Power distribution network abnormal event identification method, device, equipment and medium

Abstract

The application belongs to the technical field of electric power automation, and discloses a power distribution network abnormal event identification method, device, equipment and medium; the method comprises the following steps: continuously moving a window to online intercept power distribution network voltage data; inputting the intercepted power distribution network voltage data into a pre-trained reversible neural network model to calculate an output log-likelihood value J t ; comparing the log-likelihood value J t with an abnormal event detection threshold value J th to obtain a power distribution network abnormal event identification result; and outputting the power distribution network abnormal event identification result. The application utilizes the fitting capability of a reversible neural network for complex distribution and the reversible characteristics to realize online monitoring, positioning and risk level evaluation of abnormal events.

Description

Technical Field

[0001] This invention belongs to the field of power automation technology, and specifically relates to a method, device, equipment and medium for identifying abnormal events in power distribution networks. Background Technology

[0002] In recent years, with the integration of new power electronic devices, intermittent renewable energy sources, and electric vehicles, the structure and operation of distribution networks have become increasingly complex. Simultaneously, due to the low density and poor observability of distribution network measurements, the difficulty of analysis and operation maintenance has gradually increased. In distribution networks, real-time identification and risk level assessment of faults (such as short circuits and line breaks) and other abnormal events (such as heavy load switching) are crucial for improving fault event retrospective analysis capabilities and supporting system operation status monitoring and stability assessment. Currently, traditional mechanism-driven methods for identifying abnormal events in distribution networks often rely on varying degrees of assumptions and simplifications in event modeling, thus only identifying specific types of events and having limitations in applicable scenarios. On the other hand, the gradual improvement of distribution network measurement technology and the rapid development of artificial intelligence technology have made it possible to develop machine learning-based abnormal event identification methods within a data-driven framework, offering broader application prospects in improving algorithm universality and robustness.

[0003] The key to detecting abnormal events in distribution networks lies in pre-processing feature mining and effective modeling of measurement data (i.e., samples) under normal operating conditions. This enables early monitoring, warning, and risk level assessment of abnormal events during online application, and allows for evaluation of the effectiveness of existing response strategies through retrospective analysis of events. Due to the diversity and uncertainty of distribution network operation modes, its measurement data under normal operating conditions exhibits complex and difficult-to-simulate nonlinear distribution characteristics. Generative models are an effective method in machine learning for learning complex sample distributions. However, classic generative models (such as generative adversarial networks and variational autoencoders) cannot evaluate new samples by learning the complex distribution of normal-state data, and therefore cannot achieve online monitoring. Summary of the Invention

[0004] The purpose of this invention is to provide a method, device, equipment, and medium for identifying abnormal events in power distribution networks, thereby addressing the technical problem that existing technologies cannot evaluate new samples by learning the complex distribution of normal state data, and thus cannot achieve online monitoring. On one hand, this invention utilizes the fitting ability of reversible neural networks to complex distributions and their reversibility to achieve online monitoring, location, and risk level assessment of abnormal events. On the other hand, this invention combines the location of the abnormal event with appropriate measurement values ​​as input features, and uses pseudo-labeling technology to classify abnormal events with low labeling rates.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a method for identifying abnormal events in a power distribution network, comprising:

[0007] The distribution network voltage data is captured online using a continuously moving window.

[0008] The captured distribution network voltage data is input into a pre-trained reversible neural network model, and the log-likelihood value J is calculated. t ;

[0009] The log-likelihood value J t With the abnormal event detection threshold J th By comparison, the results of abnormal event identification in the power distribution network are obtained;

[0010] Output the results of abnormal events in the power distribution network.

[0011] In a second aspect, the present invention provides a power distribution network abnormal event identification device, comprising:

[0012] The data acquisition module is used to capture distribution network voltage data online using a continuously moving window.

[0013] The calculation module is used to input the captured distribution network voltage data into a pre-trained reversible neural network model and calculate the output log-likelihood value J. t ;

[0014] The judgment module is used to determine the log-likelihood value J. t With the abnormal event detection threshold J th By comparison, the results of abnormal event identification in the power distribution network are obtained;

[0015] The output module is used to output the identification results of abnormal events in the power distribution network.

[0016] Thirdly, the present invention provides an electronic device including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the above-described method for identifying abnormal events in a power distribution network.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the power distribution network abnormal event identification method.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] This invention provides a method, apparatus, device, and medium for identifying abnormal events in a power distribution network. The method includes: capturing power distribution network voltage data online using a continuously moving window; inputting the captured power distribution network voltage data into a pre-trained reversible neural network model; and calculating and outputting a log-likelihood value J. t; the log-likelihood value J t With the abnormal event detection threshold J th By comparison, the identification results of abnormal events in the power distribution network are obtained; the identification results of abnormal events in the power distribution network are output. This invention utilizes the fitting ability of reversible neural networks to complex distributions of measurement data. By training the reversible neural network to learn the characteristics of normal operating state data, abnormal events can be detected, located, and risk levels determined based on changes in likelihood values ​​in online applications.

[0020] Furthermore, this invention trains a convolutional neural network with a low labeling rate based on pseudo-labeling technology and achieves abnormal event type identification, which has significant engineering application value and promotion prospects. Attached Figure Description

[0021] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0022] Figure 1 This is a flowchart of the power distribution network abnormal event identification method provided in an embodiment of the present invention.

[0023] Figure 2 This is an architectural diagram of the reversible neural network in an embodiment of the present invention.

[0024] Figure 3 This is a graph showing the change in the training loss function of the reversible neural network for anomaly event detection in an IEEE 34-node system according to an embodiment of the present invention.

[0025] Figure 4 This is an architecture diagram of a convolutional neural network in an embodiment of the present invention.

[0026] Figure 5 This is a schematic diagram of the process of training a convolutional neural network based on pseudo-label technology in an embodiment of the present invention.

[0027] Figure 6 This is a schematic diagram of the likelihood value distribution for abnormal event detection in an IEEE 34-node system according to an embodiment of the present invention.

