Single - ended Quantity Protection Method for the Transmission Line of New Energy Power Stations Based on Improved PCNN

Abstract

The present invention discloses a single - end quantity protection method for the outgoing line of a new - energy power station based on an improved PCNN. The steps include: collecting the original fault - current data on the new - energy power station side, and extracting features from the original fault - current data to obtain a fault - current feature vector; using the fault - current feature vector and the original fault - current data as the input data of two input channels of an improved PCNN network model after offline training, fusing the features extracted from the two channels through a feature fusion layer after the pooling layer of the improved PCNN network model, and outputting the recognition result of the fault category; performing corresponding circuit protection actions according to the recognition results of internal and external fault categories. This single - end quantity protection method for the outgoing line of the new - energy power station can identify internal and external faults under various working conditions and has good protection performance.

Description

Technical Field

[0001] The present invention relates to a circuit protection method, in particular to a single-ended quantity protection method for the outgoing line of a new energy power station based on an improved PCNN. Background Art

[0002] To solve the problem of huge consumption of non-renewable energy, the utilization of renewable energy has developed rapidly. For example, in the power industry, environmentally friendly and green power generation technologies such as wind energy and photovoltaic energy are widely used in the power system. When centralized new energy is connected to the grid line, a large number of power electronic devices are included, and the fault current waveform is significantly affected by the control strategy in the power electronic device. The fault current exhibits characteristics such as limited amplitude, phase change, and frequency offset, while the threshold setting of traditional protection methods is complex and highly dependent on manual experience.

[0003] Currently, the protection of the outgoing line of a new energy power station refers to the experience of conventional AC lines. The main protection uses distance protection and pilot protection, and the backup protection uses current protection. The overcurrent protection method is vulnerable to the weak feed effect of new energy, resulting in a decrease in the protection range. The distance protection performance deteriorates under high transition resistance conditions. These protection methods have an adverse impact on the large-scale application of new energy.

[0004] With the development of the new generation of artificial intelligence technology, protection methods based on deep learning algorithms are applied in the field of relay protection. This type of protection method can avoid complex threshold setting. In the prior art, there is a method that uses the voltage backward traveling wave as the input stream of the deep learning model to classify fault identification and pole selection. This protection method can achieve fault discrimination in a short time; there is also a method that uses a neural network for fault detection and location, but the training workload is very large; there is also a single-ended quantity protection scheme based on wavelet energy feature extraction and support vector machine, which can accurately identify faults under various working conditions, but the accuracy is relatively low during small sample training.

[0005] The deep learning algorithm can be used for the processing of feature data, which is beneficial to the fitting of non-linear mapping relationships. Compared with threshold analysis and manual judgment, it can improve the calculation performance and recognition accuracy. Summary of the Invention

[0006] The object of the invention is to provide a single-ended quantity protection method for the outgoing line of a new energy power station based on an improved PCNN, which can identify internal and external faults under various working conditions and has good protection performance.

[0007] Technical Solution: The single-ended quantity protection method for the outgoing line of a new energy power station based on an improved PCNN according to the present invention includes the following steps:

[0008] Step 1, collect the original fault current data on the new energy power station side, and extract features from the original fault current data to obtain a fault current feature vector;

[0009] Step 2: Use the fault current feature vector and the original fault current data as the input data for the two input channels of the improved PCNN network model after offline training. After the pooling layer of the improved PCNN network model, fuse the features extracted from the two channels through the feature fusion layer, and output the recognition result of the fault type.

[0010] Step 3: Perform corresponding circuit protection actions according to the recognition results of internal and external fault types.

[0011] Furthermore, in Step 1, the specific steps for collecting the original fault current data on the new energy power station side are as follows:

[0012] Collect the current on the new energy power station side in real time, and judge whether a fault has occurred according to the protection startup criterion formula:

[0013] Δ I f >0.1 I n , where in the formula, I n is the rated current on the power station side, and Δ I f is the sudden change of the fault current on the power station side;

[0014] If the protection startup criterion formula holds, it is determined that a fault has occurred, and immediately intercept the three-phase current data on the new energy power station side within 10 ms after the current moment.

[0015] Furthermore, in Step 1, the calculation steps for the sudden change of the fault current on the power station side are as follows:

[0016] Step 1.1: Set the acquisition frequency of the current on the power station side to P 0 , and then set the sampling time width for current mutation sampling to Lm , and Lm ≥4 / P 0 ;

[0017] Step 1.2: Collect the currents of each phase on the power station side in real time according to the set acquisition frequency;

[0018] Step 1.3: Analyze and judge the collected real-time current acquisition values. If the difference between the real-time current acquisition value of a certain phase current collected in the later time and the real-time current acquisition value collected in the previous time is ≥ the amplitude growth threshold, go to Step 1.4; otherwise, return to Step 1.2;

