A new energy distribution network overcurrent protection system based on characteristic energy spectrum fault identification

Abstract

The present invention relates to the technical field of overcurrent protection of new energy distribution network, and discloses a new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification, including a multi-frequency current waveform acquisition module, a waveform sorting module, a waveform processing module, a current characteristic energy spectrum conversion module, a current characteristic energy spectrum extraction module, a current characteristic energy spectrum multi-resolution module, a fault type identification module, a current characteristic energy spectrum horizontal comparison module, a current characteristic energy spectrum vertical comparison module, a current characteristic energy spectrum database module, a current characteristic energy spectrum detection module, a new energy distribution network overcurrent protection judgment logic module and a system execution module. The system uses characteristic energy data of current waveform as a current protection criterion for new energy distribution network, makes up for the access of distributed new energy power sources in the new energy distribution network, and the distribution network is changed from a single power network to a multi-power network, which can reliably reflect various internal faults of various distributed new energy distribution networks.

Description

Technical Field

[0001] The present invention relates to the technical field of overcurrent protection for a new energy distribution network, and in particular to an overcurrent protection system for a new energy distribution network for characteristic energy spectrum fault identification. Background Art

[0002] At present, under the background of new power system, the proportion of distributed renewable energy as the main power source is constantly rising. With the impact of large-scale distributed renewable energy grid connection on the power system, the steady-state power frequency component is greatly weakened and loses stability, and various whole and non-whole harmonics are greatly enhanced. At the same time, the uncertainty of other factors such as fault type, fault location, duration, etc. makes the fault characteristics of renewable energy power sources lack regularity. Therefore, the change of fault characteristics makes the traditional distribution network current protection lose its original adaptability, the performance of existing protection is significantly reduced, and false operation or refusal to operate also occurs from time to time;

[0003] With the large-scale access of distributed renewable energy, the distribution network has changed from a traditional single-power network to a dual-power or even multi-power network. The power flow distribution law and the nature of the network load have changed, making it difficult for traditional protection schemes to adapt and new problems to emerge. A large number of distributed power sources not only have intermittent characteristics during power generation, but also make it difficult to adjust traditional overcurrent protection schemes. In addition, there are differences between inverter-type distributed power sources and traditional synchronous machine-type power sources, and the control strategies adopted often have a greater impact on their fault output characteristics. Therefore, the adaptability analysis of overcurrent protection of distribution networks containing distributed power sources not only realizes the fault safety risk analysis of overcurrent protection of distribution networks containing distributed power sources, but also has important significance for the design of overcurrent protection schemes for distribution networks after the access of renewable energy and for improving the safe and reliable operation of distribution networks. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the deficiencies in the prior art, the present invention provides a new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification, which has the advantages of improving protection speed and solves the above-mentioned technical problems.

[0006] (II) Technical solution

[0007] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification, comprising a multi-frequency current waveform acquisition module, a waveform sorting module, a waveform processing module, a current characteristic energy spectrum conversion module, a current characteristic energy spectrum extraction module, a current characteristic energy spectrum multi-resolution module, a fault type identification module, a current characteristic energy spectrum horizontal comparison module, a current characteristic energy spectrum vertical comparison module, a current characteristic energy spectrum database module, a current characteristic energy spectrum detection module, a new energy distribution network overcurrent protection judgment logic module and a system execution module;

[0008] The multi-frequency current waveform acquisition module is used to acquire and store the current waveform of the power distribution network. The waveform sorting module sorts and classifies the multi-frequency current waveform through the extreme learning machine algorithm. The multi-frequency current waveform acquisition module includes a power frequency current waveform acquisition unit and a non-power frequency current waveform acquisition unit. The power frequency current waveform acquisition unit and the non-power frequency current waveform acquisition unit respectively acquire the power frequency current waveform and the non-power frequency current waveform data.

[0009] The waveform processing module is used to pre-process the waveform data, eliminate the waveform distortion caused by electromagnetic interference and system disturbance, and decompose the multi-frequency data;

[0010] The current characteristic energy spectrum conversion module converts the decomposed waveform into a characteristic energy spectrum through a wavelet packet, the current characteristic energy spectrum extraction module uses the characteristic energy spectrum extraction amount as the input of the classifier, and the current characteristic energy spectrum multi-resolution module is used to identify the decomposition amount of the minimum frequency;

[0011] The fault type identification module identifies the fault of the new energy distribution network through the characteristic energy spectrum, and at the same time, different faults identified by the fault type identification module are further identified by the current characteristic energy spectrum horizontal comparison module and the current characteristic energy spectrum vertical comparison module respectively;

[0012] The current characteristic energy spectrum lateral comparison module is used to process the characteristic energy spectra of the first 1 / 4 cycle and the last 1 / 4 cycle of the current waveform of a specific frequency and then output a judgment instruction;

[0013] The process of further identifying faults by the current characteristic energy spectrum longitudinal comparison module is as follows: training the characteristic energy spectrum through the current characteristic energy spectrum database module, detecting the effective characteristic energy spectrum through the current characteristic energy spectrum detection module, and the current characteristic energy spectrum longitudinal comparison module longitudinally comparing with the historical characteristic energy spectrum database and outputting a judgment instruction;

[0014] The judgment instructions output by the current characteristic energy spectrum horizontal comparison module and the current characteristic energy spectrum vertical comparison module are transmitted to the new energy distribution network overcurrent protection judgment logic module, and the data of the new energy distribution network overcurrent protection judgment logic module are satisfied by the judgment instructions or current quantity judgment, and finally output to the system execution module.

[0015] As a preferred technical solution of the present invention, the power frequency current waveform acquisition unit introduces DFT orthogonal coefficients after the waveform input passes through an analog low-pass filter, and finally outputs the power frequency current waveform through a synchronous sampling unit and a power frequency vector calculation digital filter. The power frequency vector calculation digital filter uses a Hamming window low-pass filter for filtering calculation, and its specific expression is as follows:

[0016]

[0017] Where X(i) represents the independent variable of the Hanning window function, i represents the position of the window, h=0,1,2…N-1 represents the width of the window function, N represents the window length of the window function, f0 represents the sampling frequency, W k represents the coefficient of the Hamming window function, e represents the natural logarithm, the superscript j represents the imaginary unit, and Gain represents the rectangular window function;

[0018] The non-power frequency current waveform acquisition unit calculates the rotation factor of the input waveform data through the synchronous sampling unit, and then completes the non-power frequency current waveform synchronous vector acquisition in the range of 2 to 2.5 kHz by combining the windowed FFT algorithm with spectrum correction, wherein the non-power frequency current waveform synchronous acquisition data applies the Hanning window spectrum function, and its specific expression is as follows:

[0019]

[0020] Among them, W H (ω) represents the Hanning window spectrum characteristic function, ω∈[-π,π] represents the angular frequency, N represents the window length of the window function, and w R (*) represents the continuous spectrum of the Hanning window, and w R The expression of (ω) is as follows:

[0021]

[0022] Where sin(*) represents the sine function, e represents the natural logarithm, and the superscript represents the imaginary unit, ω represents the angular frequency, and N represents the window length of the window function;

[0023] When the actual frequency of the non-power frequency current signal received by the non-power frequency current waveform acquisition unit is not the frequency resolution of FFT When the actual frequency position of the signal is an integer multiple of f, the amplitude ratio of the two adjacent maximum spectral lines in the main lobe is used to estimate the actual frequency position of the signal. r The specific expression is as follows:

[0024]

[0025] in, Indicates the frequency point, Indicates The resolution of frequency points, Indicates The signal time at each frequency point, Represents an integer value, Represents the actual frequency component f r The closest FFT frequency point, Represents the actual frequency component f r relatively, The offset of δ represents the frequency offset, and its specific expression is as follows:

[0026]

[0027] in, Indicates The amplitude of the spectral line, Indicates The amplitude of the spectral line, Indicates The amplitude of the spectral line, |*| represents the absolute value of the spectral line amplitude.

[0028] As a preferred technical solution of the present invention, the extreme learning machine algorithm in the waveform sorting module includes an input layer, a hidden layer and an output layer. The mathematical model of the extreme learning machine algorithm is as follows:

[0029]

[0030] Simplifying the expression, we get:

[0031] Hβ=Y

[0032] Among them, H represents the hidden layer output matrix, β represents the hidden layer output weight matrix, Y represents the output layer expected output matrix, β σ represents the output weight from the σth hidden layer node to the output neuron, ω σ represents the input weight from the input neuron to the σth hidden layer node, b σ represents the threshold of the σth hidden layer node, y j Represents the output value of the jth training sample. The expression of the hidden layer output matrix H is as follows:

[0033]

[0034] Wherein, f(*) represents the activation function of the hidden layer, and the steps of the waveform sorting module are as follows:

[0035] A1. After information fusion diagnosis of the power frequency current waveform and the non-power frequency current waveform, the waveform is sent to the input layer. At the same time, the relevant parameters of the hidden layer are initialized, the number of neurons in the hidden layer is determined, and the input weights and the initial value of the hidden layer threshold are randomly generated. The disturbance waveform data is filed and the information fusion request of other related data is accepted;

[0036] A2. Select the hidden layer activation function and solve the output matrix H of the waveform data in the hidden layer;

[0037] A3. Output layer weight matrix Solve and complete the sorting of effective data of multi-frequency current waveforms, and output the weight matrix of the layer The calculation expression is as follows:

[0038]

[0039] Among them, H + is the Moore-Penrose generalized inverse matrix of the matrix H, Is β i Output weight inverse matrix, Y is the expected output matrix of the output layer;

[0040] The specific process of the waveform processing module is as follows:

[0041] B1. Process the waveform data of the new energy distribution network to eliminate waveform distortion caused by electromagnetic interference and system disturbance;

[0042] B2. Perform waveform peak sharpening to find hidden peaks in the current waveform, encode the parameters of the current waveform data of the new energy distribution network, set the maximum threshold of the number of fitting iterations and the standard of fitting convergence, and use real number encoding to randomly generate Data individuals are used as the initial group G(0);

[0043] B3. Decomposition of multi-frequency waveform data based on genetic algorithm.

[0044] As a preferred technical solution of the present invention, the specific expression of the total energy of the Jth layer obtained by the current characteristic energy spectrum conversion module after wavelet packet decomposition is as follows:

[0045]

[0046] Among them, S0 represents the discrete information sequence, S J,0 , S J,1 ,..., Respectively represent the 0th, 1st, ..., 2nd of the Jth layer J -1 node wavelet discrete information sequence, ||*|| 2 represents the square of the Euclid norm;

[0047] The specific implementation steps of the current characteristic energy spectrum extraction module are as follows:

[0048] C1. Perform the J-th wavelet packet decomposition on the collected multi-frequency current waveform fault signal of the new energy distribution network to extract the specific frequency characteristic energy spectrum;

[0049] C2, reconstructing the waveform of data of different frequencies generated by wavelet packet decomposition to remove interference signals;

[0050] C3. Obtain the current waveform energy spectrum contained in each reconstructed frequency. The specific expression is as follows:

[0051]

[0052] Among them, E J,n It represents the energy occupied by the specific frequency of the wavelet of the nth node in the Jth layer, represents the amplitude of k discrete points of the reconstructed signal of n nodes in the Jth layer, |*| 2 represents the energy of the wavelet packet decomposition coefficients, g and O represent the number of levels of the wavelet packet respectively;

[0053] C4. Construct the wavelet packet energy spectrum vector T. The specific expression is as follows:

[0054]

[0055] Where T represents the wavelet packet energy spectrum vector, E J,0 ,E J,1 ,..., Represents the 0th, 1st, ..., 2nd of the Jth layer respectively J -1 The energy occupied by the specific frequency of the wavelet node, and normalize the wavelet packet energy spectrum vector T:

[0056]

[0057] Among them, T' represents the energy spectrum feature as the input of the classifier, E represents energy, and E J,0 ,E J,1 ,..., Represents the 0th, 1st, ..., 2nd of the Jth layer respectively J -1 The energy occupied by a specific frequency of the node wavelet.

[0058] As a preferred technical solution of the present invention, the current characteristic energy spectrum lateral comparison module converts the characteristic energy T of the nth current 1 / 4 wave into n ' and the characteristic energy T of the n-1th current quarter wave n-1 'For horizontal comparison, when T n '≥1.5T n-1 ', it is judged that the new energy distribution network system is abnormal.

[0059] As a preferred technical solution of the present invention, the current characteristic energy spectrum database module uses the FCM clustering algorithm to train the characteristic energy spectrum library, and the specific steps are as follows:

[0060] D1. Initialization, giving the characteristic energy cluster number c, iteration standard ε, and membership matrix U 0 and number of iterations l = 1;

[0061] D2. Calculate the cluster center of the characteristic energy. The specific expression is as follows:

[0062]

[0063] Among them, V represents the membership of the τth cluster, the cluster center matrix of the feature energy after l iterations, and u τk is the membership of the kth sample data to the τth cluster, x k The kth sample data, the superscript l represents the number of iterations, and m represents the fuzziness weighting index;

[0064] D3. Calculate the membership matrix and objective function value of energy features as follows:

[0065]

[0066] Among them, u τk is the membership of the kth sample data to the τth cluster, d τk =||x k -v τ || is the Euclidean distance from the kth sample data to the τth cluster center, ||*|| represents the cluster center v τ With sample x k The superscript l indicates the number of iterations, J l (U l ,V l ) represents the optimization objective function iterated l times in the whole clustering process;

[0067] D4, iterative error judgment, until the results of the two iterations tend to be stable, that is The iteration ends, and when l=l+1, it returns to step D2 to continue the iteration;

[0068] D5. Determine the radius of each cluster, use the farthest distance within the cluster as the cluster radius, obtain the membership matrix U after iteration, and classify the energy feature sample data into the class with the largest membership;

[0069] The specific steps of the current characteristic energy spectrum detection module are as follows:

[0070] E1. The data identification method based on the measurement equation needs to statistically organize the detected current waveform characteristic energy spectrum and calculate the threshold range of the data to form a feasible domain matrix. If the identified target data cannot find the corresponding called data in the database, it needs to be transferred to the abnormal database for matching. If it still cannot be matched, it is identified as abnormal data.