[0028] Figure 7 This is a schematic diagram of the Jacobian matrix for locating abnormal events in an IEEE 34-node system according to an embodiment of the present invention.

[0029] Figure 8 This is a graph showing the change in the training loss function of the convolutional neural network for classifying abnormal events in an IEEE 34-node system, as described in this embodiment of the invention.

[0030] Figure 9 This is a graph showing the operational characteristics of receivers in anomaly event classification for the IEEE 34-node system under different proportions of labeled samples in an embodiment of the present invention.

[0031] Figure 10 This is a flowchart illustrating the distribution network abnormal event identification method in another embodiment of the present invention;

[0032] Figure 11 This is a structural block diagram of a power distribution network abnormal event identification device according to the present invention;

[0033] Figure 12 A structural block diagram of an electronic device according to the present invention. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0035] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0036] Reversible neural networks, due to their reversibility, can provide explicit likelihood values ​​for new samples, thereby measuring the probability that a new sample belongs to a normal data distribution and detecting anomalous events. Furthermore, for the problem of anomalous event type identification, on the one hand, supervised learning methods commonly used in machine learning for classification are difficult to apply due to the lack of a sufficient number of labeled event samples in power distribution networks; on the other hand, unsupervised learning can only make relatively coarse classifications of event types, failing to meet practical engineering needs. Conversely, semi-supervised classification methods based on pseudo-labeling technology simultaneously incorporate both labeled and unlabeled samples into training. Specifically, it assigns pseudo-labels to unlabeled samples and selects samples with high-accuracy pseudo-labels to participate in iterative training, thereby improving the labeling rate of the training set and ultimately achieving refined classification of anomalous events even with a low labeling rate.

[0037] Example 1

[0038] This invention provides a method for identifying abnormal events in a power distribution network, comprising the following steps:

[0039] S1. Obtain normal operating state data samples required for training the reversible neural network model through offline time-domain simulation;

[0040] Step S1 includes the following specific steps:

[0041] S11. Construct the topology of the power distribution network simulation system based on PSCAD software, and use Python to write an interface to call the simulation software, thereby controlling batch simulation under different operating scenarios;

[0042] S12. Set the simulation step size and duration parameters, obtain the three-phase voltage data of each node of the distribution network under normal operating conditions as training samples, and add Gaussian noise to simulate measurement errors.

[0043] S13. Normalize the sample data using the 0-1 standardization method:

[0044]

[0045] In the formula, M i,min and M i,max Sample M i The minimum and maximum values.

[0046] S2. Construct a reversible neural network model for detecting, locating, and assessing the risk level of abnormal events in the power distribution network and train it offline.

[0047] Step S2 includes the following specific steps:

[0048] S21. Construct a reversible neural network model for detecting abnormal events in the power distribution network. The reversible neural network model consists of multiple layers. The input sample data is processed through a compression layer and several streaming modules, and then outputs its log-likelihood value relative to a standard Gaussian distribution. The streaming module consists of an activation normalization layer, a reversible 1×1 convolutional layer, and an affine coupling layer. The affine coupling layer further includes a splitting layer, a shallow convolutional neural network, and an aggregation layer.

[0049] S22. Using the normalized sample data described in S13 As input, the inverse of the log-likelihood value described in S21 is used as the loss function to train the invertible neural network model. The parameters of the invertible neural network model are randomly initialized and iteratively updated using mini-batch gradient descent until the network converges, thus obtaining the trained invertible neural network model.

[0050] S23. Input the normalized sample data described in S13, and denote the minimum log-likelihood output of the trained invertible neural network model as J. min Take J th =λJ min The threshold for detecting abnormal events is λ = 1.2.

[0051] S3. Voltage data is captured online using a continuously moving window as input to a reversible neural network model. The output (i.e., log-likelihood value) is calculated to detect abnormal events, and the risk level of abnormal events is assessed based on the degree of deviation of the likelihood value under a t-test.

[0052] Step S3 includes the following specific steps:

[0053] S31. Use a continuously moving window of size N×K to capture the three-phase voltage data of all nodes as the input observation matrix of the reversible neural network model. Let the observation matrix acquired at time k be... Where column vector It contains the voltage variables of N measurement channels observed at time k. The observation matrix at each time moment consists of the observation vector at the current time moment and K-1 historical observation vectors.

[0054] S32. Normalize the observation matrix using the 0-1 normalization method:

[0055]

[0056] In the formula, M k,min and M k,max Sample M k The minimum and maximum values.

[0057] S33. The standardized observation matrix Input the trained reversible neural network model and calculate the corresponding output, i.e., the log-likelihood value J. t and the threshold J of S23 th Comparison: When J t <J th If an abnormal event is determined to have occurred, proceed to step S34 to assess the risk level of the abnormal event; when J t ≥J th The current state is determined to be normal; proceed to S31 to analyze the subsequent observation matrix M. k+1 .

[0058] S34. Take K' consecutive observation matrices before and after the time k when the anomalous event occurs, calculate the corresponding log-likelihood values ​​and normalize them:

[0059]

[0060] In the formula, μ(J) k ) and σ(J k ) are the mean and standard deviation of the K' log-likelihood values ​​before and after time k, respectively.

[0061] S35. Use the t-test to measure the log-likelihood value after an abnormal event occurs. The degree to which it deviates from its normal level. Under normal operating conditions, It follows a t-distribution with K'-1 degrees of freedom. After the abnormal event occurs, It no longer follows a t-distribution, therefore a one-tailed t-test is performed on it:

[0062]

[0063] In the formula, after normalization by S34, 1-α is the confidence level, t α Let P' be the lower α quantile of the t-distribution with K'-1 degrees of freedom, P{·} be the probability operator, and μ be the probability value in the t-test. The overall mean. Therefore, the confidence interval at confidence level 1-α is...

[0064] S36. Construct the confidence level 1-α corresponding to different risk levels. Based on the t-distribution table, find... The corresponding p-value is compared with α to assess the risk level of the abnormal event.

