Centralized-distributed cooperative distribution network fault detection method based on LOF-VAE-GPT

Abstract

The invention discloses a centralized-distributed cooperative distribution network fault detection method based on LOF-VAE-GPT. The method comprises the following steps: collecting physical data and network data of key edge nodes of a distribution network; based on a local outlier factor algorithm, constructing a physical anomaly detection model of the key edge node of the distribution network, and detecting an abnormal physical sample of the key edge node of the distribution network; proposing a network anomaly detection model of the key edge node of the distribution network based on a variational auto-encoder, and carrying out anomaly detection on a network sample of the key edge node of the distribution network; training and constructing a distribution network fault detection model based on a GPT model and the detected physical abnormal samples and network abnormal samples, and judging that edge nodes of the distribution network have faults when abnormal values of the nodes are detected to be greater than a threshold value. According to the method, the effectiveness of abnormal point identification in the physical data, the adaptability to the complex environment of the key edge node of the distribution network and the accuracy of fault detection of the key node of the distribution network can be improved, and the safe operation level of the distribution network is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system distribution network fault detection, and in particular to a centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT. Background Art

[0002] Detecting abnormalities and failures of key edge nodes in the distribution network is of great significance. Key edge nodes in the distribution network undertake important tasks of power transmission and distribution. Once these nodes are abnormal or fail, it may cause power supply interruption and affect the normal power consumption of users. In the existing technology, by timely detecting and handling these problems, the power outage time can be reduced and the power supply reliability can be improved; the abnormality of edge nodes may be a precursor to a larger accident. Through real-time monitoring and early warning, timely measures can be taken to prevent the expansion of accidents and ensure the safety of personnel and the integrity of equipment; through the automated monitoring system, the fault point can be quickly located, the time for on-site troubleshooting by operation and maintenance personnel can be reduced, and the efficiency of fault handling can be improved; regular detection and maintenance can reduce the cost of emergency fault repair, while extending the service life of equipment and reducing long-term operating costs; by deploying intelligent monitoring and diagnosis systems, the automation and intelligence of the distribution network can be promoted, and the scientificity and advancement of distribution network management can be improved; detection data can provide support for decision-making of the power system, such as equipment upgrades, capacity expansion, and scheduling; through the monitoring of edge nodes, energy management and optimization can be better carried out to improve the operating efficiency of the entire power grid; stable power supply can improve user satisfaction and reduce user complaints caused by power problems.

[0003] However, the existing methods for collecting network physical data and detecting anomalies at key edge nodes of distribution networks still have the following shortcomings: 1) Lack of real-time performance. Existing methods may not be able to collect and process large amounts of network physical data in real time, resulting in slow response in emergency situations. 2) Accuracy issues. Traditional methods may rely on simple threshold judgments and are easily affected by noise data and anomalies, resulting in false positives or negative positives. 3) Weak generalization capabilities. Some methods perform poorly when dealing with complex network structures and diverse data types, and are difficult to adapt to different network environments and conditions.

[0004] To this end, we designed a centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of insufficient real-time performance, insufficient accuracy and weak generalization ability in the network physical data collection and anomaly detection of key edge nodes of distribution networks in the prior art, and to propose a centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT to improve the effectiveness of anomaly point identification in physical data, improve the adaptability to the complex environment of key edge nodes of distribution networks, and improve the accuracy of fault detection of key nodes of distribution networks.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT includes the following steps:

[0008] Step 1: Collect physical data and network data of key edge nodes of the distribution network. Physical data includes current data, voltage data, power data, frequency data, and voltage phase angle data. Network data includes flow information, log information, and message information.

