Situation assessment method, system and equipment based on digital twin platform, and medium

Abstract

The invention discloses a situation assessment method, system and equipment based on a digital twin platform and a medium, and belongs to the technical field of power distribution network safety, and the method comprises the steps: collecting power distribution network GIS topological information and operation data, carrying out the preprocessing of the operation data, and obtaining basic data; constructing a graph model matrix by utilizing GIS topological information; combining the basic data with a physical structure framework of GIS topological information to construct a digital twin power distribution network model; combining the basic data with the graph model matrix to construct a multi-branch sub-model; data features are abnormal according to the multi-branch sub-model; and performing security situation assessment by using the fusion features according to the security situation level. According to the method, the problems that a single strategy is relied on, deep mining of problems is difficult, the labor cost is too high, and the real-time evaluation requirements of low time delay and high reliability cannot be met are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution network safety, and particularly relates to a situation assessment method, system, device and medium based on a digital twin platform. BACKGROUND

[0002] In recent years, with the rapid development of smart grid and Internet of Things technology, digital twin, as an advanced paradigm of coupling physical entities and virtual models, has been widely used in simulation and state monitoring of power systems. Domestic and foreign scholars have carried out operation analysis researches for digital twin in the fields of power generation, power transmission and power distribution. Through real-time collection of SCADA, PMU and sensor data, high-precision three-dimensional topology models are constructed, and real-time visualization technology is used to carry out equipment health diagnosis and fault warning. At the same time, the edge computing and cloud platform collaborative architecture gradually matures, so that the preprocessing, compression and transmission delay of massive operation data are effectively controlled, laying a foundation for online situation awareness of digital twin power distribution network. Methods based on machine learning and graph neural network (GNN) are also used for security situation assessment, which realizes the identification of typical fault modes by analyzing voltage, current and load time series signals.

[0003] However, the existing technology still has several deficiencies that need to be broken through: first, the fusion of multi-source heterogeneous data mainly depends on a single model or a flat fusion strategy, which is difficult to balance global topology information and local mutation characteristics; second, the extraction of abnormal data often stays in the threshold or the preliminary judgment based on rules, lacking deep mining of unexpected topology disturbance inside the digital twin model; third, the weight distribution of security situation assessment is fixed and the cost of manual adjustment is high, which cannot be adjusted adaptively according to the operating conditions; fourth, in the large-scale power distribution network scenario, the cloud algorithm is difficult to meet the real-time evaluation requirements of low delay and high reliability, and it is urgent to combine edge gateway for preprocessing and distributed deployment of models. SUMMARY

[0004] In view of the above problems, the present application is proposed.

[0005] Therefore, the technical problem solved by the present application is: how to realize the detection of abnormal data characteristics in the power distribution network and the accurate assessment of the security situation by a situation assessment method based on a digital twin platform.

[0006] To solve the above technical problems, the present application provides the following technical solutions: a situation assessment method based on a digital twin platform, comprising the following steps: collecting power distribution network GIS topological information and operation data, preprocessing the operation data to obtain basic data; combining the basic data with the physical structure framework of the GIS topological information to construct a digital twin power distribution network model; constructing a graph model matrix using the GIS topological information; combining the basic data with the graph model matrix to construct a multi-branch submodel; extracting first, second, and third abnormal data features of the digital twin power distribution network model according to the multi-branch submodel; fusing the first, second, and third abnormal data features to obtain abnormal data features; and referring to a safety situation level, performing safety situation assessment using the fused features.

[0007] As a preferred scheme of the situation assessment method based on the digital twin platform, the step of constructing a graph model matrix using the GIS topological information comprises: creating an adjacency matrix according to the physical structure framework; performing row vector normalization and superposition on the adjacency matrix; and obtaining the graph model matrix by performing row vector normalization on the processed adjacency matrix after trimming.

[0008] As a preferred scheme of the situation assessment method based on the digital twin platform, the step of constructing a digital twin power distribution network model comprises: determining the physical structure framework by defining the association relationship between device nodes and nodes in the GIS topological information; establishing a mapping relationship between the basic data and the physical structure framework, and combining the basic data with the physical structure framework to obtain the digital twin power distribution network model.