[0065] S4. Calculate the Jacobian matrix of the output value versus the input value of the invertible neural network model to locate abnormal events;

[0066] Step S4 includes the following specific steps:

[0067] S41. Observation matrix of abnormal events detected in step S33 Calculate the output value J of the reversible neural network model k For the input matrix The Jacobian matrix, i.e.

[0068]

[0069] In the formula, J k It is the log-likelihood value (scalar). It is the observation matrix, the Jacobian matrix R. k Each element r i,t The corresponding element m in the input matrix was measured. i,t For low likelihood values ​​J k The "contribution" is |r i,t The larger |m i,t For J k The greater the contribution.

[0070] S42, Obtain J k m, the one with the largest contribution i,t This allows us to pinpoint abnormal events.

[0071]

[0072] In the formula, β and δ represent the elements m in the input matrix, respectively. i,t The spatial location and time information can be used to locate the βth measurement channel.

[0073] S5. Perform offline simulations of various abnormal events under different operating scenarios, obtain measurement data of voltage, line current, and differential current at both ends of the disturbed line, and construct the data samples required for training the convolutional neural network based on pseudo-label technology.

[0074] Step S5 includes the following specific steps:

[0075] S51. A distribution network simulation system topology and event module are constructed based on PSCAD software, and an interface is written in Python to call the simulation software, thereby controlling batch simulations under different operating scenarios, abnormal event types, and parameter conditions. Abnormal event types include three-phase grounding, two-phase grounding, heavy load switching, and line tripping.

[0076] S52. Obtain the three-phase current amplitude (I) of the line where the abnormal event occurs. a ,I b ,I c Differential current phasor Voltage amplitude at both ends of the line (U) a1 U b1 U c1 U a2 U b2 U c2 ) and the phasor difference of the voltages at both ends of the line (referred to as differential voltage phasor in this invention, denoted as ΔU) a , ΔU b , ΔU c , ).

[0077] S53. For the abnormal event samples generated in each simulation, the measurement data obtained in S52 are concatenated into a feature matrix D, i.e., I. a ,I b ,I c , U a1 ,

[0078] U b1 U c1 U a2 U b2 U c2 ,

[0079] S6. Construct a convolutional neural network based on pseudo-label technology for the classification of abnormal events in the power distribution network, and perform offline training under different proportions of labeled samples;

[0080] Step S6 includes the following specific steps:

[0081] S6. Construct a convolutional neural network model for classifying abnormal events in the power distribution network. The convolutional neural network model consists of multiple layers, including convolutional layers, pooling layers, and fully connected layers. The input of the neural network is the feature matrix D described in S53, the output is the abnormal event category, and the loss function is cross-entropy.

[0082] S62. Divide the abnormal event samples obtained in step S5 into two parts: labeled samples and unlabeled samples. Labeled samples are denoted as L = {(x...} h ,y h ):h∈(1,…,H)}, unlabeled samples are denoted as Q={q h :h∈(1,…,ρH)}, where x h y h Let q represent the feature matrix of the labeled sample and its corresponding anomaly event category, respectively. h The feature matrix represents the unlabeled samples, where H, ρH, and ρ represent the number of labeled samples, the number of unlabeled samples, and the relative ratio of the number of labeled samples to the number of unlabeled samples, respectively.

[0083] S63. Input the feature matrices of labeled and unlabeled samples into the convolutional neural network model described in S61, respectively. For the labeled sample portion, calculate its cross-entropy loss function L1:

[0084]

[0085] In the formula c is the number of categories, p(wx) h Let p be the predicted probability vector, where each element p k (wx h Characterize sample x h The probability assigned to category k. It is a one-hot encoding of category k. Similarly, for unlabeled samples, pseudo-labels u are assigned based on the prediction results of the convolutional neural network. h And select those with a confidence level higher than the confidence threshold. Calculate the cross-entropy loss function L2 for pseudo-labeled samples:

[0086]

[0087] In the formula, This is the confidence threshold for pseudo-labeled samples. The loss function described in S61 is calculated as follows:

[0088] L=L1+ζL2

[0089] In the formula, L1 is the labeled loss function, L2 is the pseudo-label loss function, and ζ is the balance coefficient that controls the proportion of the pseudo-label loss function.

[0090] S7. Based on the abnormal event location information obtained in step S4, obtain measurement data such as voltage at both ends of the line where the abnormal event is located, line current, and differential current, and use a convolutional neural network constructed based on pseudo-label technology to identify the type of abnormal event.

[0091] Step S7 includes the following specific steps:

[0092] S71. Based on the node location information obtained in step S42, locate the line with the largest change in current amplitude associated with that node (i.e., the disturbed line). As described in step S53, obtain the measurement data of the voltage at both ends of the line and the line current to form a feature matrix D.

[0093] S72. Input the feature matrix D described in step S71 into the convolutional neural network model trained in step S6 to identify the type of abnormal event.

[0094] Example 2

[0095] like Figure 1 As shown, the present invention provides a method for identifying abnormal events in a power distribution network, which specifically includes the following steps:

[0096] S101, time-domain simulation generation of training data for a reversible neural network, using the IEEE 34-node system as an example, with photovoltaic power connected at node 814 and doubly-fed induction generators connected at nodes 856 and 890. The PSCAD simulation step size is set to 50 μs, and phasors are calculated once per cycle in a 50 Hz system. The simulation duration for each sample is set to 1 second. Assuming 17 measurement devices are deployed in the system to observe the three-phase voltage amplitude at each point, each training sample is 51 × 50 pixels. A total of 10,000 training samples under normal operating conditions are generated under different operating modes, with 50 dB of Gaussian noise added to simulate measurement errors.

[0097] S102, Training the invertible neural network model, specifically includes:

[0098] 1) Standardize the training samples generated by time-domain simulation by 0-1.

[0099] 2) Fill the sample of size 51×50 generated by S101 with 0 to form a sample of size 52×50.

