Active power distribution network fault positioning method and system based on improved convolutional neural network

Abstract

The invention discloses an active power distribution network fault positioning method and system based on an improved convolutional neural network. According to the method, fault data are obtained through an active power distribution network off-line simulation model, a band-pass filter is adopted to filter power frequency components and high-frequency interference in the fault data, then normalization processing is carried out, transient extraction transformation is utilized to convert one-dimensional fault data into a two-dimensional time-frequency graph, and a training and verification sample set is formed. A convolutional neural network model containing a channel prior convolutional attention module is built, and a K-fold cross validation method is utilized to complete model training. When the model is used for new fault data, fault line and fault section information is output at the same time, and then the fault occurrence position is determined by integrating line selection and section selection results. According to the method, a mode of combining a band-pass filtering technology and transient extraction transformation is adopted to enhance the fault feature characterization capability of a sample, and a channel prior attention mechanism is introduced, so that the model training efficiency and the fault positioning accuracy are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault line selection and section location for active distribution networks, and particularly relates to a single-phase grounding fault line selection and section location method in the scenario of clean energy access to an active distribution network. Background Technique

[0002] With the large-scale access of clean energy such as wind energy and solar energy as distributed power sources, the traditional medium-voltage distribution network has evolved into an active distribution network containing multiple power sources. The active distribution network has characteristics such as bidirectional power flow and complex topological structure. When a single-phase grounding fault occurs, the fault signal distortion is more serious, and new characteristics such as weak fault characteristics and complex transient high-frequency interference appear, making it more difficult to achieve accurate fault location.

[0003] According to the types and methods of using signals, the existing fault location methods can be divided into steady-state quantity methods, transient quantity methods, traveling wave methods, injection methods, and artificial intelligence methods.

[0004] The steady-state quantity methods include the steady-state zero-sequence power direction method, the steady-state zero-sequence current distribution characteristic method, the three-phase current amplitude difference method, and the zero-sequence admittance angle difference method, etc. Although the steady-state quantity exists throughout the fault duration, its amplitude is small, resulting in a low fault location accuracy for this type of method. When the neutral point is grounded through an arc suppression coil, the amplitude of the fault steady-state quantity is even lower, and it is easy to fail in fault location. The access of distributed power sources causes the distribution characteristics of fault steady-state quantities in the active distribution network to be different from those of the traditional distribution network. Therefore, the performance of the steady-state quantity method in the active distribution network seriously deteriorates, and the probability of fault location failure is higher.

[0005] The transient quantity methods include the transient current characteristic comparison method, the phase transient power direction method, and the time-frequency characteristic-based method, etc. Although the fault characteristic information contained in the transient quantity is obvious and rich, after the fault occurs, the transient component decays quickly, and then the transient quantity method fails. Moreover, the transient quantity method usually requires a high sampling frequency, resulting in a high implementation cost. The transient quantity method is also extremely sensitive to the transition resistance, fault initial phase angle, and strong noise. The differences in transient quantities during faults in different sections will be affected by these factors, leading to failure in fault location.

[0006] The traveling wave method determines the fault location by capturing the time difference of the arrival of the wavefront according to the reflection and refraction characteristics of the traveling wave. The traveling wave method has been successfully applied to fault location for high-voltage long-distance transmission lines. However, due to the access of distributed power sources, the active distribution network has the characteristics of multiple branches and multiple power sources, with many traveling wave reflections and refractions, difficult wavefront capture, and large time difference measurement errors, making it difficult to apply in the active distribution network.

[0007] The injection method determines the fault location by injecting signals of specific frequencies and detecting the changes in the injected signals. The injected signals can be integer or fractional harmonics, or pulse quantities. The injection method has a high positioning accuracy. However, the injected signals are generally weak. After being transmitted through the network, they are difficult to detect. Moreover, signal injection and detection equipment need to be configured, and the operation is relatively complex, making it difficult to use in the field.

[0008] In recent years, active distribution network fault location methods based on data-driven and deep learning have emerged. In principle, the fault location problem is converted into a multi-classification problem, automatically learning and extracting fault features from data, establishing a mapping relationship between fault features and fault locations, and the output result is the probability of each section being a fault section. However, when the deep learning-based fault location method is used in active distribution networks, the following technical problems exist:

[0009] (1) Currently, methods with relatively high fault location accuracy all require a large number of measurement points. In actual distribution networks, the number of measurement devices is limited, and it is difficult to meet the conditions for complete measurement. Therefore, it is impossible to obtain sufficient quantities and qualities of fault data to support the training of deep learning models. There is still a lack of methods that can complete accurate fault location using only a small number of measurement points.

[0010] (2) Using the original waveform or traditional time-frequency analysis methods to characterize fault features has disadvantages such as low frequency resolution and poor anti-noise performance. This makes it difficult for deep learning models to accurately identify fault features, resulting in a low fault location accuracy.

[0011] (3) Existing methods cannot accurately distinguish the key features beneficial for section location from the useless information in the fault signal, resulting in poor ability to distinguish the feature differences between fault sections and non-fault sections, low model training efficiency, poor generalization ability, and thus low positioning accuracy.