[0009] As a preferred scheme of the situation assessment method based on the digital twin platform, the step of constructing the multi-branch sub-model comprises: the multi-branch sub-model comprises a GCN global topology branch, a GAT local correlation branch and an LSTM time sequence dynamic branch; wherein the step of obtaining the GCN global topology branch comprises: constructing a feature matrix according to the running state of the device nodes in the physical structure framework; multiplying the graph model matrix and the feature matrix to obtain a multiplication result, and obtaining global features after the multiplication result is unified to the same dimension through a preset learning matrix; and obtaining the GCN global topology branch by sampling the global features; wherein the step of obtaining the GAT local correlation branch comprises: respectively mapping the device nodes according to the graph model matrix and the feature matrix to obtain corresponding mapping parameter features; obtaining the attention score of each device node by weighted fusion of the corresponding mapping parameter features; and obtaining the GAT local correlation branch after the attention score is spliced; wherein the step of obtaining the LSTM time sequence dynamic branch comprises: synchronizing the basic data according to a preset timestamp to obtain historical time sequence features; extracting the time sequence dynamic features of the device nodes by analyzing the change law between the historical time sequence features and the corresponding device nodes; and obtaining the LSTM time sequence dynamic branch by integrating the time sequence dynamic features. The beneficial effects of the preferred embodiment are that through the collaborative design of the GCN global topology branch, the GAT local correlation branch and the LSTM time sequence dynamic branch, the core dimensions of the spatial global topology structure, the spatial local node correlation and the time dimension dynamic change of the power distribution network are respectively focused on, the limitation that the traditional single support model can only capture single dimension features is broken, and the key information dimensions affecting the safety situation in the operation of the power distribution network are comprehensively covered.

[0010] As a preferred scheme of the situation assessment method based on the digital twin platform, wherein: the steps of extracting the first, second and third abnormal data features include: the first, second and third abnormal data features are extracted through the GCN global topology branch, the GAT local correlation branch and the LSTM time sequence dynamic branch in the multi-branch sub-model respectively; wherein, the step of extracting the first abnormal data feature from the GCN global topology branch is: projecting the feature matrix through the preset learning matrix to obtain a projection result; calculating a first difference value between the projection result and the global feature, discriminating the first difference value from a first error threshold to obtain the first abnormal data feature; when the first difference value is greater than the first error threshold, the global matrix is determined as the first abnormal data feature; when the first difference value is not greater than the first error threshold, the global matrix is eliminated; wherein, the step of extracting the second abnormal data feature from the GAT local correlation branch is: discriminating the attention score from a second error threshold to obtain the second abnormal data feature; when the attention score is greater than the second error threshold, the mapping parameter feature corresponding to the attention score is determined as the second abnormal data feature; when the attention score is not greater than the second error threshold, the mapping parameter feature corresponding to the attention score is eliminated; wherein, the step of extracting the third abnormal data feature from the LSTM time sequence dynamic branch is: calculating a third difference value between the historical time sequence feature and the current time sequence feature; discriminating the third difference value from a third error threshold to obtain the third abnormal data feature; when the third difference value is greater than the third error threshold, the time sequence dynamic feature corresponding to the historical time sequence feature is determined as the third abnormal data feature; when the third difference value is not greater than the third error threshold, the historical time sequence feature is eliminated. The beneficial effect of the preferred embodiment is to realize multi-dimensional accurate extraction of abnormal data features, through the division and cooperation of the GCN global topology branch, the GAT local correlation branch and the LSTM time sequence dynamic branch, covering the spatial global topology, spatial local node correlation and time dynamic change of the power distribution network, breaking the limitation of traditional abnormal feature extraction focusing on a single dimension, fully capturing the potential abnormalities of the power distribution network in different dimensions, and ensuring the integrity and comprehensiveness of abnormal feature extraction.