[0100] 3) Construct a reversible neural network model for anomaly detection and localization. The input sample passes through a compression layer S1, a flow module Flow1, and a flow module Flow2, and the output is the log-likelihood value of the sample, such as... Figure 2As shown in the diagram, the Flow module includes an activation normalization layer AN1, a reversible 1×1 convolutional layer IC1, and an affine coupling layer AC1. The affine coupling layer AC1 further includes a splitting layer SP1, a shallow convolutional neural network SCN1, and an aggregation layer CT1. The input feature dimension of the compression layer S1 is 52×50×1, and the output feature dimension is 26×25×4. The output feature dimensions of the activation normalization layer AN1, the reversible 1×1 convolutional layer IC1, and the affine coupling layer AC1 are all 26×25×4. The structure of the Flow module Flow2 is the same as that of the Flow module Flow1, with both its input and output feature dimensions being 26×25×4. The output of the Flow module Flow2 is substituted into the probability density function of a standard Gaussian distribution to obtain the log-likelihood value of the output. The activation function for all layers of the reversible neural network is the ReLU function.

[0101] 4) The negative of the log-likelihood function is used as the loss function for the constructed invertible neural network. The Adam algorithm is employed for optimization, ultimately yielding a model that best simulates the complex distribution of data under normal conditions. The curve showing the change in the loss function of the invertible neural network during training is shown below. Figure 3 As shown.

[0102] S103, Temporal simulation generation of convolutional neural network training data, specifically including:

[0103] 1) At different locations within the IEEE 34 nodes, event samples of various types were generated under different parameter conditions. Event types included three-phase grounding, two-phase grounding, heavy load connection, and line tripping. Three-phase grounding was randomly set at lines 816–824, 828–830, 832–858, and 812–814; the fault clearing time was 0.1 seconds, the grounding resistance was within 50 ohms, and the fault location was randomly selected within 10% to 90% of the line length. Two-phase grounding was randomly set at lines 816–824, 852–854, 888–890, and 832–858; the fault clearing time was 0.1 seconds, the grounding resistance was within 50 ohms, and the fault location was randomly selected within 10% to 90% of the line length. Heavy load connection was randomly set at nodes 844, 888, 860, and 810; the load size accounted for 10% to 30% of the total network load. Line tripping was randomly set on lines 828–830, 816–824, 832–858, and 888–890; the location was randomly selected within 10% to 90% of the line length. 1000 samples of each event type were generated. The PSCAD simulation step size was set to 50 μs, and the phasor was calculated once per cycle in a 50 Hz system. The simulation duration for each sample was set to 1 second.

[0104] 2) For each event sample, obtain the three-phase current amplitude, differential current, voltage amplitude at both ends of the line, and differential voltage of the line where the abnormal event occurs. The dimension of each sample is 21×50.

[0105] S104. Training the convolutional neural network model, specifically including:

[0106] 1) Partially label the abnormal event samples generated by S103, with the labeling ratios set to 10% and 1% respectively.

[0107] 2) Construct a convolutional neural network model for anomaly event classification. The structure of the convolutional neural network is as follows: Figure 4 As shown, the input sample passes through convolutional layer CN1, pooling layer PL1, convolutional layer CN2, pooling layer PL2, and fully connected layer FC1 to output sample category information. The activation function is the softmax function.

[0108] 3) Train the convolutional neural network based on pseudo-labeling technology. The training process is as follows: Figure 5 As shown, labeled and unlabeled samples are input into the constructed convolutional neural network, and the corresponding prediction results are output. Given a confidence threshold of 0.8 for pseudo-labeled samples and a balance coefficient of 10 controlling the proportion of the pseudo-label loss function, the convolutional neural network loss function is calculated. The Adam algorithm is used for optimization, resulting in a model that can effectively identify abnormal events. The curve of the convolutional neural network loss function during training is shown in the figure. Figure 6 As shown.

[0109] S105. Time-domain simulation generation of test set data, specifically including:

[0110] 1) As described in S101, a total of 2000 test samples under normal conditions were generated under different operating modes of the IEEE 34 nodes. The simulation time for a single sample was 3 seconds, the size was 51×150, and 50dB of Gaussian noise was added to simulate measurement error.

[0111] 2) As described in S103, abnormal event samples are generated at different locations of the IEEE 34 nodes under different parameter conditions. Event types include three-phase grounding, two-phase grounding, heavy load connection, and line tripping. 400 samples are generated for each event type, with a simulation time of 3 seconds per sample and a size of 51×150, for a total of 1600 abnormal event samples.

[0112] 3) Shuffle the samples generated in 1) and 2) to generate a test set, which contains 3600 samples.

[0113] S106. Continuously move the window to form the observation matrix, that is, use the observation matrix of size 51×50 to extract the 3600 samples of size 51×150 generated in S105.

[0114] S107. Calculate the log-likelihood value of the observation matrix, that is, input the observation matrix truncated in S106 into the invertible neural network trained in S102 to obtain the log-likelihood value of the observation matrix. Figure 7 The log-likelihood values ​​are the observation matrix under normal conditions and the observation matrix containing abnormal events in the test set samples. It can be seen that the log-likelihood value can effectively distinguish between normal samples and abnormal samples.

[0115] S108. Determine whether the log-likelihood value is lower than the threshold, specifically including:

[0116] 1) As described in S102, obtain the minimum value of the log-likelihood value calculated from the training set samples of the reversible neural network and multiply it by the coefficient 1.2 to obtain the abnormal event detection threshold.

[0117] 2) As described in S107, compare the log-likelihood value of the observation matrix with the anomaly detection threshold described in 1). If the log-likelihood value of the observation matrix is ​​lower than the threshold, an anomaly is determined to have occurred, and proceed to S109; otherwise, continue analyzing the next observation matrix.

[0118] S109. An abnormal event is detected and a risk assessment is conducted, specifically including:

[0119] 1) In the abnormal event sample, take the first observation matrix that detects the occurrence of the abnormal event, and normalize the log likelihood values ​​of the 100 observation matrices before and after it by equation (3).

