A method for identifying ground fault types based on self-organizing competitive neural network

Abstract

A method for identifying the type of grounding fault based on a self-organizing competitive neural network belongs to the field of relay protection of power systems. It is characterized by comprising the following steps: step a, collecting the zero-sequence current after the fault and obtaining the power frequency component; step b, obtaining the amplitude change trend function of the power frequency component of the zero-sequence current; step c, extracting characteristic parameters; step d, if there is no unstable grounding grounding fault, execute step e; step e, construct a neural network based on a self-organizing competitive type; step f, using the neural network based on the self-organizing competitive type to determine whether a tree-line discharge fault or other types of stable grounding grounding fault occurs in the distribution network, and output it. In this method for identifying the type of grounding fault based on a self-organizing competitive neural network, a competitive neural network is used to identify the type of grounding fault in the distribution network, especially accurately identifying the tree-line discharge fault, filling the gap in this field.

Description

Technical Field

[0001] The invention discloses a ground fault type identification method based on a self-organizing competitive neural network, belonging to the field of relay protection of power systems. Background Art

[0002] High-resistance grounding faults in medium-voltage distribution networks account for 5% to 10% of the total number of single-phase grounding faults, some of which are tree-line discharge faults where overhead lines are grounded through trees. The electrical characteristic of tree-line discharge faults is weak and often cannot reach the threshold value for the protection device to start. The long-term existence of such faults will cause the risk of forest fires, seriously threatening the safety of life and property of the people.

[0003] At present, the methods for identifying the type of grounding fault in medium-voltage distribution networks mainly use algorithms such as wavelet transform to extract the time-frequency domain characteristics of the zero-sequence information of the grounding fault as feature data, or directly use the waveform of the zero-sequence information of the grounding fault as feature data, and input the feature data into a support vector machine or a neural network to realize the type identification of the grounding fault. The data preprocessing methods of the existing methods for identifying grounding faults in medium-voltage distribution networks are too complicated, and the use of machine learning intelligent algorithms has fuzzy boundaries and uncertainties. In addition, there is no reliable and accurate method for detecting tree-line discharge faults, and it is impossible to identify whether a tree-line discharge fault has occurred. There is a gap in this research field. Summary of the invention

[0004] The technical problem to be solved by the present invention is: to overcome the shortcomings of the prior art and provide a grounding fault type identification method based on a self-organizing competitive neural network that utilizes a competitive neural network to identify the type of grounding fault in a distribution network, especially accurately identifies tree-line discharge faults.

[0005] The technical solution adopted by the present invention to solve the technical problem is: the ground fault type identification method based on the self-organizing competitive neural network is characterized by comprising the following steps:

[0006] Step a, after a ground fault occurs in the distribution line, collecting a zero-sequence current signal after the fault, and calculating a power frequency component of the zero-sequence current signal after the fault;

[0007] Step b, obtaining a trend function of the amplitude change of the power frequency component of the zero-sequence current signal according to the power frequency component of the zero-sequence current signal after the fault;

[0008] Step c, extracting characteristic parameters of the zero-sequence current signal amplitude variation trend function by curve fitting;

[0009] Step d, determining whether the type of ground fault occurring in the distribution network is a ground fault with unstable grounding. If the type of ground fault occurring in the distribution network is a ground fault with unstable grounding, the determination ends; if the type of ground fault occurring in the distribution network is not a ground fault with unstable grounding, executing step e;

[0010] Step e, constructing a neural network based on self-organizing competition;

[0011] Step f: a self-organizing competitive neural network is used to determine whether a tree-line discharge fault or other type of stable grounding fault occurs in the distribution network, and outputs the result.

[0012] Preferably, when executing step e, the following steps are included:

[0013] Step e-1, obtaining a symbol matrix and a Euclidean distance matrix of characteristic parameters of a zero-sequence current amplitude variation trend function;

[0014] Step e-2, calculating the state of the competition layer element;

[0015] Calculate the state S of neuron j in the competition layer j :

[0016] Among them, inputsum i is the i-th element of the input layer data;

[0017] Step e-3, solve the winning neuron. The winning neuron represents the classification mode of the input sample. The neuron with the largest weighted value in the competition layer wins, and the output is:

[0018]

[0019] Step e-4, calculate the weight correction of the winning neuron. For all input layer neurons i, we have:

[0020]

[0021] Among them, α is the learning coefficient, which is generally 0<α<<1; μ is the number of neurons with output 1, that is,

[0022] Step e-5, calculating the Euclidean distance between the input layer data corresponding to the neuron with an output of 1 and all the input layer data, and searching for similar data;

[0023] Step e-6, determine whether the convergence condition is met. If the convergence condition is met, execute step e-7. If the convergence condition is not met, return to step e-2.