[0011] As a preferred scheme of the situation assessment method based on the digital twin platform, wherein: the step of obtaining the abnormal data features comprises: unifying the first abnormal data feature, the second abnormal data feature and the third abnormal data feature to the same dimension after standardization processing; and obtaining the abnormal data feature by weighted fusion of the processed first abnormal data feature, the second abnormal data feature and the third abnormal data feature. The beneficial effect of the preferred embodiment is that the feature dimension difference interference is eliminated by standardization processing. The first abnormal data feature, the second abnormal data feature and the third abnormal data feature may be derived from different analysis dimensions, and the numerical dimension and value range may be significantly different. If direct fusion is used, some features may be over-amplified or weakened due to dimension imbalance. Standardization processing can unify each type of abnormal feature to the same scale to ensure that each type of abnormal feature can participate in subsequent fusion under a fair evaluation benchmark, and avoid the interference of dimension difference on the fusion result.

[0012] As a preferred scheme of the situation assessment method based on the digital twin platform, wherein: the step of safety situation assessment comprises: setting a full connection network of L levels according to the safe operation demand of the power distribution network; inputting the abnormal data features into the full connection network, and calculating each device node corresponding to each abnormal data feature to obtain an unnormalized score; converting the unnormalized score into a probability distribution through an activation function; comparing the probability distribution with the L levels to determine the safety situation.

[0013] The application provides a situation assessment system based on a digital twin platform.

[0014] To solve the above technical problems, the application further provides the following technical scheme: a situation assessment system based on a digital twin platform, comprising: a data acquisition module that acquires power distribution network GIS topological information and operation data, pre-processes the operation data, and obtains basic data; a digital twin power distribution network model construction module that combines the basic data with a physical structure framework of the GIS topological information to construct a digital twin power distribution network model; a graph model matrix construction module that constructs a graph model matrix using the GIS topological information; a multi-branch sub-model construction module that combines the basic data with the graph model matrix to construct a multi-branch sub-model; a feature extraction module that extracts first abnormal data features, second abnormal data features and third abnormal data features of the digital twin power distribution network model according to the multi-branch sub-model; an abnormal data feature acquisition module that fuses the first abnormal data features, the second abnormal data features and the third abnormal data features to obtain abnormal data features; and a safety assessment module that refers to a safety situation level and performs safety situation assessment using the fused features.

[0015] The application provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of the situation assessment method based on the digital twin platform when executing the computer program.

[0016] The application provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of the situation assessment method based on the digital twin platform.

[0017] The application has the following beneficial effects: the data quality is guaranteed through sensor data acquisition and edge gateway preprocessing, the digital twin power distribution network model is constructed in combination with GIS topology, the sudden scene is mapped and simulated, the global topology, local correlation and time sequence dynamic abnormal conditions are captured through the multi-branch sub-models of GCN global topology branch, GAT local correlation branch and LSTM time sequence dynamic branch, the integrity and sensitivity of abnormal feature extraction are improved through synergistic effect, and the false alarm and missed alarm conditions are effectively reduced, the high-risk features are highlighted through standardized unified feature dimension and weighted fusion, the digital twin power distribution network model is ensured to be stable and convergent, the full connection network evaluation is combined, and the situation judgment accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0019] Figure 1 The overall flowchart of the situation assessment method based on the digital twin platform provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor should be within the protection scope of the application.

[0021] Embodiment 1, refer to Figure 1 For the first embodiment of the application, the embodiment provides a situation assessment method based on a digital twin platform, comprising:

[0022] S100: Collect power distribution network GIS topology information and operation data, preprocess the operation data, and obtain basic data.

[0023] S200: Constructing a graph model matrix using GIS topological information.

[0024] S300: Combining the basic data with the physical structure framework of the GIS topological information to construct a digital twin power distribution network model.

[0025] S400: Combining the basic data with the graph model matrix to construct a multi-branch sub-model.

[0026] S500: Extracting first, second, and third abnormal data features of the digital twin power distribution network model according to the multi-branch sub-model.