[0120] 2) Construct confidence levels corresponding to different risk levels of abnormal events: greater than 90% and less than or equal to 96% is the precautionary state; greater than 96% and less than or equal to 99% is the high-risk state; and greater than 99% is the emergency state.

[0121] 3) Perform a one-sided t-test on the observation matrix described in 1), look up the p-value corresponding to the observation matrix in the table, and compare it with the event risk level described in 2) to determine the risk level of the detected abnormal event.

[0122] S110. Locating abnormal events based on the Jacobian matrix, specifically including:

[0123] 1) For the observation matrix of the detected abnormal event described in S109, calculate its log-likelihood value as the Jacobian matrix of the observation matrix.

[0124] 2) Obtain the row containing the element with the largest absolute value in the Jacobian matrix, thereby locating the node associated with the abnormal event, and then identifying the line with the largest change in current measurement among the lines associated with that node. A schematic diagram of the Jacobian matrix is ​​shown below. Figure 8 As shown.

[0125] S111. Obtain measurement information of the line where the abnormal event is located, that is, according to the line located in S110, obtain the three-phase current amplitude, differential current, voltage amplitude at both ends of the line and differential voltage of the line where the abnormal event is located.

[0126] S112. Identify the type of abnormal event, that is, input the abnormal event measurement information described in S111 into the convolutional neural network trained under different sample labeling ratios as described in S104, and then identify the type of abnormal event based on the output of the neural network. The Received Operational Characteristic (ROC) curves of the classifier at sample labeling ratios of 10% and 1% are as follows: Figure 9 (a) and Figure 9 As shown in (b), the area under the ROC curve (AUC) measures the classifier's performance. An area value within the range [0,1], with values ​​closer to 1 indicating better performance; that is, the further the ROC curve is from the diagonal, the better the classification performance. Figure 9 As shown, even with a low percentage of labeled samples, the convolutional neural network constructed in this paper can effectively distinguish event types.

[0127] Example 3

[0128] Please see Figure 10 The present invention provides a method for identifying abnormal events in a power distribution network, comprising:

[0129] S10. Use a continuous moving window to capture distribution network voltage data online;

[0130] S11. Input the captured distribution network voltage data into the pre-trained reversible neural network model, and calculate the output log-likelihood value J. t ;

[0131] S12, the log-likelihood value J t With the abnormal event detection threshold J th By comparison, the results of abnormal event identification in the power distribution network are obtained;

[0132] S13, Output the results of abnormal events in the power distribution network.

[0133] In one specific implementation: the pre-trained reversible neural network model is obtained through the following steps:

[0134] Normal operating state data samples for training a reversible neural network model were obtained through offline time-domain simulation.

[0135] The reversible neural network model is trained offline using normal operating state data samples to obtain a pre-trained reversible neural network model.

[0136] The reversible neural network model includes a compression layer and several streaming modules; the input training normal operation state data samples are processed through the compression layer and several streaming modules, and the output is the log-likelihood value relative to the standard Gaussian distribution.

[0137] The reversible neural network model is trained offline using normalized sample data. As input, the inverse neural network model is trained using the negative of the log-likelihood value as the loss function; the parameters of the inverse neural network model are randomly initialized and iteratively updated using mini-batch gradient descent until the network converges, thus obtaining the trained inverse neural network model.

[0138] Input normalized sample data Let J be the minimum log-likelihood value output by the trained invertible neural network model. min Take J th =λJ min λ is a coefficient used as the threshold for detecting abnormal events.

[0139] In one specific implementation: the step of using a continuously moving window to capture distribution network voltage data online specifically includes:

[0140] A continuously moving window of size N×K is used to extract the three-phase voltage data of all nodes in the distribution network as the input observation matrix of the invertible neural network model; let the observation matrix acquired at time k be... Where column vector It contains the voltage variables of N measurement channels observed at time k; the observation matrix at each time moment consists of the observation vector at the current time moment and K-1 historical observation vectors;

[0141] The observation matrix is ​​normalized using the 0-1 normalization method:

[0142]

[0143] In the formula, M k,min and M k,max Sample M k The minimum and maximum values.

[0144] In one specific implementation: the intercepted distribution network voltage data is input into a pre-trained reversible neural network model, and the log-likelihood value J is calculated and output. t The steps specifically include: standardizing the observation matrix... Input a pre-trained invertible neural network model and calculate the output log-likelihood value J. t ;

[0145] The log-likelihood value Jt With the abnormal event detection threshold J th The steps for obtaining the identification results of abnormal events in the distribution network include:

[0146] The log-likelihood value J t With the abnormal event detection threshold J th Comparison: When J t <J th Determine if an abnormal event has occurred and assess its risk level; when J t ≥J th The current status is determined to be normal;

[0147] The determination of the risk level of an abnormal event specifically includes: taking K' consecutive observation matrices before and after the time k of the abnormal event occurrence, and calculating the corresponding log-likelihood value J. k And normalize:

[0148]

[0149] In the formula, μ(J) k ) and σ(J k ) are the mean and standard deviation of the K' log-likelihood values ​​before and after time k, respectively;

[0150] The t-test is used to measure the log-likelihood after an abnormal event occurs. The degree of deviation from normal levels; under normal operating conditions, It follows a t-distribution with K'-1 degrees of freedom: After an abnormal event occurs, the log-likelihood value Perform a one-sided t-test:

[0151]

[0152] In the formula, after normalization 1-α is the confidence level, t α Let P' be the lower α quantile of the t-distribution with K'-1 degrees of freedom, P{·} be the probability operator, and μ be the probability value in the t-test. The population mean; the confidence interval at confidence level 1-α is

[0153] Construct confidence levels 1-α corresponding to different risk levels; obtain the log-likelihood value from the t-distribution table. The corresponding p-value is compared with α to determine the risk level of the abnormal event;

[0154] The results of the distribution network abnormal event identification include: whether an abnormal event has occurred in the distribution network, and the risk level of the abnormal event.

