Power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis

Abstract

The invention discloses a power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis. The method comprises the steps of S1, forming power operation data; s2, constructing a topological network model of the power grid based on the physical structure of the power grid; s3, using a first generator to generate abnormal data in the simulation power grid based on the power operation data; s4, optimization abnormal data generation is carried out; s5, embedding the power operation data and topological structure features of the power grid into the anomaly detection model through a graph convolutional network; s6, power grid anomaly detection parameters are dynamically adjusted through the adaptive learning model, so that the anomaly detection model is continuously optimized along with the change of the power grid operation condition; and S7, based on the topological information and abnormal data generated by the generative adversarial network, combining topological data analysis, and positioning a fault node or an abnormal line in the power grid. According to the invention, the accuracy of fault prediction and positioning is improved.

Description

Technical Field

[0001] The present invention relates to power grid anomaly detection and fault diagnosis in the technical field of new energy and smart grid, and in particular to a power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis. Background Art

[0002] As the scale and complexity of modern power systems continue to expand, anomaly detection and fault diagnosis in power grid operation are becoming increasingly important. Traditional power grid anomaly detection methods mainly rely on fixed rules or data monitoring based on physical sensors. Traditional methods usually make anomaly judgments based on preset static rules and cannot adapt to the complex and changeable operating environment of the power grid in real time. In addition, traditional anomaly detection methods are often limited to handling local faults. For complex multi-point faults or chain reaction faults, their detection accuracy and response speed are relatively limited, which can easily lead to major fault hazards not being discovered in time.

[0003] The existing power grid anomaly detection technology relies on data collected by physical sensors with limited accuracy, often ignoring the topological structure information of the power grid and failing to effectively capture the correlation between nodes within the power grid. Most existing detection systems perform analysis based on a single data source and have difficulty processing complex multi-source data, especially historical data in the power grid, real-time operation data, and multi-dimensional data of power grid topological structure characteristics. In addition, the existing technology cannot effectively handle multi-point faults and complex chain reaction fault modes in the power grid, resulting in inaccurate fault location and a lengthy repair process, which may extend the power grid recovery time.

[0004] In terms of abnormal data generation and fault prediction, existing power grid fault detection systems usually rely on traditional simulation methods and are unable to generate sufficiently realistic abnormal data for effective training of abnormal patterns. In addition, existing simulation data generation technologies lack dynamic adjustment and optimization capabilities, and the generated data patterns are relatively simple and cannot cover complex power grid fault types, especially the simulation capabilities of multi-point faults and chain reaction type faults.

[0005] Existing fault location technologies are usually unable to fully integrate the topological structure information of the power grid, resulting in insufficient accuracy in locating fault points in complex power grids. Especially when dealing with chain reaction faults, the identification of fault propagation paths is relatively vague, making it difficult to predict the potential risk range and impact area. Due to the lack of effective dynamic learning and adaptive adjustment mechanisms, existing power grid anomaly detection models often need to rely on manual intervention and are unable to automatically adapt to changes in power grid operating conditions, resulting in a significant decrease in the sensitivity and accuracy of anomaly detection under load changes or sudden abnormal conditions.

[0006] In summary, the existing power grid anomaly detection and fault diagnosis technologies have many shortcomings in the detection of complex fault modes, fault point location and adaptive adjustment. An innovative method combining artificial intelligence and topological data analysis is urgently needed to improve the accuracy of power grid anomaly detection, response speed and fault diagnosis accuracy. Summary of the invention

[0007] Based on the above problems, the present invention provides a power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis, aiming to improve the accuracy of power grid anomaly detection, response speed and fault diagnosis accuracy to meet the needs of expanding scale and increasing complexity of modern power systems.

[0008] A power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis according to an embodiment of the present invention comprises the following steps:

[0009] S1. Collect voltage, current and power from different nodes of the power grid through multiple sensors, and transmit the data to the data processing module to form power operation data;

[0010] S2. Construct a topological network model of the power grid based on the physical structure of the power grid. The nodes of the topological network model represent power plants, substations and key load centers, and the edges of the topological network model represent the power transmission relationship between nodes.

[0011] S3. Using the first generator to generate abnormal data in the simulated power grid based on the power operation data; using the first discriminator to discriminate the real power grid data and the generated abnormal data, and optimizing the abnormal data generated by the first generator;

[0012] S4. Based on S3, the power operation data and the abnormal data generated by the first layer are combined to generate abnormal data distribution, and the second discriminator is used to discriminate the difference between the abnormal data generated by the second layer and the real multi-source data;

[0013] S5. The power operation data and topological structure characteristics of the power grid are embedded into the anomaly detection model through a graph convolutional network, and anomaly detection is performed by combining real-time data and historical data. The anomaly pattern generated by the two-layer generative adversarial network is used to optimize the accuracy of power grid anomaly detection.

[0014] S6. Dynamically adjust the power grid anomaly detection parameters through the adaptive learning model, so that the anomaly detection model is continuously optimized as the power grid operation status changes;

[0015] S7. Based on the topology information and the abnormal data generated by the generative adversarial network, combined with the topology data analysis, locate the faulty nodes or abnormal lines in the power grid.