[0071] E2. Input the characteristic energy data into the training database and calculate each cluster center. If it falls into the radius of the corresponding cluster center, it is marked as a normal process. Otherwise, the identified bad data needs to be further matched and input into the established cluster center again. The known characteristic energy data is brought in as a new class and re-clustered using a new target optimization iterative function to generate a more accurate cluster center. If the identified bad data still does not fall into the radius of the cluster center, it is determined to be a matching failure and will be judged as abnormal data.

[0072] E3. For data that is judged to be abnormal, it is marked as interference data and the process needs to be stopped and isolated;

[0073] The current characteristic energy spectrum transverse comparison module compares the decomposed current waveform characteristic energy spectrum data of different frequencies with the data in the characteristic energy spectrum database vertically, and judges it as a sign of abnormality when T2≥1.5T1 or T2≤0.5T1 is satisfied, where T2 is the current characteristic energy value of the specific frequency, and T1 is the current characteristic energy value of the specific frequency in the current waveform characteristic energy spectrum database or historical data.

[0074] Compared with the prior art, the present invention provides a new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification, which has the following beneficial effects:

[0075] 1. The present invention decomposes the multi-frequency current waveform through wavelet packets to achieve the decomposition of the full-frequency current waveform of high frequency, medium frequency and low frequency. Secondly, the characteristic energy of the integral of the current waveform of a specific frequency is compared horizontally between the front and back half waves. Finally, a multi-frequency current characteristic energy database and a historical database are formed through cluster analysis training data for longitudinal comparison with the current characteristic energy data collected in real time, which ultimately significantly improves the protection speed.

[0076] 2. The present invention uses the characteristic energy data of the current waveform as the current protection criterion for the new energy distribution network, which makes up for the access of distributed new energy power sources in the new energy distribution network. The distribution network changes from a single power network to a multi-power network, which affects the protection measurement current, making it possible for the traditional fixed value coordination protection to refuse to operate or malfunction. The characteristic energy data of the current protection is highly sensitive and can reliably reflect various internal faults of various distributed new energy distribution networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a schematic diagram of the system of the present invention;

[0078] Figure 2 This is a schematic diagram of the power frequency current waveform acquisition unit of the present invention;

[0079] Figure 3 This is a schematic diagram of a non-industrial frequency current waveform acquisition unit of the present invention;

[0080] Figure 4 It is a structural schematic diagram of the extreme learning machine of the present invention;

[0081] Figure 5 This is a schematic diagram of the current wave characteristic energy spectrum detection process of the present invention;

[0082] Figure 6 It is a schematic diagram of the longitudinal discrimination process of the present invention;

[0083] Figure 7 This is a schematic diagram of current protection for the new energy distribution network of the present invention;

[0084] Figure 8 It is a schematic diagram of the execution flow of current protection of the new energy distribution network of the present invention. DETAILED DESCRIPTION

[0085] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] See also Figure 1-Figure 8, a new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification, including a multi-frequency current waveform acquisition module, a waveform sorting module, a waveform processing module, a current characteristic energy spectrum conversion module, a current characteristic energy spectrum extraction module, a current characteristic energy spectrum multi-resolution module, a fault type identification module, a current characteristic energy spectrum horizontal comparison module, a current characteristic energy spectrum vertical comparison module, a current characteristic energy spectrum database module, a current characteristic energy spectrum detection module, a new energy distribution network overcurrent protection judgment logic module and a system execution module;

[0087] The multi-frequency current waveform acquisition module is used to acquire and store the current waveform of the distribution network. The waveform sorting module sorts and classifies the multi-frequency current waveform through the extreme learning machine algorithm. The multi-frequency current waveform acquisition module includes a power frequency current waveform acquisition unit and a non-power frequency current waveform acquisition unit. The power frequency current waveform acquisition unit and the non-power frequency current waveform acquisition unit respectively acquire the power frequency current waveform and the non-power frequency current waveform data.

[0088] The waveform processing module is used to pre-process waveform data, eliminate waveform distortion caused by electromagnetic interference and system disturbance, and decompose multi-frequency data;

[0089] The current characteristic energy spectrum conversion module converts the decomposed waveform into a characteristic energy spectrum through wavelet packets. The current characteristic energy spectrum extraction module uses the characteristic energy spectrum extraction amount as the input of the classifier. The current characteristic energy spectrum multi-resolution module is used to identify the decomposition amount of the minimum frequency.

[0090] The fault type identification module identifies the faults of the new energy distribution network through the characteristic energy spectrum. At the same time, different faults identified by the fault type identification module are further identified by the current characteristic energy spectrum horizontal comparison module and the current characteristic energy spectrum vertical comparison module respectively;

[0091] The current characteristic energy spectrum lateral comparison module is used to process the characteristic energy spectrum of the first 1 / 4 cycle and the last 1 / 4 cycle of the current waveform of a specific frequency and then output a judgment instruction;

[0092] The process of further identifying faults by the current characteristic energy spectrum longitudinal comparison module is as follows: training the characteristic energy spectrum through the current characteristic energy spectrum database module, detecting the effective characteristic energy spectrum through the current characteristic energy spectrum detection module, and the current characteristic energy spectrum longitudinal comparison module longitudinally comparing with the historical characteristic energy spectrum database and outputting a judgment instruction;

[0093] The judgment instructions output by the current characteristic energy spectrum horizontal comparison module and the current characteristic energy spectrum vertical comparison module are transmitted to the new energy distribution network overcurrent protection judgment logic module, and the data of the new energy distribution network overcurrent protection judgment logic module are satisfied by the judgment instructions or current quantity judgment, and finally output to the system execution module.