[0155] In one specific implementation: it further includes the step of calculating the Jacobian matrix of the output value of the invertible neural network model against the input value to locate the anomalous event; specifically including:

[0156] Input matrix for detecting abnormal events Calculate the output value J of the reversible neural network model k For the input matrix Jacobi matrix:

[0157]

[0158] In the formula, the Jacobi matrix R k Each element r i,t The corresponding element m in the input matrix was measured. i,t For J k Contribution level; |r i,t The larger |m i,t For J k The greater the contribution;

[0159] Get J k m, the one with the largest contribution i,t Locate the location information of the abnormal event:

[0160]

[0161] In the formula, β and δ represent the input matrix, respectively. element m i,t The spatial location and time information are used to locate the βth measurement channel.

[0162] In one specific implementation: it further includes the step of obtaining measurement data of voltage, line current, and differential current at both ends of the line where the abnormal event occurs, based on the location information of the abnormal event, and identifying the type of abnormal event using a pre-trained convolutional neural network constructed based on pseudo-labeling technology; specifically including:

[0163] Based on the location information of the abnormal event, locate the line with the largest change in current amplitude associated with the corresponding node in the location information of the abnormal event, obtain the measurement data of voltage at both ends of the line, line current and differential current, and form a feature matrix D; input the feature matrix D into a pre-trained convolutional neural network based on pseudo-label technology to identify the type of abnormal event.

[0164] In one specific implementation: the training method for the pre-trained convolutional neural network constructed based on pseudo-label technology includes:

[0165] Offline simulation of various abnormal events under different operating scenarios is used to obtain measurement data of voltage, line current and differential current at both ends of the disturbed line, and to construct the data samples required for training the convolutional neural network based on pseudo-label technology; the measurement data of the abnormal event samples generated in each simulation are concatenated into a feature matrix D.

[0166] The required data samples for training a convolutional neural network based on pseudo-label technology are used to train the constructed convolutional neural network based on pseudo-label technology offline, thereby obtaining a pre-trained convolutional neural network constructed based on pseudo-label technology.

[0167] In one specific implementation: the various abnormal events include three-phase grounding, two-phase grounding, heavy load switching, and line tripping.

[0168] In one specific implementation: the step of using data samples required for training a convolutional neural network based on pseudo-label technology, and offline training the constructed convolutional neural network based on pseudo-label technology to obtain a pre-trained convolutional neural network based on pseudo-label technology specifically includes:

[0169] A convolutional neural network model for classifying abnormal events in a power distribution network is constructed. The convolutional neural network model includes convolutional layers, pooling layers, and fully connected layers. The input of the neural network is a feature matrix D, the output is the abnormal event category, and the loss function is cross-entropy.

[0170] The acquired abnormal event samples are divided into two parts: labeled samples and unlabeled samples. Labeled samples are denoted as L = {(x...} h ,y h ):h∈(1,…,H)}, unlabeled samples are denoted as Q={q h :h∈(1,…,ρH)}, where x h y h Let q represent the feature matrix of the labeled sample and its corresponding anomaly event category, respectively. h The feature matrix represents the unlabeled samples, where H, ρH, and ρ are the number of labeled samples, the number of unlabeled samples, and the relative ratio of the number of labeled samples to the number of unlabeled samples, respectively.

[0171] The feature matrices of labeled and unlabeled samples are input into the convolutional neural network model respectively; for the labeled sample portion, its cross-entropy loss function L1 is calculated:

[0172]

[0173] In the formula c is the number of categories, p(w|x) h Let p be the predicted probability vector, where each element p k (w|x hCharacterize sample x h The probability assigned to category k. It is a one-hot encoding of category k. For the unlabeled sample portion, pseudo-labels u are assigned based on the prediction results of the convolutional neural network. h And select those with a confidence level higher than the confidence threshold. Calculate the cross-entropy loss function L2 for pseudo-labeled samples:

[0174]

[0175] In the formula, This is the confidence threshold for pseudo-labeled samples; the loss function is calculated as follows:

[0176] L=L1+ζL2

[0177] In the formula, L1 is the labeled loss function, L2 is the pseudo-label loss function, and ζ is the balance coefficient that controls the proportion of the pseudo-label loss function.

[0178] The required data samples for training a convolutional neural network based on pseudo-labeling technology are used. The constructed convolutional neural network based on pseudo-labeling technology is trained offline until the loss function converges, thus obtaining a pre-trained convolutional neural network constructed based on pseudo-labeling technology.

[0179] Example 4

[0180] Please see Figure 11 As shown, the present invention also provides a power distribution network abnormal event identification device, comprising:

[0181] The data acquisition module is used to capture distribution network voltage data online using a continuously moving window.

[0182] The calculation module is used to input the captured distribution network voltage data into a pre-trained reversible neural network model and calculate the output log-likelihood value J. t ;

[0183] The judgment module is used to determine the log-likelihood value J. t With the abnormal event detection threshold J th By comparison, the results of abnormal event identification in the power distribution network are obtained;

[0184] The output module is used to output the identification results of abnormal events in the power distribution network.

[0185] Example 5

[0186] Please see Figure 12As shown, the present invention also provides an electronic device 100 for implementing a method for identifying abnormal events in a power distribution network; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0187] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the power distribution network abnormal event identification method described in Embodiments 1, 2, or 3 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0188] The at least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor. The processor 102 is the control center of the electronic device 100, connecting various parts of the electronic device 100 via various interfaces and lines.

[0189] The memory 101 in the electronic device 100 stores multiple instructions to implement a method for identifying abnormal events in a power distribution network, and the processor 102 can execute the multiple instructions to achieve the following:

[0190] The distribution network voltage data is captured online using a continuously moving window.

[0191] The captured distribution network voltage data is input into a pre-trained reversible neural network model, and the log-likelihood value J is calculated. t ;

[0192] The log-likelihood value J t With the abnormal event detection threshold J th By comparison, the results of abnormal event identification in the power distribution network are obtained;

[0193] Output the results of abnormal events in the power distribution network.