[0009] Step 2: Based on the local outlier factor algorithm, a physical anomaly detection model for key edge nodes of the distribution network is constructed to detect abnormal physical samples of key edge nodes of the distribution network and identify physical abnormal behaviors of key nodes in the distribution network;

[0010] Step 3: Propose a network anomaly detection model for key edge nodes of the distribution network based on variational autoencoder to perform anomaly detection on network samples of key edge nodes of the distribution network;

[0011] Step 4: Based on the GPT model and the detected physical anomaly samples and network anomaly samples, training is performed, the anomaly values ​​are calculated, and a distribution network fault detection model is constructed. The physical data and network data of the key edge nodes of the distribution network collected in real time are input into the distribution network fault detection model. When the anomaly value of the detected node is greater than the threshold, it is determined that a fault has occurred at the edge node of the distribution network.

[0012] To further illustrate, in step 1, the physical data and network data collected from the key edge nodes of the distribution network are expressed using the following formula:

[0013] X Gi (t) = [I i (t),U i (t),P i (t),f i (t),θ i (t)] (1)

[0014] X Wi (t) = [L i (t),R i (t),B i(t)] (2)

[0015] In the above formula, X Gi (t) is the physical data of the key edge node i of the distribution network collected at time t, I i (t) is the current data collected at the key edge node i of the distribution network at time t, U i (t) is the voltage data of the key edge node i of the distribution network at time t, P i (t) is the power data collected from the key edge node i of the distribution network at time t, f i (t) is the frequency data of the key edge node i of the distribution network collected at time t, θ i (t) is the voltage phase angle data of the key edge node i of the distribution network at time t; X Wi (t) is the network data of key edge node i of the distribution network collected at time t, L i (t) is the traffic information of the key edge node i of the distribution network collected at time t, R i (t) is the log information of the key edge node i of the distribution network collected at time t, B i (t) is the message information of the key edge node i of the distribution network collected at time t.

[0016] Further explanation: In step 2, based on the local outlier factor algorithm, a physical anomaly detection model of key edge nodes of the distribution network is constructed to detect abnormal physical samples of key edge nodes of the distribution network and identify physical abnormal behaviors of key nodes in the distribution network, including the following steps:

[0017] Step 21: For each physical sample of the key edge node of the distribution network, the k-distance calculation formula is used to calculate the k-distance between the physical samples:

[0018] d k (i,j)=(I i (t)-I j (t)) 2 +(U i (t)-U j (t)) 2 +(P i (t)-P j (t)) 2 +(f i (t)-f j (t)) 2 +(θ i (t)-θ j (t)) 2 (3)

[0020] Where, d k(i, j) represents the distance between the physical samples of the key edge node i and node j of the distribution network, and k represents the preset parameter; I j (t) is the current data collected at the key edge node j of the distribution network at time t, U j (t) is the voltage data of the key edge node j of the distribution network at time t, P j (t) is the power data collected from the key edge node j of the distribution network at time t, f j (t) is the frequency data of the key edge node j of the distribution network collected at time t, θ j (t) is the voltage phase angle data of the key edge node j of the distribution network at time t;

[0021] Step 22: For the physical samples of key edge nodes of the distribution network, the local reachability density of the physical samples of the key edge nodes of the distribution network is calculated using the calculation formula of the local reachability density of the physical samples of the key edge nodes of the distribution network. k (i)

[0022]

[0023] Where N k (i) is the set of k nearest neighbors of the physical sample;

[0024] Step 23, calculate the local outlier factor LOF of each physical sample k (i)

[0025]

[0026] In the formula, ρ k (j) is the local reachability density of the physical samples of the key edge node j of the distribution network;

[0027] Step 24, perform anomaly detection on the physical samples of the key edge nodes of the distribution network, and determine whether the physical samples of the key edge nodes of the distribution network are abnormal based on the local outlier factor:

[0028] LOF k (i)≥δ k (6)

[0029] In the formula, δ k is the physical sample abnormal threshold, when LOF k (i)≥δ k When the physical samples of the key edge nodes of the distribution network are marked as abnormal, the physical abnormal behaviors of the key nodes in the distribution network are identified.