[0027] S600: Fusing the first, second, and third abnormal data features to obtain abnormal data features.

[0028] S700: Referring to the safety situation level and using the fused features for safety situation assessment.

[0029] It should be noted that the existing technology still has several deficiencies that need to be addressed: first, the fusion of multi-source heterogeneous data often relies on a single model or a flat fusion strategy, making it difficult to balance global topological information and local mutation features; second, the extraction of abnormal data often stops at threshold or rule-based primary judgment, lacking deep mining of unexpected topological disturbances within the digital twin model; third, the weight distribution of safety situation assessment is fixed and the cost of manual adjustment is high, which cannot be adjusted adaptively according to the operating conditions; fourth, in large-scale power distribution network scenarios, cloud algorithms are difficult to meet the real-time evaluation needs of low latency and high reliability, and it is urgent to combine edge gateways for preprocessing and distributed deployment of models.

[0030] Therefore, to address the above problems of relying on a single strategy, difficulty in deep exploration of problems, high labor cost, and inability to meet the real-time evaluation needs of low latency and high reliability, a situation assessment method based on a digital twin platform is constructed through steps S100-S700. First, GIS topological information and operating data of the power distribution network are collected to preprocess the operating data and obtain basic data. Then, the basic data are combined with the physical structure framework of the GIS topological information to construct a digital twin power distribution network model, a graph model matrix is constructed using GIS topological information, and a multi-branch sub-model is constructed by combining the basic data with the graph model matrix. Next, first, second, and third abnormal data features of the digital twin power distribution network model are extracted according to the multi-branch sub-model, and abnormal data features are obtained by fusing the first, second, and third abnormal data features. Finally, safety situation assessment is performed using the fused features with reference to the safety situation level.

[0031] Embodiment 2, refer to Figure 1 For a second embodiment of the present application, a situation assessment method based on a digital twin platform is provided.

[0032] In an embodiment of the present application, in step S100, the GIS topological information and the operation data of the power distribution network are collected, the operation data is preprocessed, the basic data is obtained, and the following steps A1 and A2 are included:

[0033] A1: Collecting GIS topological information and operation data of the power distribution network.

[0034] Specifically, the operation data includes installing voltage and current sensors, smart meters, phasor measurement units and environmental sensors at the end of the feeder, the switch cabinet and the transformer side.

[0035] A2: Preprocessing the operation data to obtain the basic data.

[0036] Specifically, the preprocessing is performed on the edge gateway, the abnormal data is removed, the data after removing the abnormal data is smoothed by using the sliding window, and the data is normalized by Z-score after smoothing to ensure that the mean is 0 and the standard deviation is 1.

[0037] Further, the operation data is transmitted to the nearest edge gateway, the edge gateway writes the preprocessed sampling data to the time series database, and the data is uploaded to the digital twin platform through kafka.

[0038] It should be noted that voltage / current transformers (PT / CT), smart meters and PMUs are deployed at the end of the feeder, the switch cabinet and the transformer side, and meteorological information is synchronously collected at the environmental sensor to construct a multi-parameter coupled sampling system. Through multi-source information fusion, the deficiency of traditional single electrical quantity monitoring on external environment sensitivity is overcome. The fault discrimination accuracy is improved, the environmental temperature, humidity and other signals related to insulation aging are analyzed synchronously, the risk of insulation breakdown caused by overheating and humidity can be identified in advance. Reduce the time delay and bandwidth occupation: the edge gateway performs abnormal data removal, moving average and standardization on site, reduces the transmission pressure of the main station to the original big data, and improves the feasibility of real-time early warning response.

[0039] In an embodiment, the abnormal data can also be removed by the box plot method, which identifies and removes abnormal data outside the quartile range. In this scheme, the box plot method first sorts the operation data in the edge gateway, calculates the first quartile, the third quartile and the difference between the two, determines the abnormal boundary as the first quartile minus 1.5 times the interquartile range, the third quartile plus 1.5 times the interquartile range, and judges the data outside the boundary as abnormal and removes it. The processing process only depends on the calculation of basic statistics, meets the lightweight demand of the edge side, and effectively filters abnormal data.