[0019] Step 1.4: Start a timer, and when the timing of the timer reaches Lm , obtain the data from the current acquisition moment forwardLm Each real-time current acquisition value collected within the range, and a current data set corresponding to each timer is established in sequence according to the sequence of the start times of each timer;

[0020] Step 1.5, after the current data set is established, determine whether the number of all current data sets is greater than E , if the number is greater than E , then calculate the current average value of each real-time current acquisition value in the newly established current data set I 0 , and then enter Step 1.6, otherwise return to Step 1.2;

[0021] Step 1.6, analyze and judge the current average value I 0 . If the current average values corresponding to each current data set I 0 gradually increase and the maximum difference between the start times of each timer is less than the set start time difference threshold, then select the first real-time current acquisition value in the first established current data set as the first current value, and then select the last real-time current acquisition value in the newly established current data set as the second current value, and then enter Step 1.7, otherwise delete each current data set and then return to Step 1.2;

[0022] Step 1.7, after obtaining the second current value, calculate the difference between the second current value and the first current value as Δ I 0 , analyze and judge Δ I 0 . If Δ I 0 > 0.1 I n , then assign Δ I 0 to Δ I f . If Δ I 0 ≤0.1 I n , then return to Step 1.2.

[0023] Furthermore, in Step 1, the specific steps for extracting the fault current feature vector from the original fault current data are as follows:

[0024] First, extract the optimal IMF components of each group of original fault current data according to the minimum information entropy criterion;

[0025] Then, calculate the nine parameter indicators of each optimal IMF component respectively. The nine parameter indicators are mean, variance, peak value, kurtosis, effective value, peak factor, impulse factor, waveform factor, and margin factor;

[0026] Finally, use the nine parameter indicators to construct the feature vector of the original data of each group of fault currents.

[0027] Furthermore, in step 1.1, the specific steps for extracting the optimal IMF components of the original data of each group of fault currents according to the minimum information entropy criterion are as follows:

[0028] First, use the empirical mode decomposition model to decompose each IMF component from the original data of the current group of fault currents;

[0029] Then, according to the minimum information entropy criterion, find the optimal IMF component of the original data of the current group of fault currents from the decomposed IMF components;

[0030] Repeat the above two steps until the optimal IMF components of the original data of each group of fault currents are all extracted.

[0031] Furthermore, the specific steps for using the empirical mode decomposition model to decompose each IMF component from the original data of the current group of fault currents are as follows:

[0032] First, construct the empirical mode decomposition model. The specific model formula is: , , , where is the constraint condition, min {} represents taking the minimum value, u k ={ u 1 , u 2 ,…, u k} is the mode function, f （t） is the fault current signal, u k ( t ) is the original data of the fault current, k is the number of modes, δ （t） is the Dirac distribution function, ∂ t is the partial derivative, * is the convolution operation, w k represents the center frequencies of each mode, w k ={ w 1 ,w 2 ,…, w k} and using the exponent to correct so that the spectrum of each modal function is modulated to the corresponding base frequency band α is the quadratic penalty factor λ is the Lagrange multiplier is the augmented Lagrangian function;

[0033] Then, the original data of the current group of fault currents is sent into the constructed empirical mode decomposition model, and the initial value of the number of modes of the empirical mode decomposition model is set;

[0034] Finally, the empirical mode decomposition model decomposes the corresponding number of each IMF component from the original data of the current group of fault currents according to the initially set number of modes.

[0035] Furthermore, the specific steps to find the best IMF component of the original data of the current group of fault currents from the decomposed IMF components according to the minimum information entropy criterion are as follows:

[0036] First, record the kurtosis values in each IMF component of each decomposed group, and then select the number of modes corresponding to the peak of the kurtosis value as the preferred number of modes;

[0037] Then, construct a parameter optimization model to optimize the number of modes. The parameter optimization model is n a population optimization model composed of only canaries. The model formula of the population optimization model is: , where d represents the preferred number of modes d ∈[0,1,2,…] n represents the number of canaries x 1,1 ~ x n,d respectively represent each canary. The matrix form of the fitness values of all canaries is: , where f is the fitness value. The formula for updating the position of the discoverer is constructed as: , where t is the current iteration number j =1,2,…, d , x max is the maximum iteration number X i,j represents the i th j dimensional position information of the α and r 2All are warning values, α∈ [0, 1], r 2 ∈ [0, 1], CT is the safety value, CT∈ [0.5, 1], Q is a random number that follows a normal distribution, L represents a 1× d matrix, and each element in the matrix L is all 1;

[0038] Construct the position update formula for the joiner: , where, is the i new position of the j th canary in the first stage and the t dimensional, SF i,j is the selected food, r i,j is a random number between [0, 1], I i,j is a random number in the set {1, 2};

[0039] When danger is detected, the canaries at the edge of the group will quickly move towards the safe area to obtain a better position, while the canaries in the middle of the population will move randomly. The mathematical expression is: , where the canary individual with the lowest fitness value is used as the global canary optimal position, X best is the current global canary optimal position, β As the step size control parameter, it is a random number that follows a normal distribution with a mean of 0 and a variance of 1, K is a random number, K∈ [0, 1], s i is the fitness value of the current canary individual, s g and s w are the current global best and worst fitness values respectively, ε is the smallest constant to avoid a zero denominator;

[0040] Finally, take the X best value at the global optimal position d as the best component position value, and then find the best IMF component at the corresponding position from the decomposed IMF components according to the best component position value.