[0012] (4) When setting sample labels in existing methods, all fault sections of multiple lines are usually numbered one by one, resulting in a large number of labels. When the model outputs the probability values of fault sections, the discrimination is not high. The probability of the true fault section is close to or even smaller than the probabilities of other non-fault sections, resulting in failed fault location.

[0013] It can be seen that the existing fault location methods based on deep learning and time-frequency analysis generally have low accuracy and poor adaptability, and it is urgent to improve the existing technology to enhance the fault location performance of active distribution networks. Summary of the Invention

[0014] To overcome the above problems existing in the prior art, the present invention discloses an active distribution network fault location method and system based on an improved convolutional neural network.

[0015] The specific technical solutions adopted by the present invention are as follows:

[0016] A fault location method for active distribution networks based on an improved convolutional neural network, comprising the following steps:

[0017] 1. Obtain fault simulation data

[0018] The fault location method based on convolutional neural network requires a large amount of fault data to train the model. Since the amount of fault data accumulated on site is extremely small, simulation software is used to generate fault data for training and validating the convolutional neural network model.

[0019] Build a fault simulation model of the active distribution network in the simulation software, and the simulation conditions are specifically set as follows:

[0020] The grounding methods are respectively set to three types: ungrounded, grounded through an arc suppression coil, and grounded through a resistor; there are a total of L lines connected to the same bus, including overhead lines, cable lines, and overhead-cable hybrid lines; the fault points are set on each line, with a total of M1, and the percentage range of the distance from the fault point to the bus accounting for the total length of the line is 0% to 100%, with an interval of x1%; the setting range of the transition resistance is 0Ω to 2000Ω, with a total of M2 resistance values; the setting range of the distributed power source penetration rate is 0% to 100%, with a step of x2%; the fault initial phase angle is set to 0° to 360°, with a step of x3.

[0021] In the simulation, the zero-sequence current at the beginning of each line is collected as fault data, and the sampling frequency is f1Hz, and a total of M groups of fault data are obtained.

[0022] The naming rule for the saved fault data file is "*********_****.***", where the first 9 digits are the simulation parameter marks, the 4 digits after the underscore are reserved for the labels of the fault line and the fault section, and the last 3 digits are the file type suffix.

[0023] The meanings of the simulation parameter marks in order from left to right are: the first digit represents the neutral point grounding method, the second digit represents the over-compensation degree of the arc suppression coil, the third digit represents the resistance value of the transition resistance, the fourth digit represents the distributed power source penetration rate, the fifth digit is a reserved bit for expansion, and the sixth to ninth digits represent the sample serial number under the same fault condition, sorted by fault phase, fault location, fault initial phase angle, etc.

[0024] 2. Design a band-pass filter

[0025] Select a Butterworth band-pass filter to filter out the power frequency component and high-frequency interference in the zero-sequence current fault data.

[0026] Set the low-frequency cut-off frequency of the band-pass filter to f2Hz and the high-frequency cut-off frequency to (f3 + Δf h )Hz, where f3Hz is the lowest frequency of the available high-frequency signal in the zero-sequence current signal, and the high-frequency components higher than f3Hz are regarded as noise interference.

[0027] The corresponding low-frequency passband edge angular frequency is ω p1 = 2πf2, and the high-frequency passband edge angular frequency is ω p2 = 2π(f3 + Δf h ), the low-frequency stopband edge frequency is ω s1 = ω p1 - Δω1, and the high-frequency stopband edge frequency is ω s2 = ω p2 + Δω2.

[0028] Set the maximum attenuation a p of the passband and the minimum attenuation a s of the stopband of the filter, calculate the passband ripple coefficient ε1 and the stopband ripple coefficient ε2, and calculate the filter orders N1 and N2 determined by the low-frequency cut-off frequency and the high-frequency cut-off frequency respectively.

[0029] Take the order of the filter as:

[0030] N bp = max{N1, N2} (1)

[0031] 3. Generate time-frequency diagram samples

[0032] Preprocess the fault data obtained from the offline simulation in step 1 to generate time-frequency diagram samples for model training and testing, including the following steps: filter and normalize the fault data, convert the fault data into a transient extraction transform time-frequency diagram, label the classification labels for the time-frequency diagram samples, and divide the samples into a training set, a validation set, and a test set.

[0033] 3.1 Filter the fault data

[0034] For each set of simulation conditions set in step 1, a corresponding set of fault data can be obtained. Each set of fault data contains the zero-sequence current sampling value sequences of L lines, and these sequences are used as the input sequences of the filter.

[0035] Input the input sequences into the band-pass filter designed in step 2 in sequence. The output of the band-pass filter is the zero-sequence current data sequence after filtering out the power frequency component and high-frequency interference, which is called the data sample.

[0036] 3.2 Normalize the data samples

[0037] The mth data sample is expressed in matrix form as:

[0038]

[0039] where N is the number of current sampling values of each line in the data sample, specifically the product of the total number of power frequency cycles before and after the fault and the number of sampling points per power frequency cycle.