[0024] Step e-7, return to step f.

[0025] Preferably, in the step e-5, the calculation formula of the similar data is:

[0026]

[0027] Where inputsum output＝1 i is the i-th element of the input layer data corresponding to the neuron with output 1, and inputsumi is the i-th element of all input layer data corresponding to it. output＝1 When the inputsum is less than the preset threshold, the two sets of data are considered similar.

[0028] Preferably, when executing the step d, the mean square error of the two signals is obtained according to the difference between the trend function of the amplitude change of the power frequency component of the zero-sequence current signal obtained in step b and the trend function fitted in step c. If the mean square error of the two signals is less than a set threshold value, it indicates that the type of grounding fault occurring in the distribution network is an unstable grounding grounding fault; when the mean square error of the two signals is greater than or equal to the set threshold value, it indicates that the type of grounding fault occurring in the distribution network is a stable grounding grounding fault.

[0029] Preferably, in the step b, the trend function Amp_i0(k) of the amplitude change of the power frequency component of the zero-sequence current signal is:

[0030]

[0031] Where: A1 represents the amplitude of the exponential component, A2 represents the attenuation / gain coefficient of the exponential component, and A3 represents the amplitude of the constant DC component.

[0032] Preferably, the mean square error (MSE) of the two signals (Amp_i0(k), y (3) )for:

[0033]

[0034] Where n is the data length of the zero-sequence current signal, W i is the weight of the RMS difference, Amp_i0(k) is the trend function of the amplitude change of the power frequency component of the zero-sequence current signal, y (3) is the trend function obtained by curve fitting.

[0035] Preferably, when executing step c, the following steps are included:

[0036] Step c-1, fit Amp_i0(k) with a cubic polynomial and extract its characteristic parameters, and the trend function y obtained by curve fitting corresponding to the amplitude change trend function Amp_i0(k) (3) The fitting formula is:

[0037] y (3) =B3x 3 +B2x 2 +B1x+B0

[0038] Among them, B3~B0 are fitting parameters;

[0039] Step c-2, Taylor expansion is performed on the trend function of the amplitude change of the power frequency component of the zero-sequence current signal, and the high-order terms of the Taylor expansion are ignored to obtain:

[0040]

[0041] Where: A1 represents the amplitude of the exponential component, A2 represents the attenuation / gain coefficient of the exponential component, and A3 represents the amplitude of the constant DC component;

[0042] Using cubic fitting Amp_i0(k), its characteristic parameters are extracted, where the corresponding relationship between the cubic fitting parameters and the Taylor expansion parameters of Amp_i0(k) is:

[0043]

[0044] Wherein: A1 represents the amplitude of the exponential component, A2 represents the attenuation / gain coefficient of the exponential component, A3 represents the amplitude of the constant DC component; B3~B0 are fitting parameters.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] In this grounding fault type identification method based on self-organizing competitive neural network, the competitive neural network is used to identify the type of grounding fault in the distribution network, especially the accurate identification of tree-line discharge faults, which fills the gap in this field and timely and accurately identifies the occurrence of tree-line discharge faults, intermittent grounding faults and other types of grounding faults in medium-voltage distribution networks.

[0047] In this grounding fault type identification method based on self-organizing competitive neural network, the amplitude change trend of the zero-sequence current of intermittent grounding fault is approximately a square wave signal, the frequency component is complex, and it is difficult to completely fit it using curve fitting, while other fault signals can be completely fitted. The difference between the amplitude change trend of intermittent grounding fault and the zero-sequence current of different grounding faults and the root mean square difference of the amplitude change trend after fitting is used to judge whether an intermittent grounding fault occurs. This method has the advantages of low calculation amount and easy adjustment.

[0048] The characteristic parameters of the zero-sequence current signal amplitude change trend function are extracted by curve fitting methods such as cubic polynomials, and a neural network based on a self-organizing competition structure is constructed. The characteristic parameters obtained by fitting are used as the data set of the neural network input layer and further processed. The classification result is determined by the competition between the neurons in the competition layer and the weight adjustment after the competition to judge whether a tree-line discharge fault occurs. Finally, the identification of intermittent grounding faults, tree-line discharge faults and other types of grounding faults in the medium-voltage distribution network is realized with high classification and recognition accuracy.

[0049] In this ground fault type identification method based on self-organizing competitive neural network, the distribution automation device is used to collect the zero-sequence current signal within a fixed time window after the fault occurs, and FFT is used to extract the power frequency component and construct the zero-sequence current amplitude change trend function, which reduces the influence of the harmonic component of the tree-line discharge fault and reduces the calculation amount, thereby improving the calculation speed of the distribution automation device.