[0016] Optionally, the S1 step includes:

[0017] S11. Voltage sensors, current sensors and power sensors are arranged at multiple nodes of the power grid, the voltage sensor is used to collect voltage, the current sensor is used to collect current, and the power sensor is used to collect power, the nodes include power plants, substations and key load centers;

[0018] S12. Use data acquisition equipment to collect the real-time voltage data V of each node i (t), current data I i (t) and power data P i (t) continuously transmit to the power grid data processing module;

[0019] S13. According to the data transmission protocol, the power grid data processing module receives and stores all collected power operation data to form power operation data containing multiple time series:

[0020] D={(V i (t),I i (t),P i (t))|i=1,2,...,n; t=t0,t1,...,t k};

[0021] Where n represents the total number of acquisition nodes, and t is a time parameter used to represent the timestamp of the data, t0, t1, ..., t k is the data collection time at different time points.

[0022] Optionally, the S2 step includes:

[0023] S21. Define the power grid topology network model:

[0024] G = (N, E);

[0025] In the power grid topology network model, each node n is combined with the historical status of the node and the power grid operation characteristics. i ∈N state information is represented by S i (t):

[0026]

[0027] Among them, g k is a function of different historical states, used to extract historical data features, α k is the adaptive weight, f hist A nonlinear mapping function representing the historical state;

[0028] S22. In the graph neural network embedding of the power grid topology network model, the historical status and real-time power operation data of each node are mapped to a high-dimensional space. Through the dynamic update mechanism between nodes, the embedding of the node is expressed as:

[0029]

[0030] Among them, H i (t) indicates node n i The high-dimensional embedding vector at time t, W embed and W neigh are the weight matrices of the node itself and its neighbor nodes, Represents node n i The set of neighbor nodes, σ is the activation function;

[0031] S23. Using graph neural network to dynamically adjust the weight relationship between edges in the power grid topology network model to describe node n i With node n j Power transfer capability and abnormal pattern detection between:

[0032]

[0033] Among them, W ij (t) represents node n i With node n j The weights between them change dynamically with time t, H i (t)·H j (t) is the vector inner product between nodes, which is used to calculate the similarity between nodes;

[0034] S24. Combine power operation data and power grid topology network model to automatically extract potential abnormal features in the power grid through graph neural network training:

[0035]

[0036] in, Represents node n i The feature vector at layer l, W (l) and are the weight matrices of the layer respectively.

[0037] Optionally, the S3 step includes:

[0038] S31. Based on the multi-source data set of the power grid and the topological embedding features of the power grid topology network model, the first generator G1 is constructed, and the historical power operation data, the real-time collected power operation data and the topological structure embedding features are used as inputs, and the abnormal data D in the simulated power grid is generated by the generator in the generative adversarial network:

[0039] D gen =G1(D real ,T embed );

[0040] Among them, D real Represents the actual power operation data, Tembed Represents the topological structure embedding feature, the generated D gen Abnormal data is used to simulate abnormal patterns of power grids with different fault types;

[0041] S32. The first discriminator D1 receives the real abnormal data D real and the abnormal data D generated by the first generator G1 gen The goal of the first discriminator is to distinguish the generated abnormal data from the real abnormal data. The loss function is defined as:

[0042]

[0043] in, represents the loss function of the first discriminator, is the prediction accuracy of the real abnormal data, is the probability of misjudging the generated abnormal data;

[0044] S33. The goal of the first generator G1 is to maximize the error discrimination rate of the first discriminator D1, even if the generated abnormal data D gen Close to the real abnormal data D real , the first generator loss function is defined as:

[0045]

[0046] S34. During the training process of the generative adversarial network, the first generator G1 and the first discriminator D1 are iteratively trained until the abnormal data D generated by the first generator gen Compared with the real abnormal data D real Difficult to distinguish.

[0047] Optionally, the S4 step includes:

[0048] S41. In the generated abnormal data D gen Based on the above, a distribution model of abnormal data is generated by combining power operation data;

[0049] S42. Based on the distribution model, the second generator G2 is used to further generate the second layer of abnormal data reflecting different fault types and chain reactions. The second generator aims to combine the abnormal data D generated by the first layer gen and multi-source real data D real Optimize abnormal data;

[0050] S43. Use the second discriminator D2 to analyze the abnormal data of the second layer With multi-source real data D real The goal of the second discriminator is to distinguish the difference between complex abnormal data and real data:

[0051]

[0052] S44. The goal of the second generator G2 is to maximize the misjudgment rate of the second discriminator D2 and generate complex and difficult to distinguish abnormal data. The loss function of the second generator is expressed as:

[0053]

[0054] Optionally, the step S5 includes:

[0055] S51. Combine the power operation data of the power grid with the topological structure characteristics T embed By embedding the graph convolutional network into the anomaly detection model, the embedding of each node is expressed as:

[0056]

[0057] in, For node n i The embedded features in the l+1th layer of the graph convolutional network, c ij is node n i With node n j The normalization coefficient between ;

[0058] S52. The generator and discriminator of the two-layer generative adversarial network are used to optimize the power grid anomaly detection model. The first generator and the first discriminator are used to generate primary anomaly data and perform adversarial training with real data. The second generator and the second discriminator are used to generate more complex anomaly patterns and perform adversarial optimization. The optimized generative adversarial network is represented by G ipt .