[0040] For data sample I m Perform maximum-minimum normalization to obtain the normalized data sample matrix I′ m , where the l-th row represents the current sampling value sequence of the l-th line after normalization, denoted as:

[0041] I′ l = {i′ l,1 , i′ l,2 ,..., i′ l,N} (3)

[0042] 3.3 Convert the data sample into a time-frequency diagram sample of transient extraction transform

[0043] Perform short-time Fourier transform on I′ l to obtain the transformation result F STFT (n, f).

[0044] Multiply F STFT (n, f) by the transient extraction transform operation factor to obtain the transient extraction transform result F l of I′ TET (n, f).

[0045] Plot the above transient extraction transform result as a two-dimensional time-frequency diagram, where the abscissa is the current sampling value serial number and the ordinate is the frequency value.

[0046] According to the above steps, sequentially complete the transient extraction transform of the L rows of data in the data sample I′ m to obtain the time-frequency diagrams of transient extraction transform for all L lines, and splice all L time-frequency diagrams into 1 time-frequency diagram sample. When splicing, perform horizontal splicing from left to right in sequence according to the line numbers and vertical splicing from top to bottom in sequence. The number of time-frequency diagrams in the horizontal and vertical directions should be as equal as possible to make the spliced time-frequency diagram as close to a square as possible.

[0047] The file name format for saving the time-frequency diagram sample is "*********_****.jpg", and the naming rule is the same as the naming rule for the fault data saving file in step 1.

[0048] 3.4 Label classification labels for the time-frequency diagram sample

[0049] The classification labels of the time-frequency diagram samples are represented in numerical form, indicating the faulty line number and section number. The faulty line numbers are 1, 2,.., l,..., L. The faulty sections are divided according to the percentage of the line length. One section is divided every x4% of the total line length, and a total of S sections are divided. The faulty section numbers are numbered in ascending order of the distance to the bus as 1, 2,.., s,..., S. The faulty line label l indicates "fault on the l-th line", and the faulty section label s indicates "fault in the s-th section of the faulty line".

[0050] The classification label is represented by 4 digits after the underscore in the time-frequency diagram sample file name. The first 2 digits of the label represent the faulty line number, and the last 2 digits represent the faulty section number.

[0051] 3.5 Divide the time-frequency diagram samples into training set, validation set and test set

[0052] After shuffling all the time-frequency diagram samples, they are divided into 2 sets according to the ratio of p1:p2. The former is used as the first set to form the training set and validation set, and the latter is used as the second set as the test set. Then divide the former set into K T subsets with the same number of samples for model training and validation, and name them J1, J2,..., J k .

[0053] 4. Build a line selection model and a section selection model

[0054] The convolutional neural network (CNN) integrating the Channel Prior Convolutional Attention (CPCA) module is simply called the CPCA-CNN model. Accordingly, a faulty line selection model and a faulty section selection model are constructed respectively. Both the faulty line selection model and the faulty section selection model contain several convolutional layers, CPCA modules, pooling layers and fully connected layers, with the numbers being n C , n R , n P and n F , respectively determined according to the principle of the highest line selection accuracy and section selection accuracy of the validation set.

[0055] The CPCA-CNN model is composed of convolutional layers, CPCA modules, pooling layers and fully connected layers connected in series in sequence. The structures of each layer and each module are as follows.

[0056] The convolutional layer completes the convolution operation through the convolution kernel. Its input is the time-frequency diagram sample S a , and the output is the feature map

[0057] The CPCA module is composed of a channel attention module and a spatial attention module connected in series, implementing a dual attention mechanism for channels and space.

[0058] The channel attention module includes two parallel pooling layers and a shared fully connected layer. Among them, the two pooling layers use max pooling and average pooling respectively. After the results of the two pooling layers are added together, they pass through the shared fully connected layer to generate the weight vector X for each channel. a , X a And After element-wise multiplication, the channel attention feature map F is output. ca .

[0059] The spatial attention module includes three parallel depthwise separable convolutional layers, and its input is F. ca , and the output is the spatial attention feature map F. sa .

[0060] F ca And F sa After element-wise multiplication, the CPCA attention feature map F is output. cpca .

[0061] The pooling layer of the CPCA-CNN model uses max pooling, and its input is F. cpca , and the output is the feature map after dimensionality reduction.

[0062] The fully connected layer of the CPCA-CNN model uses the Softmax function as the activation function, and its input is For the fault line selection model, its output is the fault probability of each line in the fault data to be identified, and the line corresponding to the maximum probability is the fault line; for the fault section selection model, its output is the fault probability of each section in the fault data to be identified, and the section corresponding to the maximum probability is the fault section.

[0063] The CPCA-CNN model can pay more attention to the key feature differences such as frequency components, frequency change rates, and energy distributions in the time-frequency diagram of zero-sequence current before and after the fault, improving the training efficiency and recognition accuracy compared with the simple CNN model.

[0064] 5. Training and testing the line selection model and the section selection model

[0065] The hyperparameters of the CPCA-CNN model include n C , n R , n P and n FThe values of, Batch size value, and Leaning rate value. Set a reasonable value range for each hyperparameter according to experience, and enumerate all combinations of hyperparameters. Under different combinations of hyperparameter values, use the fault line selection and fault section selection models trained with the first set divided in Step 3.5 to verify the line selection and section selection accuracy rates respectively, so as to select the optimal hyperparameters for the two models.