[0030] Further explanation: In step 3, a network anomaly detection model for key edge nodes of the distribution network based on variational autoencoder is proposed to detect anomalies of network samples of key edge nodes of the distribution network, using the following steps:

[0031] Step 31, based on the distribution network key edge node network sample encoding function of the variational autoencoder, encode the network sample, the expression is as follows:

[0032] Z w,i =f VAE,BM (X W,i ) (7)

[0033] In the formula, Z w,i represents the potential sample after encoding of the key edge node i in the distribution network, f VAE,BM () is the network sample encoding function of the key edge node i of the distribution network based on the variational autoencoder, X W,i is the network sample of key edge node i of the distribution network;

[0034] Step 32, the network approximate posterior distribution of the key edge nodes of the distribution network based on the variational autoencoder is expressed by the following formula:

[0035]

[0036] μ W =f 1 (X W,i ) (9)

[0037]

[0038] In the formula, q(Z w,i |X W,i ) is the network approximate posterior distribution value of the key edge nodes of the distribution network based on the variational autoencoder, N VAE () is the approximate posterior distribution function of the key edge node network of the distribution network based on variational autoencoder, μ W is the network sample mean based on the neural network model, is the variance of the neural network model, f 1 (), f 2 () are BP neural network and convolutional neural network model functions respectively;

[0039] Step 33, based on the loss function in the variational autoencoder model training process, the gradient descent method is used to minimize the total loss of the variational autoencoder, and the variational autoencoder model is trained. After the training is completed, the network sample to be detected is input into the trained variational autoencoder model, the potential sample is output, and the potential sample is decoded to obtain the reconstructed distribution network key edge node network sample:

[0040] L C =f E [log(P(X W,i |Z W,i )]+f D [q(Zw,i |X W,i )||P(Z W,i )] (11)

[0041] X' W,i =f VAE,BM (Z W,i ) (12)

[0042] Where, L C is the loss function in the variational autoencoder model training process; P(X W,i |Z W,i ) is the network sample distribution, according to X W,i The sample distribution P(X W,i ) and Z W,i The sample distribution P(Z W,i ) is calculated by the ratio of E [], f D [] are the reconstruction loss function and KL divergence loss function respectively, X' W,i It is a network sample of key edge nodes of the reconstructed distribution network;

[0043] Step 34, construct a network anomaly detection model for key edge nodes of the distribution network, calculate the difference between the original data and the reconstructed network samples of key edge nodes of the distribution network, and perform anomaly detection on the network samples of key edge nodes of the distribution network:

[0044] S W,i =f VAE,C (X W,i ,X' W,i ) (13)

[0045] S W,i >δ W (14)

[0046] In the formula, S W,i is the abnormality score of the network sample, f VAE,C () is the scoring calculation model function, δ W is the network scoring abnormality threshold, when S W,i >δ W When the key edge node network samples of the distribution network are detected, they are judged to be abnormal.

[0047] To further illustrate, in step 4, training is performed based on the GPT model and the detected physical anomaly samples and network anomaly samples to build a distribution network fault detection model, and the expression is as follows:

[0048]

[0049] Y i ≥δ GPT (16)

[0050] Where Y i is the distribution network fault detection score, f GPT () represents the distribution network fault detection score calculation model function, For the detected physical anomaly samples, is the detected network anomaly sample, δ GPT is the scoring threshold for distribution network fault detection; the physical data and network data of key edge nodes of the distribution network collected in real time are input into the distribution network fault detection model. i ≥δ GPT , it indicates that a failure has occurred in the edge node of the distribution network.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention performs anomaly detection on the physical information data and network information data of the key edge nodes of the distribution network based on the local outlier factor algorithm (LOF) and the variational autoencoder (VAE) algorithm, trains the GPT model based on the physical anomaly and network anomaly sample data, builds a distribution network fault detection model, and detects whether there is a fault in the key edge nodes of the distribution network. It effectively makes up for the shortcomings of insufficient real-time performance, insufficient accuracy, and weak generalization ability in the network physical data collection and anomaly detection of the key edge nodes of the distribution network in the prior art. The use of the collaborative distribution network fault detection method proposed in the present invention can improve the effectiveness of abnormal point identification in physical data, improve the adaptability to the complex environment of the key edge nodes of the distribution network, improve the accuracy of fault detection of the key nodes of the distribution network, and improve the safe operation level of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flow chart of a centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT proposed in the present invention. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0054] The centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT proposed in this embodiment first collects network and physical data of key edge nodes of the distribution network, a physical anomaly detection model of key edge nodes of the distribution network based on the local outlier factor algorithm (LOF), a network anomaly detection model of key edge nodes of the distribution network based on the variational autoencoder (VAE), and a distribution network fault detection model based on the GPT model to detect the abnormal values ​​of the nodes and determine whether the edge nodes of the distribution network have failed.