[0041] Further, in step 2, the improved PCNN network model includes a feature fusion layer, a fully connected layer, a multi-head attention mechanism, a Softmax classifier, an output layer, and two convolutional pooling branches; the convolutional pooling branches include a convolutional layer, a pooling layer, and an acceleration layer connected in sequence; the two convolutional pooling branches respectively receive the fault current feature vector and the original fault current data, where the convolutional layer is used to perform convolutional calculations on the fault current feature vector and the original fault current data, and respectively outputs a set of one-dimensional vectors, the pooling layer is used to perform average pooling operations on the output vectors, the acceleration layer is used to improve the calculation efficiency, the feature fusion layer is used to splice and fuse the two sets of one-dimensional vectors and then send them to the fully connected layer, and then the fully connected layer performs probability calculation processing and sends it to the multi-head attention mechanism. After the feature enhancement processing by the multi-head attention mechanism, it is sent to the Softmax classifier, and after the probability classification processing by the Softmax classifier, it is finally output by the output layer.

[0042] Further, in step 3, when performing corresponding circuit protection actions according to the recognition results of the fault types in the zone, for a ground fault of phase A in the zone, the phase A fault protection device trips; for a ground fault of phase B in the zone, the phase B fault protection device trips; for a ground fault of phase C in the zone, the phase C fault protection device trips; for a multi-phase fault in the zone, all three-phase fault protection devices trip; for a fault outside the zone, the in-zone fault protection device does not operate.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows: By collecting the fault current on the new energy power station side, improving the variational mode decomposition (VMD) to extract the features of the fault current, and then using the extracted feature vector and the fault current data as the inputs of the two channels of the model, adding a feature fusion layer after the pooling layer, fusing the features extracted from the two channels and introducing a multi-head self-attention mechanism for training, the distinguishability of various fault characteristics is improved, and the recognition of in-zone and out-of-zone fault types and corresponding protection actions are realized.

[0044] BRIEF DESCRIPTION OF THE DRAWINGS is the flowchart of the method of the present invention;

[0045] Figure 1 is the structural diagram of the improved PCNN network model of the present invention;

[0046] Figure 2 is the outgoing line diagram of the new energy power station in the embodiment of the present invention;

[0047] Figure 3 is the curve diagram of the loss value change of the training set and the test set in the embodiment of the present invention;

[0048] Figure 4 is the curve diagram of the loss value change of the training set and the test set in the embodiment of the present invention;

[0049] Figure 5 It is the curve graph of the accuracy rate change of the training set and the test set in the embodiment of the present invention;

[0050] Figure 6 It is the classification result graph of the test set in the embodiment of the present invention;

[0051] Figure 7 It is the classification result graph of different new energy power station capacities in the embodiment of the present invention;

[0052] Figure 8 It is the test result graph of different fault positions in the embodiment of the present invention;

[0053] Figure 9 It is the classification result graph of different transition resistances in the embodiment of the present invention;

[0054] Figure 10 It is the test result graph of different noise intensities in the embodiment of the present invention;

[0055] Figure 11 It is the different network result graph in the embodiment of the present invention;

[0056] Figure 12 It is the PCNN output result graph in the embodiment of the present invention;

[0057] Figure 13 It is the single-channel CNN output result graph in the embodiment of the present invention. Detailed implementation manners

[0058] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the described embodiments.

[0059] As Figure 1 shown, the single-terminal quantity protection method for the outgoing line of a new energy power station based on an improved PCNN disclosed by the present invention includes the following steps:

[0060] Step 1, collect the original fault current data on the new energy power station side, and perform feature extraction on the original fault current data to obtain a fault current feature vector;

[0061] Step 2, use the fault current feature vector and the original fault current data as the input data of two input channels of the improved PCNN network model after offline training, and fuse the features extracted from the two channels through a feature fusion layer after the pooling layer of the improved PCNN network model, and output the recognition result of the internal and external fault categories;

[0062] Step 3, perform corresponding circuit protection actions according to the recognition result of the internal and external fault categories.

[0063] Further, in step 1, the specific steps for collecting the original fault current data on the new energy power station side are as follows:

[0064] Collect the current on the new energy power station side in real time, and determine whether a fault has occurred according to the protection startup criterion formula:

[0065] Δ I f >0.1 I n , where, I n is the rated current on the power station side, and Δ I f is the sudden change of the fault current on the power station side;

[0066] If the protection startup criterion formula holds, it is determined that a fault has occurred, and the three-phase current data on the new energy power station side within 10 ms after the current moment is immediately intercepted.