[0040] In another implementation, the data smoothing process can also be implemented by an exponential moving average method to weight and smooth the time series data, weaken random fluctuations and reduce delay. In this scheme, the exponential moving average method sets a smoothing coefficient in the edge gateway, and recursively calculates the data after excluding the abnormal data in time sequence; without storing complete sliding window data, only the last time result needs to be retained, saving edge storage resources, while quickly responding to data changes, and achieving smoothing processing.

[0041] In the embodiment of the present application, in step S200, the BIS topology information is used to construct a graph model matrix, including the following steps B1-B3:

[0042] B1: Create an adjacency matrix according to the physical structure framework.

[0043] Specifically, the dimension of the adjacency matrix is first determined. Since the physical structure framework includes N device nodes, an N-order square matrix A is created, and the rows and columns of the matrix correspond one-to-one to the device nodes sorted by unique ID in the physical structure framework, ensuring that the matrix index matches the node ID.

[0044] Further, through the above rules, the adjacency matrix A can completely quantify the correlation between the device nodes in the physical structure framework, and convert the topology information into matrix data recognizable by the graph model.

[0045] B2: Superimpose after row vector normalization of the adjacency matrix.

[0046] Specifically, for the N-order adjacency matrix generated in B1, the sum of all elements in each row is calculated. Then, each element in each row is divided by the sum of the elements in the row to complete the row vector normalization. This way, the neighborhood association weight sum of each device node is 1, avoiding the imbalance of association weights in subsequent feature calculation due to the difference in the number of associated devices of different nodes, and ensuring the fairness of topology association quantization.

[0047] Further, a Gaussian noise matrix with the same size as the normalized adjacency matrix is generated, and the elements of the matrix are used to simulate common communication jitter, short-time short-circuit and other sudden dynamic disturbances in the actual operation of the power distribution network. The Gaussian noise matrix is superimposed on the normalized adjacency matrix according to a predetermined proportion to obtain an initial enhanced matrix. After superimposition, non-negative clipping is performed on the initial enhanced matrix to adjust all elements less than 0 in the matrix to 0, avoiding negative association that does not conform to the actual logic of the power distribution network, and ensuring that the matrix can accurately reflect the topology association characteristics.

[0048] In an implementation manner, the adjacency matrix row vector normalization can also be realized by column vector normalization to make the total weight of each device node associated with 1, ensuring the fairness of the associated receiving end. In this scheme, the column vector normalization is performed on the N-order adjacency matrix generated by B1, the total sum of all elements in each column is calculated, and then each element in each column is divided by the total sum of the elements in the column, thereby completing the normalization. This way can make the total weight of the association from other nodes to each node as an associated object be 1, avoid the weight imbalance caused by the difference in the number of associations of different nodes, adapt to the scenario in which the fairness of the node receiving association in the power distribution network needs to be emphasized, and also ensure the rationality of the topology association quantization.

[0049] In another implementation manner, the adjacency matrix row vector normalization can also be realized by L2 normalization to suppress the influence of extreme association weight and ensure the stability of the weight distribution. In this scheme, the L2 normalization is performed on the N-order adjacency matrix of B1, the square sum of all elements in each row is calculated, the square root of the square sum is obtained to obtain the L2 norm of the row, and then each element in each row is divided by the L2 norm of the row. This way can avoid the suppression of extreme large values on other weights in the original row summation normalization, adapt to the scenario in which a few strong associated nodes exist in the power distribution network, ensure that the weight of the weak associated node is not diluted too much, and maintain the stability of the topology association quantization.

[0050] B3: After the processed adjacency matrix is clipped, the row vector normalization processing is performed to obtain the graph model matrix.

[0051] It should be noted that the clipping is a secondary range constraint on the initial enhanced matrix obtained in step B2, and the values of all elements in the matrix are limited in the interval [0, 1]. If there is an element greater than 1 caused by noise superposition, it is adjusted to 1.