[0055] Specifically, refer to Figure 1Next, we will introduce in detail the centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT. The method mainly includes the following steps:

[0056] Step 1: Collect physical data and network data of key edge nodes of the distribution network. Physical data includes current data, voltage data, power data, frequency data, and voltage phase angle data. Network data includes flow information, log information, and message information.

[0057] Collect the physical data and network data of key edge nodes of the distribution network and use the following formula to express it:

[0058] X Gi (t) = [I i (t),U i (t),P i (t),f i (t),θ i (t)] (1)

[0059] X Wi (t) = [L i (t),R i (t),B i (t)] (2)

[0060] In the above formula, X Gi (t) is the physical data of the key edge node i of the distribution network collected at time t, I i (t) is the current data collected at the key edge node i of the distribution network at time t, U i (t) is the voltage data of the key edge node i of the distribution network at time t, P i (t) is the power data collected from the key edge node i of the distribution network at time t, f i (t) is the frequency data of the key edge node i of the distribution network collected at time t, θ i (t) is the voltage phase angle data of the key edge node i of the distribution network at time t; X Wi (t) is the network data of key edge node i of the distribution network collected at time t, L i (t) is the traffic information of the key edge node i of the distribution network collected at time t, R i (t) is the log information of the key edge node i of the distribution network collected at time t, B i (t) is the message information of the key edge node i of the distribution network collected at time t.

[0061] Step 2: Based on the local outlier factor algorithm, a physical anomaly detection model for key edge nodes of the distribution network is constructed to detect abnormal physical samples of key edge nodes of the distribution network and identify physical abnormal behaviors of key nodes in the distribution network. Specifically, the following steps are included:

[0062] Step 21: For each physical sample of the key edge node of the distribution network, the k-distance calculation formula is used to calculate the k-distance between the physical samples:

[0063] d k (i,j)=(I i (t)-I j (t)) 2 +(U i (t)-U j (t)) 2 +(P i (t)-P j (t)) 2

[0064] +(f i (t)-f j (t)) 2 +(θ i (t)-θ j (t)) 2 (3)

[0065] Where, d k (i, j) represents the distance between the physical samples of the key edge node i and node j of the distribution network, and k represents the preset parameter; I j (t) is the current data collected at the key edge node j of the distribution network at time t, U j (t) is the voltage data of the key edge node j of the distribution network at time t, P j (t) is the power data collected from the key edge node j of the distribution network at time t, f j (t) is the frequency data of the key edge node j of the distribution network collected at time t, θ j (t) is the collected voltage phase angle data of the key edge node j of the distribution network at time t.

[0066] Step 22: For the physical samples of key edge nodes of the distribution network, the local reachability density of the physical samples of the key edge nodes of the distribution network is calculated using the calculation formula of the local reachability density of the physical samples of the key edge nodes of the distribution network. k (i)

[0067]

[0068] Where N k (i) is the set of k nearest neighbors of the physical sample;

[0069] Step 23, calculate the local outlier factor LOF of each physical sample k (i)

[0070]

[0071] In the formula, ρ k (j) is the local reachable density of physical samples of key edge node j in the distribution network.

[0072] Step 24, perform anomaly detection on the physical samples of the key edge nodes of the distribution network, and determine whether the physical samples of the key edge nodes of the distribution network are abnormal based on the local outlier factor:

[0073] LOF k (i)≥δ k (6)

[0074] In the formula, δ k is the physical sample abnormal threshold, when LOF k (i)≥δ k When the physical samples of the key edge nodes of the distribution network are marked as abnormal, the physical abnormal behaviors of the key nodes in the distribution network are identified.