[0067] Further, in step 1, the calculation steps for the sudden change of the fault current on the power station side are as follows:

[0068] Step 1.1, set the acquisition frequency of the current on the power station side to P 0 , and then set the sampling time width of the current mutation sampling to Lm , and Lm ≥4 / P 0 ;

[0069] Step 1.2, collect the currents of each phase on the power station side in real time according to the set acquisition frequency;

[0070] Step 1.3, analyze and judge the collected real-time current acquisition values. If the difference between the real-time current acquisition value of a certain phase current collected in the later time and the real-time current acquisition value of the previous collection is ≥ the amplitude growth threshold, and the amplitude growth threshold is 0.02 I n , then enter step 1.4, otherwise return to step 1.2;

[0071] Step 1.4, start a timer, and when the timing of the timer reaches Lm , obtain the real-time current acquisition values collected within the range of Lm forward from the current acquisition moment, and establish a current data set corresponding to each timer in the order of the start time of each timer;

[0072] Step 1.5, after the current data set is established, judge whether the number of all current data sets is greater than E , E Preferably 2, if the number is greater than E, the average current value of each real-time current acquisition value in the newly established current data set is calculated I 0 , then go to step 1.6, otherwise return to step 1.2;

[0073] Step 1.6, for the average current value I 0 Perform an analysis and judgment. If the average current values corresponding to each current data set I 0 gradually increase and the maximum difference between the start times of each timer is less than the set start time difference threshold, and the start time difference threshold is 0.15 s. Using the judgment of mean increase and start time difference threshold can avoid misjudging the situation where the current waveform shows a segmented abnormal mutation. Then select the first real-time current acquisition value in the first established current data set as the first current value, and then select the last real-time current acquisition value in the newly established current data set as the second current value, and then go to step 1.7, otherwise delete each current data set and return to step 1.2;

[0074] Step 1.7, after obtaining the second current value, calculate the difference between the second current value and the first current value as Δ I 0 , for Δ I 0 Perform an analysis and judgment. If Δ I 0 > 0.1 I n , then assign Δ I 0 to Δ I f , if Δ I 0 ≤0.1 I n , then return to step 1.2.

[0075] Furthermore, in step 1, the specific steps for extracting the fault current feature vector from the original fault current data are as follows:

[0076] First, extract the optimal IMF components of each group of original fault current data according to the minimum information entropy criterion;

[0077] Then calculate the nine parameter indicators of each optimal IMF component respectively. The nine parameter indicators are mean, variance, peak value, kurtosis, effective value, peak factor, impulse factor, waveform factor and margin factor;

[0078] Finally, use the nine parameter indicators to construct the feature vector of each group of original fault current data.

[0079] Further, in step 1.1, the specific steps for extracting the optimal IMF components of each group of original fault current data according to the minimum information entropy criterion are as follows:

[0080] First, use the empirical mode decomposition model to decompose each IMF component from the original fault current data of the current group;

[0081] Then, according to the minimum information entropy criterion, find the optimal IMF component of the original fault current data of the current group from the decomposed IMF components;

[0082] Repeat the above two steps until the optimal IMF components of each group of original fault current data are extracted.

[0083] Further, the specific steps for using the empirical mode decomposition model to decompose each IMF component from the original fault current data of the current group are as follows:

[0084] First, construct the empirical mode decomposition model. The specific model formula is: , , , where, is the constraint condition, min {} represents taking the minimum value, u k ={ u 1 , u 2 , …, u k} is the mode function, f （t） is the fault current signal, u k ( t ) is the original fault current data, k is the number of modes, δ （t） is the Dirac distribution function, ∂ t is the partial derivative, * is the convolution operation, w k represents the center frequency of each mode, w k ={ w 1 , w 2 , …, w k}, use the exponential correction to modulate the spectrum of each mode function to the corresponding baseband, α is the quadratic penalty factor, λ is the Lagrange multiplier, is the augmented Lagrangian function;

[0085] Then, the original data of the current group of fault currents is fed into the constructed empirical mode decomposition model, and the initial value of the number of modes of the empirical mode decomposition model is set;

[0086] Finally, the empirical mode decomposition model decomposes the corresponding number of each IMF component from the original data of the current group of fault currents according to the initially set number of modes.