[0094] Furthermore, the power frequency current waveform acquisition unit introduces the DFT orthogonal coefficients after the waveform input passes through the analog low-pass filter, and finally outputs the power frequency current waveform through the synchronous sampling unit and the power frequency vector calculation digital filter. The power frequency vector calculation digital filter uses a Hamming window low-pass filter for filtering calculation, and its specific expression is as follows:

[0095]

[0096] Where X(i) represents the independent variable of the Hanning window function, i represents the position of the window, h=0,1,2…N-1 represents the width of the window function, N represents the window length of the window function, f0 represents the sampling frequency, W k represents the coefficient of the Hamming window function, e represents the natural logarithm, the superscript j represents the imaginary unit, and Gain represents the rectangular window function;

[0097] The non-power frequency current waveform acquisition unit calculates the rotation factor of the input waveform data through the synchronous sampling unit, and then completes the non-power frequency current waveform synchronous vector acquisition in the range of 2 to 2.5kHz through the windowed FFT algorithm combined with spectrum correction. The non-power frequency current waveform synchronous acquisition data applies the Hanning window spectrum function, and its specific expression is as follows:

[0098]

[0099] Among them, W H (ω) represents the Hanning window spectrum characteristic function, ω∈[-π,π] represents the angular frequency, N represents the window length of the window function, and w R (*) represents the continuous spectrum of the Hanning window, and w R The expression of (ω) is as follows:

[0100]

[0101] Wherein, sin(*) represents the sine function, e represents the natural logarithm, ω represents the angular frequency, and N represents the window length of the window function;

[0102] When the actual frequency of the non-power frequency current signal received by the non-power frequency current waveform acquisition unit is not the frequency resolution of FFT When the actual frequency position of the signal is an integer multiple of f, the amplitude ratio of the two adjacent maximum spectral lines in the main lobe is used to estimate the actual frequency position of the signal. r The specific expression is as follows:

[0103]

[0104] in, Indicates the frequency point, Indicates The resolution of frequency points, Indicates The signal time at each frequency point, Represents an integer value, Represents the actual frequency component f r The closest FFT frequency point, Represents the actual frequency component f r relatively, The offset of δ represents the frequency offset, and its specific expression is as follows:

[0105]

[0106] in, Indicates The amplitude of the spectral line, Indicates The amplitude of the spectral line, Indicates The amplitude of the spectral line, |*| represents the absolute value of the spectral line amplitude.

[0107] Furthermore, the extreme learning machine algorithm in the waveform sorting module includes an input layer, a hidden layer, and an output layer. The mathematical model of the extreme learning machine algorithm is as follows:

[0108]

[0109] Simplifying the expression, we get:

[0110] Hβ=Y

[0111] Among them, H represents the hidden layer output matrix, β represents the hidden layer output weight matrix, Y represents the output layer expected output matrix, β σ represents the output weight from the σth hidden layer node to the output neuron, ω σ represents the input weight from the input neuron to the σth hidden layer node, b σ represents the threshold of the σth hidden layer node, y j Represents the output value of the jth training sample. The expression of the hidden layer output matrix H is as follows:

[0112]

[0113] Among them, f(*) represents the activation function of the hidden layer, and the steps of waveform sorting module are as follows:

[0114] A1. After information fusion diagnosis of the power frequency current waveform and the non-power frequency current waveform, the waveform is sent to the input layer. At the same time, the relevant parameters of the hidden layer are initialized, the number of neurons in the hidden layer is determined, and the input weights and the initial value of the hidden layer threshold are randomly generated. The disturbance waveform data is filed and the information fusion request of other related data is accepted;

[0115] A2. Select the hidden layer activation function and solve the output matrix H of the waveform data in the hidden layer;

[0116] A3. Output layer weight matrix Solve and complete the sorting of effective data of multi-frequency current waveforms, and output the weight matrix of the layer The calculation expression is as follows:

[0117]

[0118] Among them, H + is the Moore-Penrose generalized inverse matrix of the matrix H, Is β i Output weight inverse matrix, Y is the expected output matrix of the output layer;

[0119] The specific process of the waveform processing module is as follows:

[0120] B1. Process the waveform data of the new energy distribution network to eliminate waveform distortion caused by electromagnetic interference and system disturbance;

[0121] B2. Perform waveform peak sharpening to find hidden peaks in the current waveform, encode the parameters of the current waveform data of the new energy distribution network, set the maximum threshold of the number of fitting iterations and the standard of fitting convergence, and use real number encoding to randomly generate Data individuals are used as the initial group G(0);

[0122] B3. Decomposition of multi-frequency waveform data based on genetic algorithm.

[0123] The specific formula is as follows:

[0124]

[0125] In the formula, is the individual after the crossover operation, and x B are two randomly selected waveform data individuals before the crossover operation, t is the number of generations of the genetic operation, α is the constant of the crossover operation, and its value range is (0, 1). α is obtained by multiple calculations; (h) The mutation operation randomly changes the initial parameters according to certain rules to obtain new individuals. The formula for the mutation operation is as follows:

[0126]

[0127] in, is the individual after mutation operation, is the individual before mutation, k is a mutation constant with a value range of (0, 1), xmax and x min are the maximum and minimum values ​​of the parameters in the individual, respectively.

[0128] When the number of iterations reaches the set threshold or the fitting converges, the current waveform data individuals with different frequencies with the maximum fitness are output, and there is no need to provide accurate initialization. The probabilistic transfer rule replaces the deterministic rule, which has the characteristics of global optimization. The waveform after preprocessing and waveform fitting is relatively smooth;

[0129] Wavelet packet transform is to add further decomposition of high-frequency signals on the basis of wavelet transform. Therefore, the basis of wavelet packet transform is to analyze the calculation method of wavelet transform. The basic algorithm of wavelet packet transform is as follows:

[0130] For the function ψ(t), its Fourier transform is ψ(ω) if and only if its Fourier transform satisfies the condition of the inversion source function of the wavelet transform, as shown in the formula The admissible conditions shown in the figure can be used to determine the fundamental wave or mother wave of wavelet packet transform, where ω represents the angular frequency.

[0131] After scaling and translating the base wavelet ψ(t), we can get the wavelet function:

[0132]

[0133] Where: a,b (t) is a family of continuous wavelet functions or wavelet basis functions, where a represents a scaling factor for scaling transformation, b represents a translation factor for translation transformation, and t is an independent variable of the basis wavelet function; and a, b∈R and a>0.

[0134] Since the signal energy is limited, all signals satisfy the formula The function set of the condition is L2(R), then any signal f(t)∈L2(R) is decomposed under the wavelet basis function, and the continuous wavelet transform of this decomposition into functions is called W f (a,b), the expression is

[0135]

[0136] Where: W f (a, b) is the continuous wavelet transform of the function, where a represents the scaling factor for scaling transformation, b represents the translation factor for translation transformation, and t is the independent variable of the base wavelet function; and a, b∈R and a>0; ψ * is a conjugate function.

[0137] Therefore, the result of wavelet transform of any function is as shown in the above formula. Based on wavelet transform, a wavelet packet transform method for decomposing multi-frequency current waveform is proposed. The specific calculation is as follows: Since there is a dual-scale relationship between the scale function and the wavelet function:

[0138]

[0139] Where: h is the generated low-pass filter; g is the generated high-pass filter; u0(n) is the 0th layer approximation coefficient of the original signal, a0(k) is the 0th layer approximation coefficient of the original signal, describing the low-frequency part of the signal, and u1(n) is the 1st layer detail coefficient of the original signal, describing the high-frequency part of the signal.