[0052] Further, after the clipping is completed, the row vector normalization processing is performed on the matrix again: the total sum of the elements in each row of the clipped matrix is calculated, and each element in each row is divided by the total sum of the row, so that the sum of the elements in each row is normalized to 1 again.

[0053] It should be noted that after the above clipping and secondary normalization processing, the final obtained matrix is the graph model matrix A ′ .

[0054] In the embodiment of the application, in step S300, the basic data is combined with the physical structure framework of the GIS topology information to construct a digital twin power distribution network model, including the following steps B1 and B2:

[0055] C1: The physical structure framework is determined by defining the association relationship between the device nodes and the nodes in the GIS topology information.

[0056] Specifically, the device nodes in the GIS topology information include high-voltage switches, distribution transformers, feeder branch ends, switch cabinets, and reactive power compensation devices in the power distribution network. Each device node is assigned a unique ID, such as T-101 for a transformer with ID 101 or S-203 for a switch with ID 203. The physical attributes of each node, such as transformer capacity, switch rated current, and line type, are also recorded.

[0057] Further, for device nodes directly connected by transmission lines, such as switches and transformers, or feeder ends and switch cabinets, a physical connection association is defined, represented by an undirected edge. For nodes with control logic, such as the linkage relationship between a main switch and a backup switch, a control association is defined, represented by a directed edge with a control direction, such as the main switch pointing to the backup switch. The association between nodes is defined according to the physical connection method.

[0058] Further, by defining, the GIS topology information is converted into structured data containing N device nodes and the association between nodes, forming a physical structure framework with nodes as the core and association as the link. This framework can intuitively reflect the physical layout and logical connection of the power distribution network, providing clear spatial coordinates and association benchmarks for subsequent mapping with basic data.

[0059] where N is the total number of power distribution network devices.

[0060] C2: Establish a mapping relationship between the basic data and the physical structure framework, and combine the basic data with the physical structure framework to obtain a digital twin power distribution network model.

[0061] Specifically, the mapping relationship between the basic data and the physical structure framework is achieved through the unique ID of the device node. The pre-processed real-time operating data such as voltage, current, and power are matched to the corresponding device nodes in the physical structure framework according to the collection source, and a timestamp is added to each data to ensure the time consistency of the data and the node state.

[0062] Further, through the real-time synchronization mechanism of the edge gateway, the mapped basic data is dynamically written into the physical structure framework: when the operating data of the device node is updated, the state parameters of the corresponding node in the physical structure framework are updated synchronously; at the same time, based on the association between nodes, the derived state of the associated nodes is calculated and updated.

[0063] It should be noted that the final digital twin power distribution network model not only retains the topological association logic of the physical structure framework, but also contains the real-time operating state of each node, which can dynamically reproduce the real-time operating conditions of the power distribution network, providing a virtual simulation carrier highly consistent with the physical system for subsequent graph model matrix construction and multi-branch sub-model analysis.

[0064] In the embodiment of the present application, in step S400, the basic data is combined with the graph model matrix to construct a multi-branch sub-model, including the following steps D1-D3:

[0065] D1: Obtain a GCN global topology branch, including the following steps D11-D13:

[0066] D11: Construct a feature matrix according to the running state of the device nodes in the physical structure framework.

[0067] Specifically, the number of rows of the feature matrix X corresponds to the N device nodes in the physical structure framework, and the column number M is equal to the number of running state features of each device node.

[0068] Further, when constructing, the preprocessed basic data is filled according to the device node unique ID and the feature type classification, for example, the element of the i-th row and the j-th column of the matrix is assigned as the real-time value of the j-th running state feature of the node.

[0069] In an implementable manner, filling and constructing the feature matrix according to the device node unique ID and the feature type classification can also be achieved by filling and constructing a function-oriented feature matrix according to the device function partition and the feature priority order, so that the node features of the same function are concentrated. In this scheme, the N device nodes are first divided into 3-5 partitions according to the function, and the number of nodes in each partition is determined; the column number M is still equal to the number of single-node features, but is sorted according to the importance of the features to the function. The preprocessed basic data is first classified according to the function partition, then the node ID is corresponded to the row number in the partition, and the real-time values of the features are filled in the order of priority, for example, the first two columns of the power transmission line partition node row are filled with voltage and current, so as to ensure that the node feature structures of the same function are consistent, facilitating subsequent analysis of the topology correlation according to the function dimension.