[0087] Furthermore, the specific steps to find the optimal IMF component of the original data of the current group of fault currents from the decomposed IMF components according to the minimum information entropy criterion are as follows:

[0088] First, record the kurtosis values in each IMF component of each decomposed group, and then select the number of modes corresponding to the peak of the kurtosis value as the preferred number of modes;

[0089] Then, a parameter optimization model is constructed to optimize the number of modes. The parameter optimization model is n a population optimization model composed of only canaries. The model formula of the population optimization model is: , where d represents the preferred number of modes, d ∈[0,1,2,…], n represents the number of canaries, x 1,1 ~ x n,d respectively represent each canary. The matrix form of the fitness values of all canaries is: , where f is the fitness value. The discovery position update formula is constructed as: , where t is the current iteration number, j =1,2,…, d , x max is the maximum iteration number, X i,j represents the i th j dimensional position information of the α and r 2 are both early warning values, α∈ [0,1], r 2 ∈ [0,1], CT is the safety value, CT∈ [0.5,1], Q is a random number obeying the normal distribution, L represents a 1×d The matrix, where each element in the matrix L is all 1;

[0090] Construct the position update formula for the joiner: , where, for the i th canary at the first stage, the j dimensional new position, t is the current iteration number, SF i,j is the selected food, r i,j is a random number between [0, 1], I i,j is a random number in the set {1, 2};

[0091] When aware of danger, the canaries at the edge of the group will quickly move towards the safe area to obtain a better position, while the canaries located in the middle of the population will move randomly. The mathematical expression is: , where the canary individual with the lowest fitness value is used as the global canary optimal position, X best is the current global canary optimal position, β As the step size control parameter, is a random number that follows a normal distribution with a mean of 0 and a variance of 1, K is a random number, K∈ [0, 1], s i is the fitness value of the current canary individual, s g and s w are the current global best and worst fitness values respectively, ε is the smallest constant to avoid a zero denominator;

[0092] Finally, take the X best value at the global optimal position d as the best component position value, and then find the best IMF component at the corresponding position from the decomposed IMF components according to the best component position value.

[0093] Furthermore, in step 2, the improved PCNN network model includes a feature fusion layer, a fully connected layer, a multi-head attention mechanism, a Softmax classifier, an output layer, and two convolutional pooling branches, as Figure 2As shown in the figure; the convolutional pooling branch includes a convolutional layer, a pooling layer, and an acceleration layer connected in sequence; two convolutional pooling branches respectively receive the fault current feature vector and the original fault current data. The convolutional layer is used to perform convolutional calculations on the fault current feature vector and the original fault current data, and respectively output a set of one-dimensional vectors. The pooling layer is used to perform average pooling operations on the output vectors. The acceleration layer is used to improve the calculation efficiency. The feature fusion layer is used to splice and fuse the two sets of one-dimensional vectors and then send them to the fully connected layer. Then, the fully connected layer performs probability calculation processing and sends it to the multi-head attention mechanism. After feature enhancement processing by the multi-head attention mechanism, it is sent to the Softmax classifier. After probability classification processing by the Softmax classifier, it is finally output by the output layer.

[0094] Further, in step 2, the recognition results of the fault categories output include in-zone phase A ground fault, in-zone phase B ground fault, in-zone phase C ground fault, in-zone multi-phase fault, and out-of-zone fault.

[0095] Further, in step 3, when corresponding circuit protection actions are performed according to the recognition results of the in-zone fault categories, for an in-zone phase A ground fault, the phase A fault protection device trips; for an in-zone phase B ground fault, the phase B fault protection device trips; for an in-zone phase C ground fault, the phase C fault protection device trips; for an in-zone multi-phase fault, all three-phase fault protection devices trip; for an out-of-zone fault, the in-zone fault protection device does not operate.

[0096] To verify the reliability of the single-terminal quantity protection method for the outgoing line of the new energy power station of the present invention, the following simulation experiments are carried out:

[0097] First, establish the structure of the new energy outgoing line, as Figure 3 shown. Build a new energy outgoing line model in PSCAD / EMTDC. The main parameters are as follows: the voltage level of the outgoing line is 220 kV, the rated capacity of the new energy station is 100 MW. The external system capacity is 400 MVA, the positive sequence impedance is 0.5 Ω, the positive sequence impedance phase angle is 80°, the zero sequence impedance is 1 Ω, and the frequency is 50 Hz. The total length of the outgoing line is 100 km. The outgoing line is an overhead transmission line, with a positive sequence resistance of 6.76×10 -8 pu / m, a positive sequence inductive reactance of 9.6×10 -7 pu / m, a positive sequence capacitive reactance of 5.78×10 5 pu·m, a zero sequence resistance of 6.86×10 -7 pu / m, a zero sequence inductive reactance of 2.5×10 -6 pu / m, and a zero sequence capacitive reactance of 8.14×10 5The main transformer has a capacity of 500 MVA, and the rated capacity of the box-type transformer is 300 MVA. A large number of simulation experiments are carried out under different fault conditions. The sampling frequency is set to 4 kHz, the fault time is 8 s, and the original fault current data within 10 ms after the fault is intercepted. Feature extraction is performed on the original fault current data, and the fault current feature vector and the original fault current data are input into the deep learning model for waveform feature mining and training. According to the protection action principle of the 220 kV transmission line, it is divided into single-phase faults within the zone (A-phase grounding, B-phase grounding, C-phase grounding), denoted as AG, BG, CG respectively, multi-phase faults within the zone and out-of-zone faults, a total of 5 fault types. The transition resistance value range is 1 - 100 Ω, considering a total of 10 different cases; the fault distance is set to 10 - 100 km, considering a total of 10 different cases; the new energy station capacity is 75 - 150 MW, considering a total of 5 different cases. Therefore, the original fault current data is 5×10×10×5 = 2500 groups. The improved VMD is used for feature extraction to obtain the fault current feature vector. The original fault current data is divided according to the ratio of 7:3, and the division is shown in Table 1:

[0098] Table 1 is the data sample division table

[0099]

[0100] The Adam optimizer is used for optimization, and the initial learning rate is set to 0.001. The loss function is the cross-entropy function. The network training is completed after 90 iterations, and the learning rate is adjusted 60 times. The accuracy curve and the loss value curve are as Figure 4 and Figure 5 shown.

[0101] From Figure 5 it can be concluded that the prediction accuracy of the test set accounting for 30% of the sample data reaches 100%. Table 2 is the specific classification results of different fault categories in the sample test set, Figure 6 as shown in the test set classification result diagram.

[0102] Table 2 is the protection action result table for different fault categories in the sample test

[0103]

[0104] For large-scale new energy stations connected to the power grid line, to test the influence of the new energy station capacity on the protection action performance, the test samples with new energy station capacities of 120 MW, 160 MW, 180 MW, 200 MW, 220 MW, and 260 MW are input into the deep learning network for test experiments, Figure 7 are the classification results for different new energy station capacities. Table 3 is the analysis result of the protection action performance under different new energy station capacities.

[0105] Table 3 is the protection action result table for different new energy power station capacities

[0106]

[0107] From Figure 7 and Table 3, it can be concluded that under different new energy power station capacities, faults can be correctly identified and the protection operates reliably.

[0108] In the outgoing lines of new energy power stations, the protection can still operate reliably in case of different fault locations. To test the protection action performance of the entire length of the outgoing line, 10 groups of test samples with different fault locations are randomly input into the deep learning network for testing experiments. Figure 8 Table 4 is the classification result for different fault locations, and Table 4 is the analysis result of the protection action performance under different fault locations.

[0109] Table 4 is the protection action result table for different fault locations

[0110]

[0111] From Figure 8 and Table 4, it can be concluded that under different fault locations, faults can still be correctly identified and the protection operates reliably.

[0112] When a grounding resistance appears during the grounding fault of the outgoing line of a new energy power station, the maximum value of the transition resistance during the grounding fault of the outgoing line of the new energy power station is 120 Ω. To test the ability of the proposed protection method to resist the transition resistance, test samples with grounding resistances of 100 Ω, 105 Ω, 110 Ω, 115 Ω, and 120 Ω are input into the deep learning network for testing experiments. Figure 9 Table 5 is the classification result for different transition resistances, and Table 5 is the analysis result of the protection action performance under different transition resistances.

[0113] Table 5 is the protection action result table for different transition resistances

[0114]

[0115] From Figure 9 and Table 5, it can be concluded that under different fault resistances, faults can be correctly identified and the protection operates reliably.

[0116] A large amount of noise often appears in power transmission lines. To test the anti-noise ability of the proposed protection method, test samples with noise intensities of 40 db, 30 db, 20 db, and 10 db are input into the deep learning network for testing experiments. Figure 10 Table 6 is the classification result for noise tests, and Table 6 is the analysis result of the protection action performance under noise interference. From Figure 10It can be seen from Table 6 that when noise is added to the data, the faults can be correctly identified and the protection operates reliably.

[0117] Table 6 is the protection action result table for different noises

[0118]

[0119] To verify the superiority of the single-ended quantity protection based on the improved PCNN network, 10 dB noise is superimposed on the original current data of 10 ms. The data set is divided into a training set and a test set according to 7:3. The noisy data is brought into the CNN single-channel network and the PCNN network for anti-noise interference testing. The test comparison results are as Figure 11 shown. Figure 12 and Figure 13 are the output results of PCNN and single-channel CNN.

[0120] From Figure 11 and Figure 12 、 13 it can be seen that the classification accuracy of the single-channel CNN and the original PCNN is not as high as that of the method proposed in this paper. Therefore, the transmission line protection scheme for new energy power stations based on the improved PCNN can still reliably identify internal and external faults under strong noise interference, and the protection operates correctly.