[0140] Among them are, It can be inferred that:

[0141]

[0142]

[0143] Defined u n (t), n = 2m or n = 2m + 1, m = 0, 1, 2 ..., called orthogonal scaling function The wavelet packet, from which we can infer that the decomposition algorithm of the wavelet packet is:

[0144]

[0145] Where: is the J-th layer approximation coefficient of the original signal, is the detail coefficient of the J+1th layer of the original signal, where J, k are any positive integers that discretize a and b, and J is the number of decomposition layers;

[0146] The wavelet packet decomposition algorithm can be used to effectively decompose the multi-frequency current aliasing waves of the current waveform, preparing for the subsequent transformation of the current waveform characteristic energy spectrum;

[0147] Furthermore, the current characteristic energy spectrum conversion module mainly carries out energy conversion of the decomposed current waveform. Since all the information of the multi-frequency current waveform decomposed by the wavelet packet can be found in the asynchronous frequencies, the information contained in each frequency after the wavelet packet decomposition can be used as the characteristic quantity, that is, the energy contained in each frequency is used to express it. Therefore, the wavelet packet energy represents the proportion of the information carried by different frequency bands in the original information during the analysis process. The specific definition is as follows:

[0148] Let the discrete information sequence S0 = {s r ,k∈Q}, define the wavelet packet decomposition operator as follows:

[0149]

[0150] Among them, S 1,0 is the scale function, S 1,1 is the wavelet function, h r-2l is the low-pass filter coefficient, g r-2l is the high-pass filter coefficient, s r is the value of the wavelet packet information sequence at a specific frequency.

[0151] The information contained in each frequency, that is, the energy of each frequency, is defined as the square of the Euclid norm ||s0|| 2 , since the total amount of information remains unchanged, we can know that:

[0152] ||S0|| 2 =||S 1,0 || 2 +||S 1,1 || 2

[0153] It can be obtained that the total amount of energy of the Jth layer after wavelet packet decomposition is:

[0154]

[0155] Among them, S0 represents the discrete information sequence, S J,0 , S J,1 ,..., Represents the 0th, 1st, ..., 2nd of the Jth layer respectively J -1 node wavelet discrete information sequence, ||*|| 2 represents the square of the Euclid norm;

[0156] The specific implementation steps of the current characteristic energy spectrum extraction module are as follows:

[0157] C1. Perform the J-th wavelet packet decomposition on the collected multi-frequency current waveform fault signal of the new energy distribution network to extract the specific frequency characteristic energy spectrum;

[0158] C2, reconstructing the waveform of data of different frequencies generated by wavelet packet decomposition to remove interference signals;

[0159] C3. Obtain the current waveform energy spectrum contained in each reconstructed frequency. The specific expression is as follows:

[0160]

[0161] Among them, E J,n It represents the energy occupied by the specific frequency of the wavelet of the nth node in the Jth layer, represents the amplitude of k discrete points of the reconstructed signal of n nodes in the Jth layer, |*| 2represents the energy of the wavelet packet decomposition coefficients, g and O represent the number of levels of the wavelet packet respectively;

[0162] C4. Construct the wavelet packet energy spectrum vector T. The specific expression is as follows:

[0163]

[0164] Where T represents the wavelet packet energy spectrum vector, E J,0 ,E J,1 ,..., Represents the 0th, 1st, ..., 2nd of the Jth layer respectively J -1 The energy occupied by the specific frequency of the wavelet node, and normalize the wavelet packet energy spectrum vector T:

[0165]

[0166] Among them, T' represents the energy spectrum feature as the input of the classifier, E represents energy, and E J,0 ,E J,1 ,..., Represents the 0th, 1st, ..., 2nd of the Jth layer respectively J -1 The energy occupied by a specific frequency of the node wavelet.

[0167] The current characteristic energy spectrum multi-resolution module can identify the minimum frequency decomposition, that is, to scale the power grid signal to obtain different frequency bands containing different frequency signal components of the original signal, and then use the decomposition information of each frequency band to reconstruct each sub-band to obtain each harmonic in the power grid signal. Generally, the higher the resolution, the higher the accuracy. However, in addition to being affected by the resolution, the accuracy is also affected by other factors, such as poor linearity. Therefore, the modulus maximum operation needs to be performed on the obvious change points of the decomposed specific frequency current waveform, and the singular value is further identified. The closed subspace V that satisfies all the properties of the following characteristics is analyzed. J , to achieve effective identification of the characteristic energy spectrum of the current waveform. The specific adjustment steps are as follows:

[0168] The basic principle of wavelet multi-resolution is to scale the power grid signal to obtain different frequency bands containing different frequency signal components of the original signal, and then use the decomposition information of each frequency band to perform multi-resolution analysis on each sub-band.

[0169] The first step is to meet the consistent monotonicity of the current waveform at a specific frequency: It means that from negative infinity subspace, it is truly contained in the subspace larger than it until positive infinity, such as: V -1 is truly contained in V0, V1 is truly contained in V0, and so on until V -∞ , while V -∞ ={0},V∞ ={L 2 (R)}, 1, 2, ... determine the width of the information space.

[0170] The second step is to meet the asymptotic completeness of the current waveform at a specific frequency: The intersection of closed subspaces corresponding to any J is 0, and the union of closed subspaces corresponding to any J is L 2 (R) is a closed space with integrity.

[0171] The third step is to meet the regularity of the current waveform at a specific frequency: f(t) is a waveform function of a specific frequency, and the waveform has longitudinal continuous scalability as time changes;

[0172] The fourth step is to satisfy the translation invariance of the current waveform at a specific frequency: f(t) is a waveform function of a specific frequency in any closed subspace V0 given any J, which has time translation invariance;

[0173] The fifth step is to satisfy the existence of an orthogonal basis for the current waveform of a specific frequency: there exists φ(t)∈V0 such that {φ(tn)} n∈Z is an orthogonal basis of φ(t)∈V0, where the inner product vector space spanned by φ(tn) can be defined as V0, that is:

[0174]

[0175] Where: V0 is the inner product vector space, φ(tn) is the standard orthogonal polynomial of V0; if δ m,n In the space spanned by φ(tn) and φ(tm), it can be expressed as:

[0176] ∫ R φ(tn)φ(tm)dt=δ m,n

[0177] Where: φ(tn) and φ(tm) are scale functions, δ m,n is a vector space function.

[0178] The fault type identification module identifies and analyzes different fault categories, and decomposes the signal into components of different frequency channels through wavelet packets. The output frequency component information contains rich system operation information, and the energy change of a certain frequency may represent a fault. Since the fault information of the new energy distribution network system is manifested as transient aliasing information, its duration is very short, and the energy of the signal is usually concentrated on the energy part sampling points of each frequency, the signals of different frequencies in the time domain have block sparse characteristics.