[0066] The optimization process of the hyperparameters of the fault line selection model is as follows. In the order of k from small to large, successively use the J k th sample subset as the validation set, where k = 1, 2, …, (K T -1), K T ; for each value of k, use the total of (K k -1) sample subsets except the validation set J T as the training set together. Use the above validation set and training set to complete the model training, obtain K T line selection accuracy rates and take the average value. Repeat the above steps for the models with all combinations of hyperparameter values, and the hyperparameter values corresponding to the fault line selection model with the highest average accuracy rate are the optimal hyperparameters.

[0067] The hyperparameter optimization process of the fault section selection model is the same as that of the fault line selection model.

[0068] After both the fault line selection model and the fault section selection model are trained, use the samples in the second set divided in Step 3.4 to test the fault line selection accuracy rate and the fault section selection accuracy rate respectively.

[0069] 6. Use the trained model to complete fault location

[0070] For new samples, adopt the preprocessing methods in Step 3.1, Step 3.2, and Step 3.3 to transform them into the time-frequency diagram S′ to be recognized a .

[0071] Input S′ a into the trained fault line selection model. After passing through the convolutional layer, CPCA module, pooling layer, and fully connected layer, output the probability value vector X = [x1, x2,..., x L of each line having a fault, and the line corresponding to the largest element in X is the selected fault line.

[0072] Then input S′ a into the trained fault section selection model. The process is the same as that of the fault line selection model. The fault section selection model outputs the probability value vector Y = [y1, y2,..., y S of each section having a fault, and the section corresponding to the largest element in Y is the selected fault section.

[0073] Determine the fault location based on the final comprehensive fault line selection result and fault section selection result.

[0074] The beneficial effects of the present invention include:

[0075] (1) Fault section location can be completed by only using a small number of measuring devices at the head of each line to collect zero-sequence current, without arranging a large number of measurement points in the middle of the line, reducing the number of measuring devices and having higher engineering practical value.

[0076] (2) Use a band-pass filter to filter out the power frequency component and high-frequency interference in the zero-sequence current, and adopt a transient extraction transformation technology with high time-frequency resolution to form a time-frequency map sample, enhancing the fault feature characterization ability and improving the location accuracy.

[0077] (3) Embed a channel prior attention module to make the model pay more attention to key fault features, and adopt the K-fold cross-validation method to automatically find the optimal hyperparameters of the model, improving the generalization ability and training efficiency of the model.

[0078] (4) The fault line selection and fault section selection are completed by two parallel CPCA-CNN models. Divide the fault section according to the percentage of the line length, and assign the same section label to the sections with the same percentage on different lines, reducing the number of labels and effectively solving the problem of low discrimination of the model output probability values caused by too many labels. Brief Description of the Drawings

[0079] Figure 1 is the flowchart of the fault location method of the present invention;

[0080] Figure 2 is the schematic diagram of the active distribution network structure;

[0081] Figure 3 is the amplitude-frequency response curve of the Butterworth band-pass filter;

[0082] Figure 4 is the waveform diagram of the zero-sequence current;

[0083] Figure 5 is the time-frequency map of the transient extraction transformation of the zero-sequence current;

[0084] Figure 6 is the structure diagram of the CPCA-CNN model. Detailed Embodiment

[0085] The following further describes in detail the specific embodiments of the present invention in conjunction with the drawings and embodiments, but it does not limit the protection scope of the present invention. Any technical solutions obtained by equivalent replacement or equivalent transformation are within the protection scope of the present invention. The flow of the fault location method of the present invention is as shown in the attached Figure 1 figure.

[0086] Embodiment:

[0087] Take Figure 2 For example, an A-phase ground fault occurs at the 27% position of line L1 from the bus in the active distribution network shown in the appendix. The implementation process of the fault location method is described as follows. The specific implementation steps are as follows.

[0088] 1. Obtain fault simulation data

[0089] According to the appendix Figure 2 Build a simulation model in PSCAD and obtain fault data through offline simulation.

[0090] The set simulation conditions are as follows: The grounding methods are set to three types: ungrounded, grounded through an arc suppression coil, and grounded through a resistor. Among them, the compensation degree of the arc suppression coil is set to full compensation, overcompensation by 5% or overcompensation by 10%; There are a total of L = 6 lines connected to the bus. Among them, L1 is an overhead line connected to a distributed power source, L2 is a cable line, L3 is an overhead line, and L4 - L6 are cable-overhead hybrid lines; The range of the percentage of the distance from the fault point to the bus in the total length of the line is set from 0% to 100%, with an interval of x1% = 9%. Fault points are set at each line according to the above interval, and the total number of fault points is M1 = 72; The transition resistances are taken as R f = 0Ω, 1Ω, 10Ω, 100Ω, 1000Ω, 1500Ω, 2000Ω, a total of M2 = 7 resistance values; Single-phase ground faults of three phases are set respectively; The range of distributed power source penetration rate is from 0% to 100%, with a step of x2% = 20%; The initial fault phase angle is set from 0° to 360°, with a step of x3 = 30°.