[0121] In summary, aiming at the problems such as the difficulty in coordinating the four properties of the outgoing line protection for new energy power stations, the proposed single-ended quantity protection method for the outgoing line of new energy power stations based on the improved PCNN first collects the fault transient current on the new energy power station side, extracts data features, inputs the original data samples and feature vectors into the improved PCNN to deeply excavate the waveform features of the fault transient current, and finally realizes the identification of the internal and external fault categories. The results show that the proposed protection method has the following advantages:

[0122] (1) Adopting a deep learning algorithm with adaptive learning, avoiding complex calculations and manual threshold setting;

[0123] (2) Adopting a two-channel neural network. Compared with single data input, the two data types complement each other, and even if one of the data characteristics is weak, the fault characteristics can be fully excavated;

[0124] (3) As a backup protection, single-ended quantity data is used for offline training, and the detection time is 10 ms, meeting the requirement of protection quick-acting performance.

[0125] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as a limitation of the present invention itself. Various changes can be made in its form and details without departing from the spirit and scope of the present invention defined by the appended claims.

Claims

1. A single-ended quantity protection method for a new energy station transmission line based on an improved PCNN, characterized in that: The steps include: Step 1: collect the original data of fault current at the new energy station side, and extract the features of the original data of fault current to obtain the fault current feature vector; Step 2: The fault current feature vector and the original fault current data are used as the input data of the two input channels of the improved PCNN network model after offline training. After the pooling layer of the improved PCNN network model, the features extracted from the two channels are fused through the feature fusion layer to output the recognition result of the fault category. Step 3: Perform corresponding circuit protection actions according to the identification results of the fault types inside and outside the zone; In step 1, the specific steps for collecting the original data of fault current at the new energy station side are: The current on the new energy station side is collected in real time, and whether a fault occurs is determined based on the protection start criterion. The protection start criterion is: Δ I f >0.1 I n , where I n is the rated current of the station side, Δ I f is the sudden change of fault current at the station side; If the protection start criterion is established, it is determined that a fault has occurred, and the three-phase current data on the new energy station side within 10ms after the current moment is immediately intercepted; In step 1, the calculation steps of the fault current mutation amount at the station side are: Step 1.1, set the acquisition frequency of the station side current to P 0, and then set the sampling time width of the current mutation sampling to Lm ,and Lm ≥4 / P 0; Step 1.2, collect the current of each phase on the station side in real time according to the set collection frequency; Step 1.3, analyzing and judging the collected real-time current collection value, if the difference between the real-time current collection value collected the last time and the real-time current collection value collected the last time of a phase current is ≥ the amplitude growth threshold, then go to step 1.4, otherwise return to step 1.2; Step 1.4, start a timer and when the timer reaches Lm Get the number of times from the current collection time forward Lm The real-time current collection values ​​collected within the range are collected, and the current data sets corresponding to each timer are established in sequence according to the start time of each timer; Step 1.5: After the current data set is established, determine whether the number of all current data sets is greater than E , if the number is greater than E , then calculate the current average value of each real-time current acquisition value in the latest established current data set I 0, then go to step 1.6, otherwise return to step 1.2; Step 1.6, average current I 0 for analysis and judgment, if the current average value corresponding to each current data set I 0 gradually increases and the maximum difference of the start time of each timer is less than the set start time difference threshold, the first real-time current acquisition value in the first established current data set is selected as the first current value, and then the last real-time current acquisition value in the latest established current data set is selected as the second current value, and then go to step 1.7, otherwise delete each current data set and return to step 1.2; Step 1.7, after obtaining the second current value, calculate the difference between the second current value and the first current value as Δ I 0, for Δ I 0 for analysis and judgment, if Δ I 0>0.1 I n , then Δ I 0 is assigned to Δ I f , if Δ I 0≤0.1 I n , then return to step 1.

2.

2. According to the improved PCNN-based single-ended quantity protection method for new energy station transmission lines according to claim 1, it is characterized in that: In step 1, the specific steps of extracting features from the original fault current data to obtain the fault current feature vector are as follows: Firstly, the best IMF component of each group of fault current raw data is extracted according to the minimum information entropy criterion; Then, nine parameter indicators of each optimal IMF component are calculated respectively, and the nine parameter indicators are mean, variance, peak value, kurtosis, effective value, peak factor, impulse factor, waveform factor and margin factor; Finally, nine parameter indicators are used to construct the characteristic vector of each group of fault current raw data.

3. The single-ended quantity protection method for the new energy station transmission line based on the improved PCNN according to claim 2 is characterized in that: In step 1.1, the specific steps of extracting the best IMF component of each group of fault current raw data according to the minimum information entropy criterion are as follows: Firstly, the empirical mode decomposition model is used to decompose the various IMF components from the original data of the current group fault current; Then, the optimal IMF component of the original data of the current group fault current is found from the decomposed IMF components according to the minimum information entropy criterion; Repeat the above two steps until the best IMF components of each group of fault current raw data are extracted.