[0059] Optionally, the step S6 includes:

[0060] S61. Constructing an adaptive learning model M for dynamically adjusting power grid anomaly detection parameters during power grid anomaly detection adapt The adaptive learning model is based on real-time power operation data D real (t) and historical power operation data D hist (t), combined with a two-layer generative adversarial network G opt The optimization result of (t) updates the parameter set Θ(t) of the anomaly detection model in real time:

[0061]

[0062] Among them, Θ(t) represents the parameters of the anomaly detection model at time t, η1 is the learning rate, and L det is the anomaly detection loss function;

[0063] S62. The adaptive learning model dynamically adjusts the threshold parameters of the anomaly detection model by continuously monitoring the power operation data of the power grid, so that the anomaly detection model adjusts the detection sensitivity according to the load change or abnormal mode of the power grid;

[0064] S63. When the grid operation status changes, the adaptive learning model automatically adjusts the feature weight W(t) in the anomaly detection model according to the real-time power operation data to optimize the response time:

[0065]

[0066] Among them, W(t) represents the feature weight matrix and β is the learning rate.

[0067] Optionally, the step S7 includes:

[0068] S71. Based on the topological information of the power grid, the second layer abnormal data generated by the two-layer generative adversarial network is combined with the topological data analysis method to locate the faulty nodes or abnormal lines in the power grid, and define the abnormal scoring function A of the node i (t):

[0069]

[0070] Among them, A i (t) represents node n i The anomaly score at time t, W ij (t) is node n i Its neighbor node n j The transmission weight between det is an abnormality scoring function, which is used to comprehensively evaluate the abnormal state of the node;

[0071] S72. By calculating the abnormal score A of each node i (t) Identify nodes n with high anomaly scores in the power grid fault , nodes with higher abnormality scores are regarded as potential fault nodes, and the identification rule of abnormal nodes is defined as:

[0072]

[0073] Among them, n fault Indicates the node that is located as a faulty node;

[0074] S73. Based on topological data analysis, identify the lines or areas where the fault may propagate by calculating the transmission weight changes between adjacent nodes and the power transmission flow between nodes;

[0075] S74. Combine topological data analysis with anomaly patterns generated by generative adversarial networks to ultimately locate faulty nodes and abnormal lines in the power grid and predict potential fault propagation paths.

[0076] The beneficial effects of the present invention are:

[0077] (1) The present invention embeds the topological structure information of the power grid into the graph neural network, which can automatically extract the potential relationships and abnormal patterns between the nodes of the power grid. The historical status and real-time data of each node are mapped to the high-dimensional space through the graph embedding mechanism of the graph neural network, which automatically captures the potential fault characteristics and complex abnormal propagation patterns in the power system. Compared with traditional single sensor or physical rule-based anomaly detection, the present invention can dynamically adjust the relationship weights between nodes, further improving the accuracy of anomaly detection, especially when dealing with complex multi-point failures and chain reactions.

[0078] (2) The present invention uses an adaptive learning model to dynamically adjust the parameters of the power grid anomaly detection model in real time. By combining real-time data with historical data, the detection model can automatically adapt to changes in the power grid operating conditions and adjust the parameters, weights and thresholds of the detection model. This can ensure that when the power grid load changes or emergencies occur, the system can respond in real time and automatically optimize the detection sensitivity and accuracy, thereby reducing the need for manual intervention and improving the efficiency and accuracy of fault detection.

[0079] (3) The present invention uses a two-layer generative adversarial network to generate more complex and realistic abnormal data for training the power grid anomaly detection model. The first generator and discriminator generate primary abnormal data, and the second generator and discriminator further optimize the abnormal data to make the generated data more consistent with the actual power grid failure mode, multi-point failure and chain reaction simulation, combined with topological data analysis, can identify potential fault nodes or abnormal lines in the power grid in advance, thereby improving the accuracy of fault prediction and location.

[0080] (4) The present invention accurately locates faulty nodes or abnormal lines in the power grid through topological data analysis and abnormal patterns generated by generative adversarial networks. By utilizing the abnormal scores of nodes and the weight changes of neighboring nodes, the system can quickly identify fault points in the power grid and predict the paths along which the faults may propagate, thereby greatly improving the efficiency of power grid fault handling, shortening the time for fault location and repair, and helping to improve the overall operational stability and safety of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0082] Figure 1 A flowchart of a power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis proposed by the present invention;

[0083] Figure 2 A schematic diagram of positioning power grid fault nodes and abnormal lines in a power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis proposed by the present invention. DETAILED DESCRIPTION

[0084] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0085] refer to Figure 1-2 , a power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis, comprising the following steps:

[0086] S1. Collect voltage, current and power from different nodes of the power grid through multiple sensors, and transmit the data to the data processing module to form power operation data;

[0087] S2. Construct a topological network model of the power grid based on the physical structure of the power grid. The nodes of the topological network model represent power plants, substations and key load centers, and the edges of the topological network model represent the power transmission relationship between nodes.