[0179] When the distribution network system fails, the collected distribution network system information is usually composed of multi-frequency current fault waves and interference waves, namely:

[0180] y=x+e

[0181] Among them, y represents the original signal collected, x represents the multi-frequency current fault wave, and e represents the interference wave. After processing the original information and filtering out the interference wave, the characteristic energy spectrum obtained by decomposing the multi-frequency fault current wave has a periodic block sparse structure. At a specific frequency, a binary periodic sequence is used to describe the periodic block sparse structure. In each period, there will be a binary sequence p to describe the block sparseness, that is:

[0182]

[0183] Where p is the period of the specific frequency wave, [*] T Represents the transposition operation, assuming that if each periodic fault current wave appears in a size of N l In the block, followed by N2=R-N1 zeros, where R=N1+N2, the fault wave block sparse structure that combines the binary sequence within the period can be used to judge the fault through the sparse block structure shown in the specific frequency and the changes in the energy sparse blocks of several fault types that often occur in the new energy distribution network system. That is, when a single-phase ground short circuit occurs, the characteristic energy sparse block of the fault phase current becomes significantly larger. When a two-phase short circuit occurs, the characteristic energy sparse block of the fault two-phase current becomes significantly larger, and the other phases remain unchanged. When a two-phase ground short circuit occurs, the characteristic energy sparse block of the current of the two fault phases becomes significantly larger. When a three-phase short circuit occurs, the characteristic energy sparse block of the three-phase current becomes significantly larger.

[0184] Furthermore, the current characteristic energy spectrum lateral comparison module mainly uses Newton-Cotes to perform integration processing on the current waveform of a specific frequency.

[0185] The general integral form of the multi-frequency waveform decomposed by wavelet packet is:

[0186]

[0187] In the formula, from the perspective of approximate calculation, it is to appropriately select several points x on the interval [a, b] i , and use the function values ​​f(x i ) is used to obtain the approximate value of the definite integral by the weighted average method, λ i (i=0,1,...,n) is called the quadrature coefficient, x i (i=0,1,...,n) is called the quadrature node, and there is

[0188] a≤x0≤x1≤...≤x n ≤ b

[0189] λ i (i=0,1,2,.......n) is the quadrature coefficient and has nothing to do with the integrand f(x). The truncation error of the quadrature formula is

[0190]

[0191] Where R n is the cutoff error value.

[0192] The Newton-Cotes quadrature formula can be derived using the interpolation method.

[0193]

[0194] The quadrature coefficient is

[0195]

[0196] In the formula, It is called the Cotes coefficient, i = 0, 1, ..., n, j = 0, 1, 2, ..., n, divide the integral interval [a, b] n into equal parts, h = (ba) / n

[0197] By integrating the current waveform of a specific frequency through the above integral formula, the integrated area data is the characteristic energy spectrum.

[0198] The characteristic energy T of the nth current quarter wave n ' and the characteristic energy T of the n-1th current quarter wave n-1 'For horizontal comparison, when T n '≥1.5T n-1 'When judging that the new energy distribution network system is abnormal, or T n ≥1.5T n-1 It can judge whether there is an abnormality in the new energy distribution network system.

[0199] Furthermore, the current characteristic energy spectrum database module uses the FCM clustering algorithm to train the characteristic energy spectrum library. The specific sample setting number is n, the characteristic index is p, and therefore p = 3. The energy characteristic clustering sample matrix is ​​as follows:

[0200]

[0201] After clustering, the energy feature samples are divided into c (2≤c≤n) fuzzy classes, and their cluster centers are expressed by the cluster center matrix V as follows:

[0202] V={v1,v2,…,v c}

[0203] Among them, V τ =(v τ1 ,v τ2 ,…,v τp ),τ=1,2,…,c.

[0204] The membership matrix u τk Representing the degree of membership of sample data and clustering, the following formula can be obtained:

[0205]

[0206] Where: u τk is the membership of the kth energy feature sample data to the τth cluster. The objective function of the FCM algorithm is as follows:

[0207]

[0208] Where: d τk =||x k -v τ || is the Euclidean distance from the kth characteristic energy sample data to the τth cluster center; m is the fuzziness weighting index, which is generally 2. Here, each dimension of the original energy feature sample, that is, the attribute of each system call, is given the same weight, without highlighting the system calls that play a decisive role in classification. In order to better distinguish normal and abnormal energy feature samples, different system call data need to be given different weights. Set the weight vector W = {w1,w2,…,w s}, the new error sum of squares formula is:

[0209]

[0210] The "." here refers to the Hadamard (or Schur) Product in matrix or vector operations;

[0211] The specific steps based on training are as follows:

[0212] D1. Initialization, giving the characteristic energy cluster number c, iteration standard ε, and membership matrix U 0 and number of iterations l = 1;

[0213] D2. Calculate the cluster center of the characteristic energy. The specific expression is as follows:

[0214]

[0215] Among them, V represents the membership of the τth cluster, the cluster center matrix of the feature energy after l iterations, and u τk is the membership of the kth sample data to the τth cluster, x kThe kth sample data, the superscript l represents the number of iterations, and m represents the fuzziness weighting index;

[0216] D3. Calculate the membership matrix and objective function value of energy features as follows:

[0217]

[0218] Among them, u τk is the membership of the kth sample data to the τth cluster, d τk =||x k -v τ || is the Euclidean distance from the kth sample data to the τth cluster center, ||*|| represents the cluster center v τ With sample x k The superscript l indicates the number of iterations, J l (U l ,V l ) represents the optimization objective function iterated l times in the whole clustering process;

[0219] D4, iterative error judgment, until the results of the two iterations tend to be stable, that is The iteration ends, and when l=l+1, it returns to step D2 to continue the iteration;

[0220] D5. Determine the radius of each cluster, use the farthest distance within the cluster as the cluster radius, obtain the membership matrix U, and classify the energy feature sample data into the class with the largest membership;

[0221] The specific steps of the current characteristic energy spectrum detection module are as follows:

[0222] E1. The data identification method based on the measurement equation needs to statistically organize the detected current waveform characteristic energy spectrum and calculate the threshold range of the data to form a feasible domain matrix. If the identified target data cannot find the corresponding called data in the database, it needs to be transferred to the abnormal database for matching. If it still cannot be matched, it is identified as abnormal data.

[0223] E2. Input the characteristic energy data into the training database and calculate each cluster center. If it falls into the radius of the corresponding cluster center, it is marked as a normal process. Otherwise, the identified bad data needs to be further matched and input into the established cluster center again. The known characteristic energy data is brought in as a new class and re-clustered using a new target optimization iterative function to generate a more accurate cluster center. If the identified bad data still does not fall into the radius of the cluster center, it is determined to be a matching failure and will be judged as abnormal data.