[0091] Collect the zero-sequence current at the beginning of each line as fault data. The sampling frequency is f1 = 4000Hz. According to the above simulation conditions, a total of M = 5×72×7×3×6×12 = 544320 groups of fault data can be obtained.

[0092] When saving the fault data, the file naming rule is "*********_****.***", where the first 9 digits are the simulation parameter marks, the 4 digits after the underscore are reserved for the labels of the fault line and the fault section, and the last 3 digits are the file type suffixes.

[0093] The meanings of the simulation parameter marks from left to right are as follows: the first digit represents the neutral grounding method, 0 - represents ungrounded neutral, 1 - represents neutral grounded through an arc suppression coil, 2 - represents neutral grounded through a resistor; the second digit represents the over - compensation degree of the arc suppression coil, 0 - represents full compensation, 1 - represents over - compensation degree of 5%, 2 - represents over - compensation degree of 10%; the third digit represents the resistance value of the transition resistor, 0 - represents 0Ω, 1 - represents 1Ω, 2 - represents 10Ω, 3 - represents 100Ω, 4 - represents 1000Ω, 5 - represents 1500Ω, 6 - represents 2000Ω; the fourth digit represents the distributed power penetration rate, 0 - represents 0%, 1 - represents 20%, 2 - represents 40%, 3 - represents 60%, 4 - represents 80%, 5 - represents 100%; the fifth digit is a reserved bit for extension, not used in this embodiment, and is defaulted to 0; the sixth to ninth digits represent the sample number under the same fault condition, sorted by fault phase, fault location, and fault initial phase angle, with values ranging from 0000 to 9999. For example, the file named "001100100_****.dat" represents the 100th group of fault data obtained under the simulation conditions of ungrounded neutral, transition resistor of 1Ω, and distributed power penetration rate of 20%, and the saved format is "dat".

[0094] 2. Design a band - pass filter

[0095] Select a Butterworth band - pass filter to filter out the power - frequency component and high - frequency interference in the zero - sequence current fault data.

[0096] To filter out the power - frequency component, set the low - frequency cut - off frequency f2 = 200Hz.

[0097] Assume that the lowest frequency of the high - frequency signal available in the zero - sequence current signal is f3 = 4000Hz, then the high - frequency cut - off frequency is f3+Δf h = 4000 + 500 = 4500Hz.

[0098] Set the maximum attenuation of the filter passband to a p = 2dB, and the minimum attenuation of the stopband to a s = 20dB, and calculate the passband ripple coefficient ε1 and the stopband ripple coefficient ε2:

[0099]

[0100] The low - frequency passband edge angular frequency is ω p1 = 2πf2 = 400πrad;

[0101] The high - frequency passband edge angular frequency is ω p2 = 2π(f3+Δf h ) = 9000πrad.

[0102] Set the edge offset:

[0103]

[0104] Then the low - frequency stop - band edge frequency ω s1 and the high - frequency stop - band edge frequency ω s2 are:

[0105]

[0106] Calculate the minimum order N1 corresponding to the low - frequency cut - off frequency of the filter and the minimum order N2 corresponding to the high - frequency cut - off frequency:

[0107]

[0108] In the formula, the symbol represents rounding up. Take the order of the filter as N bp = max{N1, N2}=8.

[0109] The amplitude - frequency response curve of the designed Butterworth band - pass filter is as shown in the appendix Figure 3 as follows.

[0110] 3. Generate time - frequency map samples

[0111] 3.1 Fault data filtering

[0112] The zero - sequence current waveform diagram collected in step 1 is as shown in the appendix Figure 4 as follows. Each group of fault data contains the zero - sequence current sampling value sequences of L = 6 lines. The zero - sequence current sampling value sequences are input into the band - pass filter designed in step 2 in sequence. Filter out the power - frequency components and high - frequency interference. The zero - sequence current data sequence output by the band - pass filter is called the data sample.

[0113] 3.2 Data sample normalization

[0114] The sampling frequency is f1 = 4000Hz, then one power - frequency cycle contains N s = 4000 / 50 = 80 current sampling values; It is set that one data sample contains 4 power - frequency cycles before the fault and 6 power - frequency cycles after the fault. The number of current sampling values of one line calculated is N = 80×(4 + 6)=800.

[0115] The m - th group of data samples can be expressed as:

[0116]

[0117] The data sample after maximum - minimum normalization is I′ m .

[0118] The matrix I′ mThe l-th row in represents the current sampling value sequence of the l-th line after normalization, denoted as I′ l ={i′ l,1 ,i′ l,2 ,...,i′ l,N}.

[0119] 3.3 Convert data samples into transient extraction time-frequency map samples

[0120] Perform transient extraction transformation on I′ m in the data sample I′ l .

[0121] First, obtain the short-time Fourier transform result of I′ l :

[0122]

[0123] where n is the time-domain sampling point, n = 1, 2,..., 800; f is the frequency-domain sampling point, f = 1, 2,..., 800; g(·) is the window function; k1 is the cumulative sequence number of the summation operation.