[0075] Step 3: Propose a network anomaly detection model for key edge nodes of the distribution network based on variational autoencoder to perform anomaly detection on network samples of key edge nodes of the distribution network.

[0076] Use the following steps:

[0077] Step 31, based on the distribution network key edge node network sample encoding function of the variational autoencoder, encode the network sample, the expression is as follows:

[0078] Z w,i =f VAE,BM (X W,i ) (7)

[0079] In the formula, Z w,i represents the potential sample after encoding of the key edge node i in the distribution network, f VAE,BM () is the network sample encoding function of the key edge node i of the distribution network based on the variational autoencoder, X W,i is the network sample of key edge node i of the distribution network;

[0080] Step 32, the network approximate posterior distribution of the key edge nodes of the distribution network based on the variational autoencoder is expressed by the following formula:

[0081]

[0082] μ W =f 1 (X W,i ) (9)

[0083]

[0084] In the formula, q(Z w,i|X W,i ) is the network approximate posterior distribution value of the key edge nodes of the distribution network based on the variational autoencoder, N VAE () is the approximate posterior distribution function of the key edge node network of the distribution network based on variational autoencoder, μ W is the network sample mean based on the neural network model, is the variance of the neural network model, f 1 (), f 2 () are BP neural network and convolutional neural network model functions respectively;

[0085] Step 33, based on the loss function in the variational autoencoder model training process, use the gradient descent method to minimize the total loss of the variational autoencoder (the sum of the reconstruction loss and the KL divergence loss), train the variational autoencoder model, and after the training, input the network sample to be tested into the trained variational autoencoder model, output the potential sample, and decode the potential sample to obtain the reconstructed distribution network key edge node network sample:

[0086] L C =f E [log(P(X W,i |Z W,i )]+f D [q(Z w,i |X W,i )||P(Z W,i )] (11)

[0087] X' W,i =f VAE,BM (Z W,i ) (12)

[0088] Where, L C is the loss function in the training process of the variational autoencoder model; P(X W,i |Z W,i ) is the network sample distribution, according to X W,i The sample distribution P(X W,i ) and Z W,i The sample distribution P(Z W,i ) is calculated by the ratio of E [], f D [] are the reconstruction loss function and KL divergence loss function respectively, X' W,i This is a sample of the reconstructed key edge node network of the distribution network.

[0089] Step 34, construct a network anomaly detection model for key edge nodes of the distribution network, calculate the difference between the original data and the reconstructed network samples of key edge nodes of the distribution network, and perform anomaly detection on the network samples of key edge nodes of the distribution network:

[0090] S W,i =f VAE,C (X W,i ,X' W,i ) (13)

[0091] S W,i >δ W (14)

[0092] In the formula, S W,i is the abnormality score of the network sample, f VAE,C () is the scoring calculation model function, δ W is the network scoring abnormality threshold, when S W,i >δ W When the key edge node network samples of the distribution network are detected, they are judged to be abnormal.

[0093] Step 4: Based on the GPT model, the physical anomaly samples detected in step 2 are and the network anomaly samples detected in step 3 It is trained based on the GPT model and the detected physical anomaly samples and network anomaly samples to build a distribution network fault detection model, which is expressed as follows:

[0094]

[0095] Y i ≥δ GPT (16)

[0096] Where Y i is the distribution network fault detection score, f GPT () represents the distribution network fault detection score calculation model function, For the detected physical anomaly samples, is the detected network anomaly sample, δ GPT This parameter is the scoring threshold for distribution network fault detection.

[0097] The physical data and network data of the key edge nodes of the distribution network collected in real time are input into the distribution network fault detection model. i ≥δ GPT , it indicates that a failure has occurred in the edge node of the distribution network.

[0098] Through the above steps, effective detection of key edge node failures in the distribution network can be achieved.