[0224] E3. For data that is judged to be abnormal, it is marked as interference data and the process needs to be stopped and isolated;

[0225] The current characteristic energy spectrum longitudinal comparison module forms a new energy distribution network current waveform characteristic energy spectrum database after optimizing and training different clustering data, and longitudinally compares the decomposed current waveform characteristic energy spectrum data with the data in the characteristic energy spectrum database, that is, by regularly detecting the current energy characteristics of the same frequency, observe whether its energy characteristics have an obvious upward or downward trend. If the change value exceeds 50%, it can be determined that the new energy distribution network system is abnormal or faulty. At the same time, through real-time comparison and judgment with the historical current waveform characteristic energy spectrum, if the current characteristic energy of the distribution network system shows a trend of increasing or decreasing, it means that the system has shown signs of abnormality. It longitudinally compares the decomposed current waveform characteristic energy spectrum data with the data in the characteristic energy spectrum database. When T2≥1.5T1 or T2≤0.5T1 is satisfied, it is judged as a sign of abnormality, where T2 is the current characteristic energy value of the current specific frequency, and T1 is the current characteristic energy value of the current waveform characteristic energy spectrum database or the historical data specific frequency current characteristic energy value;

[0226] Current protection scheme for new energy distribution network Figure 7 As shown, in order to further ensure the reliability of the current characteristic energy data judgment, the current protection of the new energy distribution network takes the conventional current overcurrent protection as an auxiliary judgment criterion, and the current waveform characteristic energy spectrum data as the main judgment criterion. When any of the current waveform transverse characteristic energy spectrum, the current waveform longitudinal characteristic energy spectrum, and the conventional overcurrent protection conditions are activated, the current protection of the new energy distribution network will be delayed or the outlet will trip;

[0227] Energy distribution network current protection execution process is as follows Figure 8 As shown, the sampling current waveform data is collected, and the sampling current and current waveform characteristic energy data are judged to be greater than the preset threshold value through the new energy distribution current protection scheme. When it is greater than the set threshold value, the system outputs instructions to the execution module;

[0228] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification, characterized by: It includes a multi-frequency current waveform acquisition module, a waveform sorting module, a waveform processing module, a current characteristic energy spectrum conversion module, a current characteristic energy spectrum extraction module, a current characteristic energy spectrum multi-resolution module, a fault type identification module, a current characteristic energy spectrum horizontal comparison module, a current characteristic energy spectrum vertical comparison module, a current characteristic energy spectrum database module, a current characteristic energy spectrum detection module, a new energy distribution network overcurrent protection judgment logic module and a system execution module; The multi-frequency current waveform acquisition module is used to acquire and store the current waveform of the power distribution network. The waveform sorting module sorts and classifies the multi-frequency current waveform through the extreme learning machine algorithm. The multi-frequency current waveform acquisition module includes a power frequency current waveform acquisition unit and a non-power frequency current waveform acquisition unit. The power frequency current waveform acquisition unit and the non-power frequency current waveform acquisition unit respectively acquire the power frequency current waveform and the non-power frequency current waveform data. The waveform processing module is used to pre-process the waveform data, eliminate the waveform distortion caused by electromagnetic interference and system disturbance, and decompose the multi-frequency data; The current characteristic energy spectrum conversion module performs characteristic energy spectrum conversion on the decomposed waveform through wavelet packets, the current characteristic energy spectrum extraction module uses the characteristic energy spectrum extraction amount as the input of the classifier, and the current characteristic energy spectrum multi-resolution module is used to identify the decomposition amount of the minimum frequency; The fault type identification module identifies the fault of the new energy distribution network through the characteristic energy spectrum, and at the same time, different faults identified by the fault type identification module are further identified by the current characteristic energy spectrum horizontal comparison module and the current characteristic energy spectrum vertical comparison module respectively; The current characteristic energy spectrum lateral comparison module is used to process the characteristic energy spectra of the first 1 / 4 cycle and the last 1 / 4 cycle of the current waveform of a specific frequency and then output a judgment instruction; The process of further identifying faults by the current characteristic energy spectrum longitudinal comparison module is as follows: training the characteristic energy spectrum through the current characteristic energy spectrum database module, detecting the effective characteristic energy spectrum through the current characteristic energy spectrum detection module, and the current characteristic energy spectrum longitudinal comparison module longitudinally comparing with the historical characteristic energy spectrum database and outputting a judgment instruction; The judgment instructions output by the current characteristic energy spectrum horizontal comparison module and the current characteristic energy spectrum vertical comparison module are transmitted to the new energy distribution network overcurrent protection judgment logic module, and the data of the new energy distribution network overcurrent protection judgment logic module are satisfied by the judgment instructions or current quantity judgment, and finally output to the system execution module.

2. The new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification according to claim 1 is characterized by: The power frequency current waveform acquisition unit introduces the DFT orthogonal coefficients after the waveform input passes through the analog low-pass filter, and finally outputs the power frequency current waveform through the synchronous sampling unit and the power frequency vector calculation digital filter. The power frequency vector calculation digital filter uses a Hamming window low-pass filter for filtering calculation, and its specific expression is as follows: Where X(i) represents the independent variable of the Hanning window function, i represents the position of the window, h=0,1,2…N-1 represents the width of the window function, N represents the window length of the window function, f0 represents the sampling frequency, W k represents the coefficient of the Hamming window function, e represents the natural logarithm, and the superscript represents the imaginary unit, Gain represents the rectangular window function; The non-power frequency current waveform acquisition unit calculates the rotation factor of the input waveform data through the synchronous sampling unit, and then completes the non-power frequency current waveform synchronous vector acquisition in the range of 2 to 2.5 kHz by combining the windowed FFT algorithm with spectrum correction, wherein the non-power frequency current waveform synchronous acquisition data applies the Hanning window spectrum function, and its specific expression is as follows: Among them, W H (ω) represents the Hanning window spectrum characteristic function, ω∈[-π,π] represents the angular frequency, N represents the window length of the window function, and w R (*) represents the continuous spectrum of the Hanning window, and w R The expression of (ω) is as follows: Wherein, sin(*) represents the sine function, e represents the natural logarithm, ω represents the angular frequency, and N represents the window length of the window function; When the actual frequency of the non-power frequency current signal received by the non-power frequency current waveform acquisition unit is not the frequency resolution F of FFT n When the actual frequency position of the signal is an integer multiple of f, the amplitude ratio of the two adjacent maximum spectral lines in the main lobe is used to estimate the actual frequency position of the signal. r The specific expression is as follows: f r =(i m +δ)F n Among them, F n =1 / T n , F n Indicates the frequency resolution, T n represents the signal time, i m Represents an integer value, i m F n Represents the actual frequency component f r The closest FFT frequency, δF n Represents the actual frequency component f r Relatively speaking, m F n The offset of δ represents the frequency offset, and its specific expression is as follows: Among them, X m represents the amplitude of the mth spectral line, X m+1 represents the amplitude of the m+1th spectral line, X m-1 represents the amplitude of the m-1th spectral line, and |*| represents the absolute value of the spectral line amplitude.