[0124] Then calculate the transient extraction transformation result of I′ l :

[0125] F TET (n,f)=F STFT (n,f)δ[t - t0(n,f)] (10)

[0126] where δ[t - t0(n,f)] is the transient extraction transformation operation factor, and its value is:

[0127]

[0128] The calculation process of t0(n,f) is:

[0129]

[0130] where Δf is the frequency difference between two adjacent frequency sampling points.

[0131] Plot the above transient extraction transformation result as a two-dimensional time-frequency map, with the abscissa being the current sampling value serial number and the ordinate being the frequency value.

[0132] According to the above steps, sequentially complete the transient extraction transformation of the 6 rows of data in the data sample I′ m , obtain the transient extraction transformation time-frequency maps of 6 lines, and horizontally splice the 6 time-frequency maps in the order of line numbers from left to right in a 2-row and 3-column manner to obtain 1 time-frequency map sample, as shown in the appendix Figure 5 .

[0133] Repeat the above steps to obtain a total of M = 544320 time-frequency diagram samples.

[0134] When saving the time-frequency diagram samples, the file name is "*********_****.jpg", and the saved format is "jpg". The naming rule is the same as the naming rule for the fault data saved in step 1.

[0135] 3.4 Label classification labels for time-frequency diagram samples

[0136] Appendix Figure 2 There are a total of L = 6 faulty lines in the model shown. The line numbers are 1, 2,.., 6 respectively; the faulty sections are divided according to the percentage of the line length. A section is divided every x4% = 10% of the total line length, and a total of S = 10 sections are divided. The faulty sections are numbered 1, 2,.., 10 in the order from near to far from the bus.

[0137] The classification label is represented by the 4 digits "****" after the underscore in the time-frequency diagram sample file name. The first 2 digits of the label represent the faulty line number, with values ranging from 01 to 06; the last 2 digits represent the faulty section number, with values ranging from 01 to 10.

[0138] Example: The time-frequency diagram sample corresponding to the file name "223300050_0106.jpg" has a fault in "the 1st line, the 6th section".

[0139] 3.5 Divide the time-frequency diagram samples into a training set, a validation set, and a test set

[0140] After shuffling all the time-frequency diagram samples, divide them into 2 sets according to the ratio of p1:p2 = 9:1. After division, the first set has 489890 samples, and the second set has 54430 samples.

[0141] Then divide the first set into K T = 10 subsets with the same number of samples for model training and validation. Each subset contains 48989 samples, and they are named J1, J2,..., J 10 .

[0142] The second set is used as the test set.

[0143] 4. Build a line selection model and a section selection model

[0144] Use a convolutional neural network with a fusion channel prior convolutional attention module to construct a line selection model and a section selection model.

[0145] The fault line selection model and the fault section selection model each contain several convolutional layers, CPCA modules, pooling layers, and fully connected layers, with the numbers being n C , n R , nP and n F Both are determined according to the principle of the highest line selection accuracy and section selection accuracy in the validation set.

[0146] As shown in the Figure 6 appendix, the CPCA-CNN model is composed in series in the order of convolutional layer, CPCA module, pooling layer and fully connected layer. The structures of each layer and each module are as follows.

[0147] The convolutional layer completes the convolution operation through the convolution kernel, and its input is the time-frequency map sample S a , and the output is the feature map

[0148] The CPCA module is composed of 1 channel attention module and 1 spatial attention module in series.

[0149] The channel attention module includes 2 parallel pooling layers and 1 shared fully connected layer. Among them, the 2 pooling layers use max pooling and average pooling respectively. After the results of the 2 pooling layers are added, the shared fully connected layer generates the weight vector X of each channel a , X a and are multiplied element by element and then the channel attention feature map F is output ca .

[0150] The spatial attention module includes 3 parallel depthwise separable convolutional layers. The convolutional kernel sizes in the 3 parallel depthwise separable convolutional layers are selected as 7, 11 and 21 respectively. The input of the spatial attention module is F ca , and the output is F sa .

[0151] F ca and F sa are multiplied element by element, and the CPCA attention feature map F cpca is output.

[0152] The pooling layer of the CPCA-CNN model uses max pooling, and its input is F cpca , and the output is the feature map after dimensionality reduction

[0153] In the fully connected layer of the CPCA-CNN model, the activation function is the Softmax function, and its input is For the fault line selection model, its output is the fault probability of each line in the fault data to be recognized, and the line corresponding to the maximum probability is the fault line; for the fault section selection model, its output is the fault probability of each section in the fault data to be recognized, and the section corresponding to the maximum probability is the fault section.

[0154] 5. Training and testing the line selection model and the section selection model

[0155] Use the time-frequency map samples in the training set obtained in step 3 to train the line selection model and the segment selection model respectively.

[0156] Set the value range of hyperparameters according to experience: n C = n R = n P The value is 1, 2, 3, 4, 5, n F The value is 1, 2, the Batch size value is 4, 8, 12, 16, 20, and the Leaning rate value is 0.1, 0.01, 0.001, 0.0001. Enumerate all combinations of hyperparameters, and a total of 5×2×5×4 = 200 different hyperparameter value schemes can be obtained.