[0099] The present invention performs abnormal detection on the physical information and network information of key edge nodes of the distribution network based on the local outlier factor algorithm (LOF) and the variational autoencoder (VAE) algorithm, detects physical and network anomalies based on the GPT model, and detects whether there is a fault in the distribution network node. It can improve the effectiveness of abnormal point identification in physical data, improve the adaptability to the complex environment of key edge nodes of the distribution network, improve the accuracy of fault detection of key nodes of the distribution network, and improve the safe operation level of the distribution network.

[0100] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT, characterized in that: The following steps are involved: Step 1: Collect physical data and network data of key edge nodes of the distribution network. Physical data includes current data, voltage data, power data, frequency data, and voltage phase angle data. Network data includes flow information, log information, and message information. Step 2: Based on the local outlier factor algorithm, a physical anomaly detection model for key edge nodes of the distribution network is constructed to detect abnormal physical samples of key edge nodes of the distribution network and identify physical abnormal behaviors of key nodes in the distribution network; Step 3: Propose a network anomaly detection model for key edge nodes of the distribution network based on variational autoencoder to perform anomaly detection on network samples of key edge nodes of the distribution network; Step 4: Based on the GPT model and the detected physical anomaly samples and network anomaly samples, training is performed to build a distribution network fault detection model. The physical data and network data of the key edge nodes of the distribution network collected in real time are input into the distribution network fault detection model. When the abnormal value of the detected node is greater than the threshold, it is determined that a fault has occurred at the edge node of the distribution network.

2. According to claim 1, a centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT is characterized in that: In step 1, the physical data and network data of the key edge nodes of the distribution network are collected and expressed by the following formula: X Gi (t)=[I i (t),U i (t),P i (t),f i (t),θ i (t)] (1) X Wi (t)=[L i (t),R i (t),B i (t)] (2) In the above formula, X Gi (t) is the physical data of the key edge node i of the distribution network collected at time t, I i (t) is the current data collected at the key edge node i of the distribution network at time t, U i (t) is the voltage data of the key edge node i of the distribution network at time t, P i (t) is the power data collected from the key edge node i of the distribution network at time t, f i (t) is the frequency data of the key edge node i of the distribution network collected at time t, θ i (t) is the voltage phase angle data of the key edge node i of the distribution network at time t; X Wi (t) is the network data of key edge node i of the distribution network collected at time t, L i (t) is the traffic information of the key edge node i of the distribution network collected at time t, R i (t) is the log information of the key edge node i of the distribution network collected at time t, B i (t) is the message information of the key edge node i of the distribution network collected at time t.

3. According to claim 2, a centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT is characterized in that: In step 2, based on the local outlier factor algorithm, a physical anomaly detection model for key edge nodes of the distribution network is constructed to detect abnormal physical samples of key edge nodes of the distribution network and identify physical abnormal behaviors of key nodes in the distribution network, including the following steps: Step 21: For each physical sample of the key edge node of the distribution network, the k-distance calculation formula is used to calculate the k-distance between the physical samples: d k (i,j)=(I i (t)-I j (t)) 2 +(U i (t)-U j (t)) 2 +(P i (t)-P j (t)) 2 +(f i (t)-f j (t)) 2 +(θ i (t)-θ j (t)) 2 (3) Where, d k (i, j) represents the distance between the physical samples of the key edge node i and node j of the distribution network, and k represents the preset parameter; I j (t) is the current data collected at the key edge node j of the distribution network at time t, U j (t) is the voltage data of the key edge node j of the distribution network at time t, P j (t) is the power data collected from the key edge node j of the distribution network at time t, f j (t) is the frequency data of the key edge node j of the distribution network collected at time t, θ j (t) is the voltage phase angle data of the key edge node j of the distribution network at time t; Step 22: For the physical samples of key edge nodes of the distribution network, the local reachability density of the physical samples of the key edge nodes of the distribution network is calculated using the calculation formula of the local reachability density of the physical samples of the key edge nodes of the distribution network. k (i) Where N k (i) is the set of k nearest neighbors of the physical sample; Step 23, calculate the local outlier factor LOF of each physical sample k (i) In the formula, ρ k (j) is the local reachability density of the physical samples of the key edge node j of the distribution network; Step 24, perform anomaly detection on the physical samples of the key edge nodes of the distribution network, and determine whether the physical samples of the key edge nodes of the distribution network are abnormal based on the local outlier factor: LOF k (i)≥δ k (6) In the formula, δ k is the physical sample abnormal threshold, when LOF k (i)≥δ k When the physical samples of the key edge nodes of the distribution network are marked as abnormal, the physical abnormal behaviors of the key nodes in the distribution network are identified.