3. The new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification according to claim 1 is characterized by: The extreme learning machine algorithm in the waveform sorting module includes an input layer, a hidden layer and an output layer. The mathematical model of the extreme learning machine algorithm is as follows: Simplifying the expression, we get: Hβ=Y Among them, H represents the hidden layer output matrix, β represents the hidden layer output weight matrix, Y represents the output layer expected output matrix, β σ represents the output weight from the σth hidden layer node to the output neuron, ω σ represents the input weight from the input neuron to the σth hidden layer node, b σ represents the threshold of the σth hidden layer node, y j Represents the output value of the jth training sample. The expression of the hidden layer output matrix H is as follows: Wherein, f(*) represents the activation function of the hidden layer, and the steps of the waveform sorting module are as follows: A1. After information fusion diagnosis of the power frequency current waveform and the non-power frequency current waveform, the waveform is sent to the input layer. At the same time, the relevant parameters of the hidden layer are initialized, the number of neurons in the hidden layer is determined, and the input weights and the initial value of the hidden layer threshold are randomly generated. The disturbance waveform data is filed and the information fusion request of other related data is accepted; A2. Select the hidden layer activation function and solve the output matrix H of the waveform data in the hidden layer; A3. Output layer weight matrix Solve and complete the sorting of effective data of multi-frequency current waveforms, and output the weight matrix of the layer The calculation expression is as follows: Among them, H + is the Moore-Penrose generalized inverse matrix of the matrix H, Is β i Output weight inverse matrix, Y is the expected output matrix of the output layer; The specific process of the waveform processing module is as follows: B1. Process the waveform data of the new energy distribution network to eliminate waveform distortion caused by electromagnetic interference and system disturbance; B2. Perform waveform peak sharpening to find hidden peaks in the current waveform, encode the parameters of the current waveform data of the new energy distribution network, set the maximum threshold of the number of fitting iterations and the standard of fitting convergence, and use real number encoding to randomly generate Data individuals are used as the initial group G(0); B3. Decomposition of multi-frequency waveform data based on genetic algorithm.

4. The new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification according to claim 3 is characterized by: The specific expression of the total energy of the Jth layer obtained by the current characteristic energy spectrum conversion module after wavelet packet decomposition is as follows: Among them, S0 represents the discrete information sequence, Represents the 0th, 1st, ..., 2nd of the Jth layer respectively J -1 node wavelet discrete information sequence, ||*|| 2 represents the square of the Euclid norm; The specific implementation steps of the current characteristic energy spectrum extraction module are as follows: C1. Perform the J-th wavelet packet decomposition on the collected multi-frequency current waveform fault signal of the new energy distribution network to extract the specific frequency characteristic energy spectrum; C2, reconstructing the waveform of data of different frequencies generated by wavelet packet decomposition to remove interference signals; C3. Obtain the current waveform energy spectrum contained in each reconstructed frequency. The specific expression is as follows: Among them, E J,n It represents the energy occupied by the specific frequency of the wavelet of the nth node in the Jth layer, represents the reconstruction signal of n nodes in the Jth layer, |*| 2 represents the energy of the wavelet packet decomposition coefficients, g and O represent the number of levels of the wavelet packet respectively; C4. Construct the wavelet packet energy spectrum vector T. The specific expression is as follows: Where T represents the wavelet packet energy spectrum vector, Represents the 0th, 1st, ..., 2nd of the Jth layer respectively J -1 The energy occupied by the specific frequency of the wavelet node, and normalize the wavelet packet energy spectrum vector T: Among them, T' represents the energy spectrum feature as the input of the classifier, E represents energy, Represents the 0th, 1st, ..., 2nd of the Jth layer respectively J -1 The energy occupied by a specific frequency of the node wavelet.

5. The new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification according to claim 4 is characterized in that: The current characteristic energy spectrum lateral comparison module converts the characteristic energy T of the nth current 1 / 4 wave into n ' and the characteristic energy T of the n-1th current quarter wave n-1 'For horizontal comparison, when T n '≥1.5T n-1 ', it is judged that the new energy distribution network system is abnormal.

6. The new energy distribution network overcurrent protection system for characteristic energy spectrum fault identification according to claim 1 is characterized by: The current characteristic energy spectrum database module uses the FCM clustering algorithm to train the characteristic energy spectrum library. The specific steps are as follows: D1. Initialization, giving the characteristic energy cluster number c, iteration standard ε, and membership matrix U 0 and number of iterations l = 1; D2. Calculate the cluster center of the characteristic energy. The specific expression is as follows: Among them, V represents the membership of the τth cluster, the cluster center matrix of the feature energy after l iterations, and u τk is the membership of the kth sample data to the τth cluster, x k The kth sample data, the superscript l represents the number of iterations, and m represents the fuzziness weighting index; D3. Calculate the membership matrix and objective function value of energy features as follows: Among them, u τk is the membership of the kth sample data to the τth cluster, d τk =||x k -v τ || is the Euclidean distance from the kth sample data to the τth cluster center, ||*|| represents the cluster center v τ With sample x k The superscript l indicates the number of iterations, J l (U l ,V l ) represents the optimization objective function iterated l times in the whole clustering process; D4, iterative error judgment, until the results of the two iterations tend to be stable, that is The iteration ends, and when l=l+1, it returns to step D2 to continue the iteration; D5. Determine the radius of each cluster, use the farthest distance within the cluster as the cluster radius, obtain the membership matrix U, and classify the energy feature sample data into the class with the largest membership; The specific steps of the current characteristic energy spectrum detection module are as follows: E1. The data identification method based on the measurement equation needs to statistically organize the detected current waveform characteristic energy spectrum and calculate the threshold range of the data to form a feasible domain matrix. If the identified target data cannot find the corresponding called data in the database, it needs to be transferred to the abnormal database for matching. If it still cannot be matched, it is identified as abnormal data. E2. Input the characteristic energy data into the training database and calculate each cluster center. If it falls within the radius of the corresponding cluster center, it is marked as a normal process. Otherwise, the identified bad data needs to be further matched. If the match fails, it will be judged as abnormal data. E3. For data that is judged to be abnormal, it is marked as interference data and the process needs to be stopped and isolated; The current characteristic energy spectrum transverse comparison module compares the decomposed current waveform characteristic energy spectrum data of different frequencies with the data in the characteristic energy spectrum database vertically, and judges it as a sign of abnormality when T2≥1.5T1 or T2≤0.5T1 is satisfied, where T2 is the current characteristic energy value of the specific frequency, and T1 is the current characteristic energy value of the specific frequency in the current waveform characteristic energy spectrum database or historical data.

Images