[0088] S3. Using the first generator to generate abnormal data in the simulated power grid based on the power operation data; using the first discriminator to discriminate the real power grid data and the generated abnormal data, and optimizing the abnormal data generated by the first generator;

[0089] S4. Based on S3, the power operation data and the abnormal data generated by the first layer are combined to generate abnormal data distribution, and the second discriminator is used to discriminate the difference between the abnormal data generated by the second layer and the real multi-source data;

[0090] S5. The power operation data and topological structure characteristics of the power grid are embedded into the anomaly detection model through a graph convolutional network, and anomaly detection is performed by combining real-time data and historical data. The anomaly pattern generated by the two-layer generative adversarial network is used to optimize the accuracy of power grid anomaly detection.

[0091] S6. Dynamically adjust the power grid anomaly detection parameters through the adaptive learning model, so that the anomaly detection model is continuously optimized as the power grid operation status changes;

[0092] S7. Based on the topology information and the abnormal data generated by the generative adversarial network, combined with the topology data analysis, locate the faulty nodes or abnormal lines in the power grid.

[0093] In this implementation, step S1 includes:

[0094] S11. Voltage sensors, current sensors and power sensors are deployed at multiple nodes of the power grid. The voltage sensors are used to collect voltage, the current sensors are used to collect current, and the power sensors are used to collect power. The nodes include power plants, substations and key load centers.

[0095] S12. Use data acquisition equipment to collect the real-time voltage data V of each node i (t), current data I i (t) and power data P i (t) continuously transmit to the power grid data processing module;

[0096] S13. According to the data transmission protocol, the power grid data processing module receives and stores all collected power operation data to form power operation data containing multiple time series:

[0097] D={(V i (t),I i (t),P i (t))|i=1,2,...,n; t=t0,t1,...,t k};

[0098] Where n represents the total number of acquisition nodes, and t is a time parameter used to represent the timestamp of the data, t0, t1, ..., t k is the data collection time at different time points.

[0099] In this embodiment, step S2 includes:

[0100] S21. Define the power grid topology network model:

[0101] G = (N, E);

[0102] In the power grid topology network model, each node n is combined with the historical status of the node and the power grid operation characteristics. i ∈N state information is represented by S i (t):

[0103]

[0104] Among them, g k is a function of different historical states, used to extract historical data features, α k is the adaptive weight, f hist A nonlinear mapping function representing the historical state;

[0105] S22. In the graph neural network embedding of the power grid topology network model, the historical status and real-time power operation data of each node are mapped to a high-dimensional space. Through the dynamic update mechanism between nodes, the embedding of the node is expressed as:

[0106]

[0107] Among them, H i (t) indicates node n i The high-dimensional embedding vector at time t, W embed and W neigh are the weight matrices of the node itself and its neighbor nodes, Represents node n i The set of neighbor nodes, σ is the activation function;

[0108] S23. Using graph neural network to dynamically adjust the weight relationship between edges in the power grid topology network model to describe node n i With node n j Power transfer capability and abnormal pattern detection between:

[0109]

[0110] Among them, W ij (t) represents node n i With node n j The weights between them change dynamically with time t, H i (t)·H j (t) is the vector inner product between nodes, which is used to calculate the similarity between nodes;

[0111] S24. Combine power operation data and power grid topology network model to automatically extract potential abnormal features in the power grid through graph neural network training:

[0112]

[0113] in, Represents node n i The feature vector at layer l, W (l) and are the weight matrices of the layer respectively.

[0114] In this implementation, step S3 includes:

[0115] S31. Based on the multi-source data set of the power grid and the topological embedding features of the power grid topology network model, the first generator G1 is constructed, and the historical power operation data, the real-time collected power operation data and the topological structure embedding features are used as inputs, and the abnormal data D in the simulated power grid is generated by the generator in the generative adversarial network:

[0116] D gen =G1(D real ,T embed );

[0117] Among them, D real Represents the actual power operation data, Tembed Represents the topological structure embedding feature, the generated D gen Abnormal data is used to simulate abnormal patterns of power grids with different fault types;

[0118] S32. The first discriminator D1 receives the real abnormal data D real and the abnormal data D generated by the first generator G1 gen The goal of the first discriminator is to distinguish the generated abnormal data from the real abnormal data. The loss function is defined as:

[0119]

[0120] in, represents the loss function of the first discriminator, is the prediction accuracy of the real abnormal data, is the probability of misjudging the generated abnormal data;

[0121] S33. The goal of the first generator G1 is to maximize the error discrimination rate of the first discriminator D1, even if the generated abnormal data D gen Close to the real abnormal data D real , the first generator loss function is defined as:

[0122]

[0123] S34. During the training process of the generative adversarial network, the first generator G1 and the first discriminator D1 are iteratively trained until the abnormal data D generated by the first generator gen Compared with the real abnormal data D real Difficult to distinguish.