[0050] In this grounding fault type identification method based on a self-organizing competitive neural network, the data input to the input layer of the self-organizing competitive neural network is further classified and processed, and the symbol matrix and the Euclidean distance matrix are extracted, which solves the problems of high discreteness of the original data, scattered inherent features, slow neural network classification and large computational complexity; and by finding the corresponding relationship between the input layer data corresponding to the neuron with an output of 1 and all the input layer data, the learning parameters are adaptively adjusted to improve the classification speed and accuracy of the neural network.

[0051] The zero-sequence current signal can be acquired through traditional power frequency sensor collection or three-phase synthesis, without adding additional primary equipment or cooperating with other primary equipment, and has high practical application value.

[0052] The characteristic parameters of the zero-sequence current signal amplitude change trend function are extracted by curve fitting methods such as polynomial, Fourier series fitting and Prony algorithm, and the symbol matrix of the characteristic parameters is constructed. The criterion is constructed based on whether there are negative elements or the product or ratio of elements in the symbol matrix, so as to detect whether tree-line discharge faults and other grounding faults occur. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The flowchart of the ground fault type identification method based on the self-organizing competitive neural network is shown in FIG.

[0054] Figure 2 A flow chart of neural network construction for ground fault type identification method based on self-organizing competitive neural network. DETAILED DESCRIPTION

[0055] Figures 1-2 The best embodiment of the present invention is shown below in conjunction with the attached Figures 1-2 The present invention is further described.

[0056] Step 1001, start;

[0057] The distribution automation device arranged in the distribution network determines whether a ground fault occurs in the distribution network. When a ground fault occurs, step 1002 is executed.

[0058] Step 1002, collecting a zero-sequence current signal within a fixed time window;

[0059] The distribution automation device in the distribution network collects the zero-sequence current signal within the fixed time window after the fault occurs according to the zero-sequence mutual inductor set in the distribution network. The zero-sequence current signal after the grounding fault of the distribution network is defined as i0(n), where n is the data length of the zero-sequence current signal, which depends on the fixed time window length and the distribution automation sampling device. Generally, the fixed time window length is greater than or equal to 150s, and the sampling frequency is greater than or equal to 6400Hz.

[0060] Step 1003, extracting the power frequency component of the zero-sequence current signal;

[0061] The distribution automation device in the distribution network uses FFT to extract the power frequency component i0(k) of the zero-sequence current signal i0(n) after the ground fault of the distribution network occurs, that is:

[0062]

[0063] Wherein, n represents the data length of the zero-sequence current signal, k represents the data length of the power frequency component, and j represents the imaginary part.

[0064] Step 1004, obtaining a function of a change trend of an amplitude of a power frequency component of a zero-sequence current signal;

[0065] The frequency and amplitude characteristics of the corresponding frequency contained in the power frequency component i0(k) of the zero-sequence current signal after the ground fault occurs, and then the trend function Amp_i0(k) of the amplitude change of the power frequency component of the zero-sequence current signal is obtained:

[0066]

[0067] Where: A1 represents the amplitude of the exponential component, A2 represents the attenuation / gain coefficient of the exponential component, and A3 represents the amplitude of the constant DC component.

[0068] Step 1005, extracting characteristic parameters of the zero-sequence current signal amplitude variation trend function;

[0069] The characteristic parameters of the zero-sequence current signal amplitude change trend function are extracted by using curve fitting methods such as polynomial fitting, Fourier series fitting or Prony algorithm;

[0070] The specific steps include:

[0071] Step 1005-1, obtaining a fitting formula Amp_i0(k) of the amplitude change of the power frequency component of the zero-sequence current signal;

[0072] Taking the method of fitting Amp_i0(k) through a cubic polynomial and extracting its characteristic parameters as an example, the fitting formula of the amplitude change trend function Amp_i0(k) is:

[0073] y (3) =B3x 3 +B2x 2 +B1x+B0

[0074] Among them, B3~B0 are fitting parameters.