4. According to claim 1, a centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT is characterized in that: In step 3, a network anomaly detection model for key edge nodes of the distribution network based on variational autoencoder is proposed to detect anomalies of network samples of key edge nodes of the distribution network. The following steps are adopted: Step 31, based on the distribution network key edge node network sample encoding function of the variational autoencoder, encode the network sample, the expression is as follows: Z w,i =f VAE,BM (X W,i ) (7) In the formula, Z w,i represents the potential sample after encoding of the key edge node i in the distribution network, f VAE,BM () is the network sample encoding function of the key edge node i of the distribution network based on the variational autoencoder, X W,i is the network sample of key edge node i of the distribution network; Step 32, the network approximate posterior distribution of the key edge nodes of the distribution network based on the variational autoencoder is expressed by the following formula: μ W =f1(X W,i ) (9) In the formula, q(Z w,i |X W,i ) is the network approximate posterior distribution value of the key edge nodes of the distribution network based on the variational autoencoder, N VAE () is the approximate posterior distribution function of the key edge node network of the distribution network based on variational autoencoder, μ W is the network sample mean based on the neural network model, is the variance based on the neural network model, f1() and f2() are the BP neural network and convolutional neural network model functions respectively; Step 33, based on the loss function in the variational autoencoder model training process, the gradient descent method is used to minimize the total loss of the variational autoencoder, and the variational autoencoder model is trained. After the training is completed, the network sample to be detected is input into the trained variational autoencoder model, the potential sample is output, and the potential sample is decoded to obtain the reconstructed distribution network key edge node network sample: L C =f E [log(P(X W,i |Z W,i )]+f D [q(Z w,i |X W,i )||P(Z W,i )] (11) X' W,i =f VAE,BM (Z W,i ) (12) Where, L C is the loss function in the training process of the variational autoencoder model; P(X W,i |Z W,i ) is the network sample distribution, f E [], f D [] are the reconstruction loss function and KL divergence loss function respectively, X' W,i It is a network sample of key edge nodes of the reconstructed distribution network; Step 34, construct a network anomaly detection model for key edge nodes of the distribution network, calculate the difference between the original data and the reconstructed network samples of key edge nodes of the distribution network, and perform anomaly detection on the network samples of key edge nodes of the distribution network: S W,i =f VAE,C (X W,i ,X' W,i ) (13) S W,i >d W (14) In the formula, S W,i is the abnormality score of the network sample, f VAE,C () is the scoring calculation model function, δ W is the network scoring abnormality threshold, when S W,i >δ W When the key edge node network samples of the distribution network are detected, they are judged to be abnormal.

5. According to claim 1, a centralized-distributed collaborative distribution network fault detection method based on LOF-VAE-GPT is characterized in that: Step 4: Based on the GPT model and the detected physical anomaly samples and network anomaly samples, training is performed, the anomaly values ​​are calculated, and a distribution network fault detection model is constructed. The expression is as follows: AND i ≥δ GPT (16) Where Y i is the distribution network fault detection score, f GPT () represents the distribution network fault detection score calculation model function, For the detected physical anomaly samples, is the detected network anomaly sample, δ GPT is the scoring threshold for distribution network fault detection; the physical data and network data of key edge nodes of the distribution network collected in real time are input into the distribution network fault detection model. i ≥δ GPT , it indicates that a failure has occurred in the edge node of the distribution network.

Images