[0124] In this embodiment, step S4 includes:

[0125] S41. In the generated abnormal data D gen Based on the above, a distribution model of abnormal data is generated by combining power operation data;

[0126] S42. Based on the distribution model, the second generator G2 is used to further generate the second layer of abnormal data reflecting different fault types and chain reactions. The second generator aims to combine the abnormal data D generated by the first layer gen and multi-source real data D real Optimize abnormal data;

[0127] S43. Use the second discriminator D2 to analyze the abnormal data of the second layer With multi-source real data D real The goal of the second discriminator is to distinguish the difference between complex abnormal data and real data:

[0128]

[0129] S44. The goal of the second generator G2 is to maximize the misjudgment rate of the second discriminator D2 and generate complex and difficult to distinguish abnormal data. The loss function of the second generator is expressed as:

[0130]

[0131] In this implementation, step S5 includes:

[0132] S51. Combine the power operation data of the power grid with the topological structure characteristics T embed By embedding the graph convolutional network into the anomaly detection model, the embedding of each node is expressed as:

[0133]

[0134] in, For node n i The embedded features in the l+1th layer of the graph convolutional network, c ij is node n i With node n j The normalization coefficient between ;

[0135] S52. The generator and discriminator of the two-layer generative adversarial network are used to optimize the power grid anomaly detection model. The first generator and the first discriminator are used to generate primary anomaly data and perform adversarial training with real data. The second generator and the second discriminator are used to generate more complex anomaly patterns and perform adversarial optimization. The optimized generative adversarial network is represented by G opt .

[0136] In this implementation, step S6 includes:

[0137] S61. Constructing an adaptive learning model M for dynamically adjusting power grid anomaly detection parameters during power grid anomaly detection adapt , the adaptive learning model is based on real-time power operation data D real (t) and historical power operation data D hist (t), combined with a two-layer generative adversarial network G opt The optimization result of (t) updates the parameter set Θ(t) of the anomaly detection model in real time:

[0138]

[0139] Among them, Θ(t) represents the parameters of the anomaly detection model at time t, η1 is the learning rate, and L det is the anomaly detection loss function;

[0140] S62. The adaptive learning model dynamically adjusts the threshold parameters of the anomaly detection model by continuously monitoring the power operation data of the power grid, so that the anomaly detection model adjusts the detection sensitivity according to the load change or abnormal mode of the power grid;

[0141] S63. When the grid operation status changes, the adaptive learning model automatically adjusts the feature weight W(t) in the anomaly detection model according to the real-time power operation data to optimize the response time:

[0142]

[0143] Among them, W(t) represents the feature weight matrix and β is the learning rate.

[0144] In this implementation, step S7 includes:

[0145] S71. Based on the topological information of the power grid, the second layer abnormal data generated by the two-layer generative adversarial network is combined with the topological data analysis method to locate the faulty nodes or abnormal lines in the power grid, and define the abnormal scoring function A of the node i (t):

[0146]

[0147] Among them, A i (t) represents node n i The anomaly score at time t, W ij (t) is node n i Its neighbor node n j The transmission weight between det is an abnormality scoring function, which is used to comprehensively evaluate the abnormal state of the node;

[0148] S72. By calculating the abnormal score A of each node i (t) Identify nodes n with high anomaly scores in the power grid fault , nodes with higher abnormality scores are regarded as potential fault nodes, and the identification rule of abnormal nodes is defined as:

[0149]

[0150] Among them, n fault Indicates the node that is located as a faulty node;

[0151] S73. Based on topological data analysis, identify the lines or areas where the fault may propagate by calculating the transmission weight changes between adjacent nodes and the power transmission flow between nodes;

[0152] S74. Combine topological data analysis with anomaly patterns generated by generative adversarial networks to ultimately locate faulty nodes and abnormal lines in the power grid and predict potential fault propagation paths.

[0153] Embodiment 1:

[0154] In Example 1, the present invention is applied to the abnormality detection and diagnosis process of a real power grid to verify the feasibility and effectiveness of the method of the present invention. In the simulation scenario, the topology of the power grid is complex and covers multiple key nodes of the urban backbone network, including power plants, substations and load centers. The power grid experiences abnormal fluctuations during a summer peak load period. Several typical faults are selected for detailed simulation to demonstrate the entire process of fault detection and location.

[0155] At 14:35 on July 15, 2024, a sudden abnormal voltage fluctuation was detected in the southern area of ​​the city power grid. The system collected data from the node located in the power plant "Plant_A" in real time through monitoring. Its voltage data suddenly dropped from the normal value of 230V to 190V, accompanied by abnormal current fluctuations. At the same time, the substation node "Substation_12" in the area where the power plant belongs was overloaded, and the current rose sharply from the normal value of 1200A to 1600A, causing a chain reaction with the load center "LoadCenter_C" in the northern area. The traditional system failed to capture the potential risks of the chain reaction in time until a large-scale power outage occurred.