[0075] Step 1005-2, Taylor expansion is performed on the trend function of the amplitude change of the power frequency component of the zero-sequence current signal to obtain:

[0076]

[0077] In the above Taylor expansion, for high-resistance grounding faults, especially tree-line discharge faults, the attenuation / gain coefficient A2 has a large value, so the high-order terms in the above Taylor expansion can be ignored. The above Taylor expansion can be approximated as follows:

[0078]

[0079] Using cubic fitting Amp_i0(k), its characteristic parameters are extracted, where the corresponding relationship between the cubic fitting parameters and the Taylor expansion parameters of Amp_i0(k) is:

[0080]

[0081] Step 1006, determining whether the fault in the distribution network is an intermittent ground fault, if it is an intermittent ground fault, executing step 1009, if it is not an intermittent ground fault, executing step 1007;

[0082] Calculate the trend function of zero-sequence current amplitude change Amp_i0(k) and the trend function y obtained by curve fitting (3) The difference between the two signals is obtained by taking the mean square error MSE (Amp_i0(k), y (3) ):

[0083]

[0084] Where n is the data length of the zero-sequence current signal, W i is the weight of the RMS difference, which is usually set to 1.

[0085] When MSE(Amp_i0(k),y (3) ) is less than the set threshold value θ, it indicates that the type of ground fault in the distribution network is an unstable ground fault, and step 1009 is executed; when MSE(Amp_i0(k), y (3) ) is greater than or equal to the set threshold value θ, it indicates that the type of ground fault in the distribution network is a stable ground fault, and step 1007 is executed. θ is a preset threshold value, which can be set by oneself, such as θ=0.6.

[0086] Step 1007, constructing a neural network based on self-organizing competition;

[0087] Construct a neural network based on self-organizing competition, including an input layer and a competition layer, where the number of neurons in the input layer N is consistent with the number of characteristic parameters, and the number of competition layers M is generally greater than twice the number of neurons in the input layer N. The weight of the connection between the input layer and the competition layer is ω ij (i=1:N;j=1:M) which satisfies the constraint condition

[0088] like Figure 2 As shown, this step includes the following steps:

[0089] Step 1007-1, obtaining a symbol matrix and a Euclidean distance matrix of characteristic parameters of a zero-sequence current amplitude variation trend function;

[0090] The input layer data data(B)=[B0 B1 B2 B3] is further processed to obtain its symbol matrix sgn(B), where:

[0091] sgn(B)=[sgn(B0) sgn(B1) sgn(B2) sgn(B3)]

[0092] Calculate the Euclidean distance ED(B) between B1, B2, B3 and B0 = [ED(B0B1) ED(B0B2) ED(B0B3)], where:

[0093]

[0094] Therefore, the data input to the input layer of the neural network is:

[0095] inputsum=[data(B) sgn(B) ED(B)].

[0096] Step 1007-2, calculating the state of the competition layer element;

[0097] Calculate the state S of neuron j in the competition layer j :

[0098]

[0099] Among them, inputsum i is the i-th element of the input layer data.

[0100] Step 1007-3, solving the winning neuron;

[0101] Solve for the winning neuron. The winning neuron represents the classification mode of the input sample. The neuron with the largest weight in the competition layer wins, and the output is:

[0102]

[0103] Step 1007-4, calculating the weight correction of the winning neuron;

[0104] Calculate the weight correction of the winning neuron. For all input layer neurons i, we have:

[0105]

[0106] Among them, α is the learning coefficient, which is generally 0<α<<1; μ is the number of neurons with output 1, that is,

[0107] Step 1007-5, adaptive learning parameters;

[0108] Calculate the Euclidean distance between the input layer data corresponding to the neuron with an output of 1 and all input layer data, and seek similar data. The calculation formula is:

[0109]

[0110] Where inputsum output＝1 i is the i-th element of the input layer data corresponding to the neuron with output 1, and inputsumi is the i-th element of all input layer data corresponding to it. output＝1 When the inputsum is less than the preset threshold, the two sets of data are considered similar, and the number of similar data is set to γ. With γ as the reference, the learning coefficient α is proportional to γ, and the adjustment is made adaptively.

[0111] Step 1007-6, whether the convergence condition is met;

[0112] Determine whether the convergence condition is met. If the convergence condition is met, execute step 1007-7. If the convergence condition is not met, return to step 1007-2.

[0113] Step 1007-7, return to the main process; return to the main process and execute step 1008.

[0114] Step 1008, outputting the result of the ground fault type;

[0115] In the loop of steps 1007 - 3 to 1007 - 5 in step 1007 , the self-organizing competitive neural network determines whether a tree-line discharge fault or other type of stable grounding fault occurs in the distribution network, and outputs the result.