[0157] The specific process of hyperparameter optimization for the fault line selection model is as follows:

[0158] Take 1 set of hyperparameter value combinations. First, use the sample subset J1 as the validation set, and use the sample subsets J2, J3,..., J 10 as the training set to obtain the first line selection accuracy rate (47700 / 48989)×100% = 97.38%; then use the sample subset J2 as the validation set, and use the sample subsets J1, J3,..., J 10 as the training set to obtain the second line selection accuracy rate (47847 / 48989)×100% = 97.67%. And so on, obtain K T = 10 line selection accuracy rates, take the average of the obtained line selection accuracy rates, and obtain the average accuracy rate of this hyperparameter value combination as 98.59%.

[0159] Repeat the above process for all hyperparameter value combinations, and take the hyperparameter value combination corresponding to the highest average accuracy rate as the optimal hyperparameter. It can be obtained that the highest average accuracy rate of the fault line selection model is 99.50%, and the corresponding optimal hyperparameter value combination is n R = n C = n P = 3, n F = 1, Batch size = 16, and Leaning rate = 0.0001.

[0160] The hyperparameter optimization steps for the fault segment selection model are the same as those for the fault line selection model. Through the same steps as above, it can be obtained that the highest average accuracy rate of the fault segment selection model is 98.17%, and the corresponding optimal hyperparameter combination is n R = n C = n P = 3, n F = 2, Batch size = 20, and Leaning rate = 0.0001.

[0161] After the training of both the fault line selection and fault section selection models is completed, the accuracy rates of the fault line selection and fault section selection models are respectively tested using 54,430 time-frequency diagrams in the test set.

[0162] 6. Use the trained model to complete fault location

[0163] For the data of the A-phase ground fault occurring at the 27% position from the bus on L1 in this embodiment, the preprocessing methods of steps 3.1, 3.2, and 3.3 are adopted to be transformed into the time-frequency diagram S' to be recognized. a 。

[0164] First, the time-frequency diagram S' to be recognized a is input into the trained fault line selection model, and the processing procedures of each layer and each module in the line selection model are as follows.

[0165] Input S' a into the convolutional layer. After the convolution operation of the convolution kernel, the feature map S' is obtained a [1]

[0166] The output of the convolutional layer is input to the channel attention module for processing. First Pass through 2 pooling layers, and respectively output the maximum pooling feature vector and the average pooling feature vector; the 2 feature vectors are subjected to feature extraction through the shared fully connected layer to obtain 2 feature vectors. After adding these 2 feature vectors element by element and passing through the Sigmoid function, the weight vector X of each channel is obtained a 。X a is multiplied element by element with to generate the channel attention feature map F ca 。

[0167] Then input F ca into the spatial attention module. First, pass through 1 convolutional layer to obtain the intermediate feature map F m , input F m into 3 parallel depthwise separable convolutional layers to respectively obtain feature maps of different scales Add the feature maps F m 、 element by element, and then pass through 1 convolutional layer to finally output the spatial attention feature map F sa 。

[0168] The channel attention feature map F ca is multiplied element by element with the spatial attention feature map F sa to generate the CPCA attention feature map F cpca 。

[0169] Input F cpcaInput into the pooling layer, and after max-pooling processing, output the feature map

[0170] Feature map After passing through the fully connected layer, first obtain the value a of each neuron g , the number of neurons is the same as the number of transmission lines, and the neuron serial number takes values g = 1, 2, 3, 4, 5, 6, and the results are a1 = 9.1, a2 = 2.6, a3 = 4.0, a4 = 1.9, a5 = 2.3, a6 = 3.7. Then use a g And calculate the fault probability value vector of each line in the data of the fault to be identified through the Softmax function:

[0171] X = [0.9885, 0.0021, 0.0032, 0.0015, 0.0018, 0.0029] (13)

[0172] Among them, the line L1 with the highest probability is the selected fault line.

[0173] Then input S' a Into the trained fault section selection model, the process is the same as that of the fault line selection model. The fault section selection model outputs the probability value vector of each section where a fault occurs:

[0174] Y = [0.0078, 0.0089, 0.9787, 0.0013, 0.0009, 0.0005, 0.0012, 0.0001, 0.0002, 0.0004] (14)

[0175] Among them, the section S3 with the highest probability is the selected fault section.

[0176] Combining the fault line selection result and the fault section selection result, it is concluded that a single-phase ground fault has occurred in the S3 section of line L1.