[0194] Example 6

[0195] If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0196] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0197] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0198] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0199] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0200] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying abnormal events in a power distribution network, characterized in that, include: The distribution network voltage data is captured online using a continuously moving window. The captured distribution network voltage data is input into a pre-trained reversible neural network model, and the log-likelihood value is calculated. ; log-likelihood value With abnormal event detection threshold By comparison, the results of abnormal event identification in the power distribution network are obtained; Output the results of abnormal events in the power distribution network; The step of using a continuously moving window to capture distribution network voltage data online specifically includes: Use size is N × K A continuously moving window is used to extract the three-phase voltage data of all nodes in the distribution network as the input observation matrix of the invertible neural network model; denoted as at time t... k The obtained observation matrix is , where column vectors Including time k Observed N The voltage variables of each measurement channel; the observation matrix at each time step consists of the observation vector at the current time step and... K -1 historical observation vectors constitute; The observation matrix is ​​normalized using the 0-1 normalization method: （2） In the formula, and Samples The minimum and maximum values; The intercepted distribution network voltage data is input into a pre-trained reversible neural network model, and the log-likelihood value is calculated and output. The steps specifically include: standardizing the observation matrix... Input a pre-trained invertible neural network model and calculate the output log-likelihood value. ; The log-likelihood value With abnormal event detection threshold The steps for obtaining the identification results of abnormal events in the distribution network include: log-likelihood value With abnormal event detection threshold Comparison: When Determine if an abnormal event has occurred and assess its risk level; when The current status is determined to be normal; The determination of the risk level of an abnormal event specifically includes: at the time the abnormal event occurs. k Take the consecutive values ​​before and after. For each observation matrix, calculate the corresponding log-likelihood value. And normalize: （3） In the formula, and They are time points k front and back The mean and standard deviation of the log-likelihood values; The t-test is used to measure the log-likelihood after an abnormal event occurs. The degree of deviation from normal levels; under normal operating conditions, obey t-distribution of degrees of freedom: After an abnormal event occurs, the log-likelihood value is... Perform a one-sided t-test: （4） In the formula, after normalization , , For confidence level, for The lower bound of the t-distribution of degrees of freedom Quantiles For probability operators, In the t-test The overall mean; confidence level The confidence interval below is ; Construct confidence levels corresponding to different risk levels Based on the t-distribution table, find the log-likelihood value. corresponding p value and Compare and determine the risk level of abnormal events; The results of the distribution network abnormal event identification include: whether an abnormal event has occurred in the distribution network, and the risk level of the abnormal event.

2. The method for identifying abnormal events in a power distribution network according to claim 1, characterized in that, The pre-trained reversible neural network model is obtained through the following steps: Normal operating state data samples for training a reversible neural network model were obtained through offline time-domain simulation. The reversible neural network model is trained offline using normal operating state data samples to obtain a pre-trained reversible neural network model. The reversible neural network model includes a compression layer and several streaming modules; the input training normal operation state data samples are processed through the compression layer and several streaming modules, and the output is the log-likelihood value relative to the standard Gaussian distribution. The reversible neural network model is trained offline using normalized sample data. As input, the inverse neural network model is trained using the negative of the log-likelihood value as the loss function; the parameters of the inverse neural network model are randomly initialized and iteratively updated using mini-batch gradient descent until the network converges, thus obtaining the trained inverse neural network model. Input normalized sample data The minimum log-likelihood of the trained invertible neural network model is denoted as... ,Pick As an abnormal event detection threshold is a coefficient.

3. The method for identifying abnormal events in a power distribution network according to claim 1, characterized in that, It also includes the step of calculating the Jacobian matrix of the output value versus the input value of the invertible neural network model to locate anomalous events; specifically including: Input matrix for detecting abnormal events Calculate the output value of the reversible neural network model For the input matrix Jacobi matrix: （5） In the formula, the Jacobi matrix Each element The corresponding elements in the input matrix were measured. for The degree of contribution; The larger, right The greater the contribution; Get the The biggest contributor Locate the location information of the abnormal event: （6） In the formula, and Representing the input matrix respectively medium elements Spatial location and temporal information, by Locate the first One measurement channel.

4. The method for identifying abnormal events in a power distribution network according to claim 3, characterized in that, It also includes the steps of obtaining measurement data of voltage, line current, and differential current at both ends of the line where the abnormal event occurs, based on the location information of the abnormal event, and identifying the type of abnormal event using a pre-trained convolutional neural network constructed based on pseudo-labeling technology; specifically including: Based on the location information of the abnormal event, locate the line with the largest change in current amplitude associated with the corresponding node in the location information of the abnormal event, obtain the measurement data of voltage at both ends of the line, line current and differential current, and form a feature matrix D; input the feature matrix D into a pre-trained convolutional neural network based on pseudo-label technology to identify the type of abnormal event.

5. The method for identifying abnormal events in a power distribution network according to claim 4, characterized in that, The training method for the pre-trained convolutional neural network constructed based on pseudo-labeling technology includes: Offline simulation of various abnormal events under different operating scenarios is used to obtain measurement data of voltage, line current and differential current at both ends of the disturbed line, and to construct the data samples required for training the convolutional neural network based on pseudo-label technology; the measurement data of the abnormal event samples generated in each simulation are concatenated into a feature matrix D. The required data samples for training a convolutional neural network based on pseudo-label technology are used to train the constructed convolutional neural network based on pseudo-label technology offline, thereby obtaining a pre-trained convolutional neural network constructed based on pseudo-label technology.

6. The method for identifying abnormal events in a power distribution network according to claim 5, characterized in that, The various abnormal events include three-phase grounding, two-phase grounding, heavy load switching, and line tripping.