[0156] In the present invention, a power grid anomaly detection method combining graph neural network and generative adversarial network is used to perform anomaly detection and diagnosis in complex scenarios.

[0157] First, the voltage, current and power data from multiple nodes are collected in real time by deploying sensors at each node. The specific data is shown in Table 1 below:

[0158] Table 1 The whole process of anomaly detection

[0159]

[0160] Based on real-time data, the model quickly generates anomaly scores by combining the power operation data between nodes with topological embedding features through graph neural networks. On the node "Plant_A", the anomaly score reached 0.85 at 14:36, exceeding the 0.75 threshold set by the system. The score is calculated through the hierarchical embedding mechanism of the graph convolutional network. At this time, the system generates an anomaly warning, warning that the "Plant_A" node may have a risk of power failure, and the chain reaction may affect the surrounding nodes.

[0161] The first generator of the generative adversarial network generates simulated abnormal data based on multi-source data sets. The comparison between the abnormal data and the real data is shown in Table 2:

[0162] Table 2. Simulation process of generating abnormal data

[0163]

[0164]

[0165] After generating abnormal data, the system optimized the anomaly detection model using the second generator. By comparing the real data with the simulated abnormal data, the system further confirmed the characteristics of the faulty node. For the "Plant_A" node, the system detected that the fluctuations of its voltage and current exceeded the normal range. The simulated data anomaly score increased to 0.90, further verifying the possibility of a node failure.

[0166] After the abnormal voltage fluctuation was identified, the system further analyzed the "Substation_12" substation and the "LoadCenter_C" load center through the graph neural network. The abnormal scores showed that these two nodes might be affected by the chain reaction of the "Plant_A" failure. By calculating the weight changes of the transmission lines, the system confirmed that the power transmission line from "Substation_12" to "LoadCenter_C" had a risk of fault propagation. The score of the abnormal line was 0.85, which exceeded the 0.8 threshold set by the system.

[0167] Based on this, the system generated a detailed fault diagnosis report and provided the following repair suggestions:

[0168] Immediately stabilize the voltage at the "Plant_A" node, reduce the grid load, and prevent further current fluctuations.

[0169] By cutting off part of the power transmission line from "Substation_12" to "LoadCenter_C", the chain reaction can be prevented from spreading and the risk of power outage can be reduced.

[0170] Implement regional power rationing for LoadCenter_C to ensure power supply in other areas.

[0171] In order to verify the effectiveness of the present invention, a comparison is made with the traditional power grid anomaly detection method. The following Table 3 is the comparison data of the present invention and the traditional method in the fault detection process:

[0172] Table 3 Comparison between the method of the present invention and the traditional method

[0173] Comparison indicators Traditional methods Method of the present invention Fault identification time 20 minutes 5 minutes Fault location accuracy 80％ 95％ Response time 30 minutes 10 minutes Recovery time 2 hours 45 minutes Multi-point fault detection capability Low high

[0174] In this power grid simulation, 5,000 sets of real power grid fault data were used to train the generative adversarial network, and the parameters were dynamically adjusted through the adaptive learning model. The following Table 4 shows some of the training data:

[0175] Table 4 Simulation data display

[0176]

[0177] Through training with 5,000 sets of data, the method of the present invention has shown excellent accuracy and stability in dealing with multi-point faults and chain reaction faults. Compared with traditional methods, the present invention significantly shortens the fault identification and response time, and greatly improves the accuracy of fault detection and the stability of power grid operation.

[0178] Through the detailed simulation process of this embodiment, it can be concluded that the power grid anomaly detection and diagnosis method of the present invention effectively solves the shortcomings of traditional detection methods, especially in the detection and processing of multi-point faults and chain reaction faults. By combining graph neural network and generative adversarial network technology, the present invention can provide early warning of potential faults, accurately locate faulty nodes and provide intelligent repair suggestions, which greatly improves the safety and stability of power grid operation.

[0179] The present invention embeds the topological structure information of the power grid into the graph neural network, which can automatically extract the potential relationships and abnormal patterns between power grid nodes. The historical status and real-time data of each node are mapped to the high-dimensional space through the graph embedding mechanism of the graph neural network, and the potential fault characteristics and complex abnormal propagation patterns in the power system are automatically captured. Compared with traditional single sensor or physical rule-based anomaly detection, the present invention can dynamically adjust the relationship weights between nodes, further improving the accuracy of anomaly detection, especially when dealing with complex multi-point failures and chain reactions.

[0180] The present invention uses an adaptive learning model to dynamically adjust the parameters of the power grid anomaly detection model in real time. Through the combination of real-time data and historical data, the detection model can automatically adapt to changes in the power grid operation status and adjust the parameters, weights and thresholds of the detection model. This can ensure that when the power grid load changes or emergencies occur, the system can respond in real time, automatically optimize the detection sensitivity and accuracy, reduce the need for manual intervention, and improve the efficiency and accuracy of fault detection.