[0177] As shown above, only the preferred embodiments of the present invention are shown, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for locating faults in active distribution networks based on an improved convolutional neural network, characterized in that: The following steps are involved: Step 1: Obtain fault simulation data An active distribution network model was built using simulation software, and simulation experiments were carried out under different fault conditions. The zero-sequence current at the beginning of each line was collected as fault data, with a sampling frequency of f1Hz. Step 2: Design the Bandpass Filter The Butterworth bandpass filter is used to filter out the power frequency component and high-frequency interference in the zero-sequence current fault data. The order of the filter is calculated according to the cut-off frequency, and the coefficient of the digital filter is designed. Step 3: Generate time-frequency graph samples The collected fault data is bandpass filtered to form a data sample I including the zero-sequence current of all L lines. m , and then normalized to obtain data sample I′ m , where the lth row vector I′ l Corresponding to the zero-sequence current of the lth line; for I′ l Implement time-frequency transformation to transform the one-dimensional zero-sequence current data into a two-dimensional time-frequency diagram; follow the above steps to complete the data sample I′ in sequence m Perform the time-frequency transformation of L rows of data to obtain L time-frequency graphs, and then splice these L time-frequency graphs into one time-frequency graph sample; Label the time-frequency graph samples with classification labels; repeat the above process to obtain the time-frequency graph samples corresponding to all fault data; divide them into two sets according to the ratio of p1:p2, the former is used to form the training set and the validation set, and the latter is used as the test set; Step 4: Build line selection model and segment selection model The channel prior convolutional attention module, referred to as CPCA-CNN model, is embedded in the convolutional neural network, and the fault line selection model and fault section selection model based on CPCA-CNN are built; Step 5: Train and test the line selection model and segment selection model According to experience, set the reasonable value range of the model hyperparameters, and exhaustively enumerate all combinations of hyperparameters. Under different hyperparameter value combinations, use the first set divided in step 3 to train the fault line selection model and the fault section selection model respectively. According to the highest accuracy criterion, determine the optimal hyperparameters, and use the second set to test the model. Step 6: Use the trained model to complete fault location The newly collected fault data is converted into time-frequency graph samples to be identified, and input into the fault line selection model, which outputs the probability value of each line being a faulty line, and the line corresponding to the maximum probability value is the faulty line; the time-frequency graph samples to be identified are then input into the fault section selection model, which outputs the probability value of each section being a faulty section, and the section corresponding to the maximum probability value is the faulty section; the fault location is determined by combining the line selection results and fault results.

2. The active distribution network fault location method based on improved convolutional neural network according to claim 1 is characterized in that: The method for determining the filter order in step 2 is: The ideal order N1 of the filter is calculated based on the low-frequency cutoff frequency, the low-frequency passband edge angular frequency and the low-frequency edge bias. The ideal order N2 of the filter is calculated based on the high-frequency cutoff frequency, the high-frequency passband edge angular frequency and the high-frequency edge bias. The larger value of N1 and N2 is taken as the order of the bandpass filter.

3. The active distribution network fault location method based on improved convolutional neural network according to claim 1 is characterized in that: The classification labels marked on the time-frequency graph samples in step 3 are combined with the fault conditions in step 1 and reflected in the time-frequency graph sample file name at the same time, separated by underscores. The fault conditions are set before the underscore, and the classification labels are after the underscore, including the fault line number and the fault section number.

4. The active distribution network fault location method based on improved convolutional neural network according to claim 1 is characterized in that: The classification labels marked in step 3 include fault section labels. The fault section is divided as follows: the sections are divided according to the percentage of the length of each line. Each line is divided into S sections. Sections with the same percentage on different lines are assigned the same section label. The section labels are numbered in order from near to far from the busbar, 1, 2, .., S.

5. The active distribution network fault location method based on improved convolutional neural network according to claim 1, characterized in that: In step 3, the one-dimensional zero-sequence current data is transformed into a two-dimensional time-frequency diagram by using transient extraction transformation.

6. The active distribution network fault location method based on improved convolutional neural network according to claim 1, characterized in that: The CPCA-CNN model built in step 4 consists of convolutional layers, CPCA modules, pooling layers and fully connected layers connected in series in sequence; among them, the CPCA module is composed of 1 channel attention module and 1 spatial attention module in series, the channel attention module is composed of 2 parallel pooling layers and 1 shared fully connected layer, and the spatial attention module contains 3 parallel depth-separable convolutional layers; the output of the CPCA module is obtained by element-by-element multiplication of the output of the channel attention module and the output of the spatial attention module.

7. The active distribution network fault location method based on improved convolutional neural network according to claim 1, characterized in that: In step 5, the K-fold cross-validation method is used to determine the optimal hyperparameters of the CPCA-CNN model. The sample set with a proportion of p1 in step 3 is divided into K T When one of the subsets with the same number of samples is selected as the validation set, the remaining (K T -1) subsets are used as training sets.

8. An active distribution network fault location system based on improved convolutional neural network, characterized in that: The system is implemented by the active distribution network fault location method based on the improved convolutional neural network according to any one of claims 1 to 7, and the system comprises: Data acquisition module: configured at the beginning of each line, used to collect zero-sequence current fault data of the distribution network line; Data preprocessing module: configured in the central processing device of the fault location system, used to filter, normalize, transform time and frequency, label and divide the zero-sequence current fault data into sample sets; Fault location model training module: It is configured in the central processing device of the fault location system and is used to build a CPCA-CNN-based fault line selection model and a fault section selection model, and train the model using time-frequency graph samples and the K-fold cross-validation method; Fault data identification module: It is configured in the central processing device of the fault location system. It inputs the on-site fault data into the trained fault line selection model and fault section selection model after the data preprocessing module, outputs the fault line and fault section results, and determines the fault location by combining the two results.

Images