7. The method for identifying abnormal events in a power distribution network according to claim 5, characterized in that, The steps of using data samples required for training a convolutional neural network based on pseudo-labeling technology, and offline training the constructed convolutional neural network based on pseudo-labeling technology to obtain a pre-trained convolutional neural network based on pseudo-labeling technology, specifically include: A convolutional neural network model for classifying abnormal events in a power distribution network is constructed. The convolutional neural network model includes convolutional layers, pooling layers, and fully connected layers. The input of the neural network is a feature matrix D, the output is the abnormal event category, and the loss function is cross-entropy. The acquired abnormal event samples are divided into two parts: labeled samples and unlabeled samples. Labeled samples are denoted as... Unlabeled samples are denoted as ,in , These represent the feature matrices of labeled samples and their corresponding anomaly event categories, respectively. The feature matrix representing unlabeled samples, , , These represent the number of labeled samples, the number of unlabeled samples, and the relative ratio of labeled samples to unlabeled samples, respectively. The feature matrices of labeled and unlabeled samples are input into the convolutional neural network model respectively; for the labeled sample portion, its cross-entropy loss function is calculated. : （7） In the formula , For the number of categories, For the prediction probability vector, each element Characterization of samples Assigned to category The probability, It is a category One-hot encoding, For unlabeled samples, pseudo-labels are assigned based on the prediction results of the convolutional neural network. And select the pseudo-label sample with the highest confidence to calculate the cross-entropy loss function. : （8） In the formula, This is the confidence threshold for pseudo-labeled samples; the loss function is calculated as follows: In the formula, It is a balancing coefficient that controls the proportion of the pseudo-label loss function; The required data samples for training a convolutional neural network based on pseudo-label technology are used. The constructed convolutional neural network based on pseudo-label technology is trained offline until the loss function converges, thus obtaining a pre-trained convolutional neural network constructed based on pseudo-label technology.

8. A power distribution network abnormal event identification device, characterized in that, include: The data acquisition module is used to capture distribution network voltage data online using a continuously moving window. The calculation module is used to input the captured distribution network voltage data into a pre-trained reversible neural network model and calculate the output log-likelihood value. ; The judgment module is used to determine the log-likelihood value. With abnormal event detection threshold By comparison, the results of abnormal event identification in the power distribution network are obtained; The output module is used to output the identification results of abnormal events in the power distribution network; The acquisition module uses a continuously moving window to capture distribution network voltage data online, specifically including: Use size is N × K A continuously moving window is used to extract the three-phase voltage data of all nodes in the distribution network as the input observation matrix of the invertible neural network model; denoted as at time t... k The obtained observation matrix is , where column vectors Including time k Observed N The voltage variables of each measurement channel; the observation matrix at each time step consists of the observation vector at the current time step and... K -1 historical observation vectors constitute; The observation matrix is ​​normalized using the 0-1 normalization method: （2） In the formula, and Samples The minimum and maximum values; The calculation module inputs the captured distribution network voltage data into a pre-trained reversible neural network model and calculates the log-likelihood value. Specifically, this includes: standardizing the observation matrix Input a pre-trained invertible neural network model and calculate the output log-likelihood value. ; The log-likelihood value With abnormal event detection threshold The steps for obtaining the identification results of abnormal events in the distribution network include: log-likelihood value With abnormal event detection threshold Comparison: When Determine if an abnormal event has occurred and assess its risk level; when The current status is determined to be normal; The determination of the risk level of an abnormal event specifically includes: at the time the abnormal event occurs. k Take the consecutive values ​​before and after. For each observation matrix, calculate the corresponding log-likelihood value. And normalize: （3） In the formula, and They are time points k front and back The mean and standard deviation of the log-likelihood values; The t-test is used to measure the log-likelihood after an abnormal event occurs. The degree of deviation from normal levels; under normal operating conditions, obey t-distribution of degrees of freedom: After an abnormal event occurs, the log-likelihood value is... Perform a one-sided t-test: （4） In the formula, after normalization , , For confidence level, for The lower bound of the t-distribution of degrees of freedom Quantiles For probability operators, In the t-test The overall mean; confidence level The confidence interval below is ; Construct confidence levels corresponding to different risk levels Based on the t-distribution table, find the log-likelihood value. corresponding p value and Compare and determine the risk level of abnormal events; The results of the distribution network abnormal event identification include: whether an abnormal event has occurred in the distribution network, and the risk level of the abnormal event.

9. The power distribution network abnormal event identification device according to claim 8, characterized in that, The pre-trained reversible neural network model is obtained through the following steps: Normal operating state data samples for training a reversible neural network model were obtained through offline time-domain simulation. The reversible neural network model is trained offline using normal operating state data samples to obtain a pre-trained reversible neural network model. The reversible neural network model includes a compression layer and several streaming modules; the input training normal operation state data samples are processed through the compression layer and several streaming modules, and the output is the log-likelihood value relative to the standard Gaussian distribution. The reversible neural network model is trained offline using normalized sample data. As input, the inverse neural network model is trained using the negative of the log-likelihood value as the loss function; the parameters of the inverse neural network model are randomly initialized and iteratively updated using mini-batch gradient descent until the network converges, thus obtaining the trained inverse neural network model. Input normalized sample data The minimum log-likelihood of the trained invertible neural network model is denoted as... ,Pick As an abnormal event detection threshold is a coefficient.

10. The power distribution network abnormal event identification device according to claim 8, characterized in that, The judgment module is also used to calculate the Jacobian matrix of the output value of the invertible neural network model against the input value to locate abnormal events; specifically including: Input matrix for detecting abnormal events Calculate the output value of the reversible neural network model For the input matrix Jacobi matrix: （5） In the formula, the Jacobi matrix Each element The corresponding elements in the input matrix were measured. for The degree of contribution; The larger, right The greater the contribution; Get the The biggest contributor Locate the location information of the abnormal event: （6） In the formula, and Representing the input matrix respectively medium elements Spatial location and temporal information, by Locate the first One measurement channel.

11. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the power distribution network abnormal event identification method as described in any one of claims 1 to 7.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the power distribution network abnormal event identification method as described in any one of claims 1 to 7.

Images