[0116] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for identifying ground fault types based on a self-organizing competitive neural network, characterized in that: The steps include: Step a, after a ground fault occurs in the distribution line, collecting a zero-sequence current signal after the fault, and calculating a power frequency component of the zero-sequence current signal after the fault; Step b, obtaining a trend function of the amplitude change of the power frequency component of the zero-sequence current signal according to the power frequency component of the zero-sequence current signal after the fault; Step c, extracting characteristic parameters of the zero-sequence current signal amplitude variation trend function by curve fitting; Step d, determining whether the type of ground fault occurring in the distribution network is an unstable ground fault, if the type of ground fault occurring in the distribution network is an unstable ground fault, then the determination ends; If the type of ground fault occurring in the distribution network is not a non-stable ground fault, execute step e; Step e, constructing a neural network based on self-organizing competition; Step f, using a self-organizing competitive neural network to determine whether a tree-line discharge fault or other types of stable grounding fault occurs in the distribution network, and outputting the result; When executing the step d, the mean square error of the two signals is obtained according to the difference between the trend function of the amplitude change of the power frequency component of the zero-sequence current signal obtained in step b and the trend function fitted in step c. If the mean square error of the two signals is less than the set threshold value, it indicates that the type of grounding fault occurring in the distribution network is a non-stable grounding grounding fault; when the mean square error of the two signals is greater than or equal to the set threshold value, it indicates that the type of grounding fault occurring in the distribution network is a stable grounding grounding fault.

2. The ground fault type identification method based on the self-organizing competitive neural network according to claim 1 is characterized in that: When executing step e, the following steps are included: Step e-1, obtaining a symbol matrix and a Euclidean distance matrix of characteristic parameters of a zero-sequence current amplitude variation trend function; Step e-2, calculating the state of the competition layer element; Calculate the state S of neuron j in the competition layer j : Among them, inputsum i is the i-th element of the input layer data; Step e-3, solve the winning neuron. The winning neuron represents the classification mode of the input sample. The neuron with the largest weighted value in the competition layer wins, and the output is: Step e-4, calculate the weight correction of the winning neuron. For all input layer neurons i, we have: Among them, α is the learning coefficient, which is generally 0<α<<1; μ is the number of neurons with output 1, that is, ω ij Represents the weight of the connection between the input layer and the competition layer; Step e-5, calculating the Euclidean distance between the input layer data corresponding to the neuron with an output of 1 and all the input layer data, and searching for similar data; Step e-6, determining whether the convergence condition is met, if the convergence condition is met, executing step e-7, if the convergence condition is not met, returning to step e-2; Step e-7, return to step f.

3. The ground fault type identification method based on the self-organizing competitive neural network according to claim 2 is characterized in that: In the step e-5, the calculation formula of the similar data is: Where inputsum output＝1 i is the i-th element of the input layer data corresponding to the neuron with output 1, and inputsumi is the i-th element of all input layer data corresponding to it. output＝1 When the inputsum is less than the preset threshold, the two sets of data are considered similar.

4. The ground fault type identification method based on the self-organizing competitive neural network according to claim 1 is characterized in that: In the step b, the trend function Amp_i0(k) of the amplitude change of the power frequency component of the zero-sequence current signal is: Wherein: A1 represents the amplitude of the exponential component, A2 represents the attenuation / gain coefficient of the exponential component, A3 represents the amplitude of the constant DC component, and k represents the data length of the power frequency component.

5. The ground fault type identification method based on self-organizing competitive neural network according to claim 1 is characterized in that: The mean square error value MSE(Amp_i0(k),y (3) )for: Where n is the data length of the zero-sequence current signal, W i is the weight of the RMS difference, Amp_i0(k) is the trend function of the amplitude change of the power frequency component of the zero-sequence current signal, y (3) is the trend function obtained by curve fitting.

6. The ground fault type identification method based on a self-organizing competitive neural network according to claim 1 is characterized in that: When executing step c, the following steps are included: Step c-1, fit Amp_i0(k) with a cubic polynomial and extract its characteristic parameters, and the trend function y obtained by curve fitting corresponding to the amplitude change trend function Amp_i0(k) (3) The fitting formula is: y (3) =B3x 3 +B2x 2 +B1x+B0 Among them, B3~B0 are fitting parameters; Step c-2, Taylor expansion is performed on the trend function of the amplitude change of the power frequency component of the zero-sequence current signal, and the high-order terms of the Taylor expansion are ignored to obtain: Where: A1 represents the amplitude of the exponential component, A2 represents the attenuation / gain coefficient of the exponential component, A3 represents the amplitude of the constant DC component; k represents the data length of the power frequency component; Using cubic fitting Amp_i0(k), its characteristic parameters are extracted, where the corresponding relationship between the cubic fitting parameters and the Taylor expansion parameters of Amp_i0(k) is: Wherein: A1 represents the amplitude of the exponential component, A2 represents the attenuation / gain coefficient of the exponential component, A3 represents the amplitude of the constant DC component; B3~B0 are fitting parameters.

Images