * q2-q * …q n -q * ]; Step S5.4: The decision logic for establishing the RRBF fault area detection model is as follows: In formula (16), i is the node of the distribution system, is the absolute value of the residual vector corresponding to the i-th node, δ i (i=1, 2…n) is a threshold value associated with the residual output of the RRBF fault area detection model; Specifically refers to the node health status of the system at each moment; T s is the sampling time of the power distribution system; It is the time to detect and isolate the fault that occurred; Step S5.5: Determine whether r1 is greater than δ i ; When the judgment result is yes, it is determined that the distribution line i is faulty, so as to detect the fault area of ​​the distribution system by using the synchronous phasor information of the distribution system under the supervised multi-model residual generation classification learning framework.

Images