[0181] The present invention uses a two-layer generative adversarial network to generate more complex and realistic abnormal data for training the power grid anomaly detection model. The first generator and discriminator generate primary abnormal data, and the second generator and discriminator further optimize the abnormal data to make the generated data more consistent with the actual power grid failure mode, the simulation of multi-point failures and chain reactions, and the combination of topological data analysis can identify potential fault nodes or abnormal lines in the power grid in advance, thereby improving the accuracy of fault prediction and location.

[0182] The present invention accurately locates faulty nodes or abnormal lines in the power grid through topological data analysis and abnormal patterns generated by generative adversarial networks. By utilizing the abnormal scores of nodes and the weight changes of neighboring nodes, the system can quickly identify fault points in the power grid and predict the paths along which the faults may propagate, greatly improving the efficiency of power grid fault handling, shortening the time for fault location and repair, and helping to improve the overall operational stability and safety of the power grid.

[0183] 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 power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis, characterized in that: The steps include: S1. Collect voltage, current and power from different nodes of the power grid through multiple sensors, and transmit the data to the data processing module to form power operation data; S2. Construct a topological network model of the power grid based on the physical structure of the power grid. The nodes of the topological network model represent power plants, substations and key load centers, and the edges of the topological network model represent the power transmission relationship between nodes. S3. Using the first generator to generate abnormal data in the simulated power grid based on the power operation data; using the first discriminator to discriminate the real power grid data and the generated abnormal data, and optimizing the abnormal data generated by the first generator; S4. Based on S3, the power operation data and the abnormal data generated by the first layer are combined to generate abnormal data distribution, and the second discriminator is used to discriminate the difference between the abnormal data generated by the second layer and the real multi-source data; S5. The power operation data and topological structure characteristics of the power grid are embedded into the anomaly detection model through a graph convolutional network, and anomaly detection is performed by combining real-time data and historical data. The anomaly pattern generated by the two-layer generative adversarial network is used to optimize the accuracy of power grid anomaly detection. S6. Dynamically adjust the power grid anomaly detection parameters through the adaptive learning model, so that the anomaly detection model is continuously optimized as the power grid operation status changes; S7. Based on the topology information and the abnormal data generated by the generative adversarial network, combined with the topology data analysis, locate the faulty nodes or abnormal lines in the power grid.

2. A method for detecting and diagnosing power grid anomalies based on artificial intelligence and topological data analysis according to claim 1, characterized in that: The S1 step includes: S11. Voltage sensors, current sensors and power sensors are arranged at multiple nodes of the power grid, the voltage sensor is used to collect voltage, the current sensor is used to collect current, and the power sensor is used to collect power, the nodes include power plants, substations and key load centers; S12. Use data acquisition equipment to collect the real-time voltage data V of each node i (t), current data I i (t) and power data P i (t) continuously transmit to the power grid data processing module; S13. According to the data transmission protocol, the power grid data processing module receives and stores all collected power operation data to form power operation data containing multiple time series: D={(V i (t),I i (t),P i (t))|i=1,2,...,n;t=t0,t1,...,t k }; Where n represents the total number of acquisition nodes, and t is a time parameter used to represent the timestamp of the data, t0, t1, ..., t k is the data collection time at different time points.

3. The power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis according to claim 1 is characterized in that: The S2 step includes: S21. Define the power grid topology network model: G = (N, E); In the power grid topology network model, each node n is combined with the historical status of the node and the power grid operation characteristics. i ∈N state information is represented by S i (t): Among them, g k is a function of different historical states, used to extract historical data features, α k is the adaptive weight, f hist A nonlinear mapping function representing the historical state; S22. In the graph neural network embedding of the power grid topology network model, the historical status and real-time power operation data of each node are mapped to a high-dimensional space. Through the dynamic update mechanism between nodes, the embedding of the node is expressed as: Among them, H i (t) indicates node n i The high-dimensional embedding vector at time t, W embed and W neigh are the weight matrices of the node itself and its neighbor nodes, Represents node n i The set of neighbor nodes, σ is the activation function; S23. Using graph neural network to dynamically adjust the weight relationship between edges in the power grid topology network model to describe node n i With node n j Power transfer capability and abnormal pattern detection between: Among them, W ij (t) represents node n i With node n j The weights between them change dynamically with time t, H i (t)·H j (t) is the vector inner product between nodes, which is used to calculate the similarity between nodes; S24. Combine power operation data and power grid topology network model to automatically extract potential abnormal features in the power grid through graph neural network training: in, Represents node n i The feature vector at layer 1, W (l) and are the weight matrices of the layer respectively.