4. The single-ended quantity protection method for the new energy station transmission line based on the improved PCNN according to claim 3 is characterized in that: The specific steps of decomposing each IMF component from the original data of the current group fault current using the empirical mode decomposition model are as follows: First, the empirical mode decomposition model is constructed. The specific model formula is: , , , where As constraints, min {} means taking the minimum value, u k ={ u 1, u 2,…, u k } is the modal function, f （t） is the fault current signal, u k ( t ) is the original data of fault current, k is the number of modes, δ （t） is the Dirac distribution function, For partial guidance, is the convolution operation, w k represents the center frequency of each mode, w k ={ w 1, w 2,…, w k }, using index Correction, so that the spectrum of each mode function is modulated to the corresponding baseband, α is the quadratic penalty factor, λ is the Lagrange multiplication operator, is the augmented Lagrangian function; Then the original data of the current group fault current is sent to the constructed empirical mode decomposition model, and the initial value of the number of modes of the empirical mode decomposition model is set; Finally, the empirical mode decomposition model decomposes the corresponding number of IMF components from the original data of the current group fault current according to the initially set number of modes.

5. The single-ended quantity protection method for the new energy station transmission line based on the improved PCNN according to claim 3 is characterized in that: The specific steps of finding the best IMF component of the current group fault current raw data from the decomposed IMF components according to the minimum information entropy criterion are as follows: First, record the kurtosis value of each group of IMF components that have been decomposed, and then select the mode number corresponding to the peak of the kurtosis value as the preferred mode number; Then, a parameter optimization model is constructed to optimize the number of modes. The parameter optimization model is: n The population optimization model consists of canaries. The model formula of the population optimization model is: , where d represents the number of preferred modes, d ∈[0,1,2,…], n represents the number of canaries, x 1,1 ~ x n,d Represents each canary separately, and the matrix form of the fitness value of all canaries is: , where f is the fitness value, and the formula for updating the discoverer position is: , where t is the current iteration number, j =1,2,…, d , x max is the maximum number of iterations, X i,j Indicates i Canary j Location information in the dimension, α and r 2 are warning values. α∈ [0,1], r 2 ∈ [0,1], CT is a safe value, CT∈ [0.5,1], Q is a random number that follows a normal distribution, L Represents a 1× d The matrix, matrix L All elements in are 1; Construct the position update formula of the joiner: , where is i The canary in the first stage, j The new location of the t is the current iteration number, SF i,j For the selected food, r i,j is a random number between [0,1], I i,j is a random number in the set {1,2}; When aware of danger, the canaries at the edge of the group will quickly move to a safe area to get a better position, while the canaries in the middle of the group will move randomly. The mathematical expression is: , where the canary individual with the lowest fitness value is taken as the global canary optimal position, X best is the current global canary optimal position, β As the step size control parameter, it is a random number that follows a normal distribution with a mean of 0 and a variance of 1. K is a random number, K∈ [0,1], s i is the fitness value of the current canary individual, s g and s w are the current global best and worst fitness values, respectively. ε is the smallest constant that avoids zero in the denominator; Finally, the global optimal position X best Where d The value is used as the optimal component position value, and then the optimal IMF component at the corresponding position is found from the decomposed IMF components according to the optimal component position value.

6. The single-ended quantity protection method for the new energy station transmission line based on the improved PCNN according to claim 1 is characterized in that: In step 2, the improved PCNN network model includes a feature fusion layer, a fully connected layer, a multi-head attention mechanism, a Softmax classifier, an output layer and two convolutional pooling branches; the convolutional pooling branch includes sequentially connected convolutional layers, pooling layers and acceleration layers; the two convolutional pooling branches receive the fault current feature vector and the fault current original data respectively, wherein the convolutional layer is used to perform convolution calculations on the fault current feature vector and the fault current original data, and output a group of one-dimensional vectors respectively, the pooling layer is used to perform average pooling operations on the output vectors, the acceleration layer is used to improve the calculation efficiency, the feature fusion layer is used to concatenate and fuse the two groups of one-dimensional vectors and send them to the fully connected layer, and then the fully connected layer performs probability calculation processing and then sends them to the multi-head attention mechanism, and then the multi-head attention mechanism performs feature enhancement processing and then sends them to the Softmax classifier, and the Softmax classifier performs probability classification processing and finally outputs them from the output layer.

7. The single-ended quantity protection method for the new energy station transmission line based on the improved PCNN according to claim 1 is characterized in that: In step 2, the identification results of the output fault categories include an A-phase grounding fault within the zone, a B-phase grounding fault within the zone, a C-phase grounding fault within the zone, a multi-phase fault within the zone, and an out-of-zone fault.

8. The single-ended quantity protection method for the new energy station transmission line based on the improved PCNN according to claim 7 is characterized in that: In step 3, when the corresponding circuit protection action is performed according to the identification result of the fault type in the zone, if there is a phase A grounding fault in the zone, the phase A fault protection device will trip, if there is a phase B grounding fault in the zone, the phase B fault protection device will trip, if there is a phase C grounding fault in the zone, the phase C fault protection device will trip, if there is a multi-phase fault in the zone, all three-phase fault protection devices will trip, and if there is a fault outside the zone, the fault protection device in the zone will not operate.

Images