4. The power grid anomaly detection and diagnosis method based on artificial intelligence and topological data analysis according to claim 1 is characterized in that: The S3 step includes: S31. Construct a first generator G1 based on the multi-source data set of the power grid and the topological embedding features of the power grid topology network model, take the historical power operation data, the real-time collected power operation data and the topological structure embedding features as input, and generate abnormal data D in the simulated power grid through the generator in the generative adversarial network: D gen =G1(D real ,T embed ); Among them, D real Represents the real power operation data, T embed Represents the topological structure embedding feature, the generated D gen Abnormal data is used to simulate abnormal patterns of power grids with different fault types; S32. The first discriminator D1 receives the real abnormal data D real and the abnormal data D generated by the first generator G1 gen The goal of the first discriminator is to distinguish the generated abnormal data from the real abnormal data. The loss function is defined as: in, represents the loss function of the first discriminator, is the prediction accuracy of the real abnormal data, is the probability of misjudging the generated abnormal data; S33. The goal of the first generator G1 is to maximize the error discrimination rate of the first discriminator D1, even if the generated abnormal data D gen Close to the real abnormal data D real , the first generator loss function is defined as: S34. During the training process of the generative adversarial network, the first generator G1 and the first discriminator D1 are iteratively trained until the abnormal data D generated by the first generator gen Compared with the real abnormal data D real Difficult to distinguish.

5. The method for detecting and diagnosing power grid anomalies based on artificial intelligence and topological data analysis according to claim 1, characterized in that: The S4 step includes: S41. In the generated abnormal data D gen Based on the above, a distribution model of abnormal data is generated by combining power operation data; S42. Based on the distribution model, the second generator G2 is used to further generate the second layer of abnormal data reflecting different fault types and chain reactions. The second generator aims to combine the abnormal data D generated by the first layer gen and multi-source real data D real Optimize abnormal data; S43. Use the second discriminator D2 to analyze the abnormal data of the second layer With multi-source real data D real The goal of the second discriminator is to distinguish the difference between complex abnormal data and real data: S44. The goal of the second generator G2 is to maximize the misjudgment rate of the second discriminator D2 and generate complex and difficult to distinguish abnormal data. The loss function of the second generator is expressed as:

6. The method for detecting and diagnosing power grid anomalies based on artificial intelligence and topological data analysis according to claim 1, characterized in that: The S5 step includes: S51. Combine the power operation data of the power grid with the topological structure characteristics d By embedding the graph convolutional network into the anomaly detection model, the embedding of each node is expressed as: in, For node n i The embedded features in the 1+1 layer of the graph convolutional network, c ii is node n i With node n j The normalization coefficient between ; S52. The generator and discriminator of the two-layer generative adversarial network are used to optimize the power grid anomaly detection model. The first generator and the first discriminator are used to generate primary anomaly data and conduct adversarial training with real data. The second generator and the second discriminator are used to generate more complex anomaly patterns and conduct adversarial optimization. The optimized generative adversarial network is represented by G opt .

7. The method for detecting and diagnosing power grid anomalies based on artificial intelligence and topological data analysis according to claim 1, characterized in that: The S6 step includes: S61. Constructing an adaptive learning model M for dynamically adjusting power grid anomaly detection parameters during power grid anomaly detection adapt The adaptive learning model is based on real-time power operation data D real (t) and historical power operation data D hist (t), combined with a two-layer generative adversarial network G opt The optimization result of (t) updates the parameter set Θ(t) of the anomaly detection model in real time: Among them, Θ(t) represents the parameters of the anomaly detection model at time t, η1 is the learning rate, and L det is the anomaly detection loss function; S62. The adaptive learning model dynamically adjusts the threshold parameters of the anomaly detection model by continuously monitoring the power operation data of the power grid, so that the anomaly detection model adjusts the detection sensitivity according to the load change or abnormal mode of the power grid; S63. When the grid operation status changes, the adaptive learning model automatically adjusts the feature weight W(t) in the anomaly detection model according to the real-time power operation data to optimize the response time: Among them, W(t) represents the feature weight matrix and β is the learning rate.

8. The method for detecting and diagnosing power grid anomalies based on artificial intelligence and topological data analysis according to claim 1, characterized in that: The S7 step includes: S71. Based on the topological information of the power grid, the second layer abnormal data generated by the two-layer generative adversarial network is combined with the topological data analysis method to locate the faulty nodes or abnormal lines in the power grid, and define the abnormal scoring function A of the node i (t): Among them, A i (t) represents node n i The anomaly score at time t, W ij (t) is the node ni and its neighbor node n j The transmission weight between det is an abnormality scoring function, which is used to comprehensively evaluate the abnormal state of the node; S72. By calculating the abnormal score A of each node i (t) Identify nodes n with high anomaly scores in the power grid fault , nodes with higher abnormality scores are regarded as potential fault nodes, and the identification rule of abnormal nodes is defined as: Among them, n fault Indicates the node that is located as a faulty node; S73. Based on topological data analysis, identify the lines or areas where the fault may propagate by calculating the transmission weight changes between adjacent nodes and the power transmission flow between nodes; S74. Combine topological data analysis with anomaly patterns generated by generative adversarial networks to ultimately locate faulty nodes and abnormal lines in the power grid and predict potential fault propagation paths.

Images