[0070] In another implementable manner, filling and constructing the feature matrix according to the device node unique ID and the feature type classification can also be achieved by filling and constructing a feature matrix containing time sequence information through time sequence window slicing and node feature splicing, so as to retain the time correlation of the node state. In this scheme, the number of rows is still N device node numbers, and the column number M is expanded to the product of the number of single-node features and the length of the time sequence window. The preprocessed basic data is grouped according to the node ID, and the running feature values of the last 5 time points of each node are extracted and spliced in time sequence to fill the corresponding columns of the i-th row of the matrix, so that each row is associated with the node ID and contains time sequence information, adapting to the subsequent multi-branch feature fusion scene that needs to combine the time sequence.

[0071] D12: Multiply the graph model matrix and the feature matrix to obtain a multiplication result, and obtain the global feature after unifying the multiplication result to the same dimension through a preset learning matrix.

[0072] Specifically, the graph model data is the graph model matrix A obtained in step S200 ′, the feature matrix is D11, and X is constructed, and the dimension is N*M.

[0073] Further, the specific form of the multiplication of the enhanced graph matrix and the feature matrix and the unified preset learning matrix is:

[0074] Z=A′×X×W gcn ;

[0075] In the formula, A' is the enhanced graph matrix; X is the feature matrix; W gcn is the initialization projection matrix, that is, the preset learning matrix, which adopts the Xavier uniform distribution, and the preset learning matrix hidden dimension d is set to 64; Z is the multiplication result.

[0076] D13: Obtain the GCN global topology branch by sampling the global feature.

[0077] Specifically, the sampling result of the multiplication result unified in D11 is sampled, and the specific form is:

[0078]

[0079] In the formula, H gcn represents the GCN feature matrix obtained after the ReLU activation function processing; is the space where H gcn is located, which is a real number space, N represents the total number of device nodes in the physical structure framework, and d represents the dimension of the feature after being unified by the preset learning matrix.

[0080] It should be noted that, during training, the Dropout (regularization technique) is performed on H gcn to obtain the final output After the ReLU activation and down-sampling operation dropout, the feature representation of the GCN global topology branch is obtained.

[0081] D2: Obtain the GAT local correlation branch, including steps D21-D23:

[0082] D21: According to the graph model matrix and the feature matrix, respectively map the device nodes to obtain the corresponding mapping parameter features.

[0083] Specifically, the feature projection of each device node is calculated in the mapping, and the specific form is:

[0084]

[0085] In the formula, is the initialization mapping matrix, represents the linear mapping matrix of the GAT branch; represents the feature vector of each node after mapping.

[0086] D22: Obtain the attention score of each device node by weighted fusion of corresponding mapping parameter features.

[0087] D23: Obtain the GAT local correlation branch after splicing the attention score.

[0088] Specifically, the feature vector of each node after mapping is processed through linear transformation and activation function to calculate the attention score, and the specific form is:

[0089]

[0090] In the formula, is a learnable vector of the attention mechanism; is the feature splicing of nodes i and j, and ‖ is splicing; LeakyReLU(·) is a ReLU activation function with a negative slope, which is used to calculate the attention score before splicing.

[0091] It should be noted that the neighboring nodes of each device node need to be normalized before splicing to ensure that the attention score is in the same dimension.

[0092] Further, the specific form of the normalization processing is:

[0093]

[0094] In the formula, α ij is the normalized attention weight, e ij is the non-normalized attention score of node i to j; is the neighbor set of node i.

[0095] Further, the feature representation of the GAT local correlation branch is obtained by weighting the device nodes after complete normalization processing, and the specific form is:

[0096]

[0097] In the formula, is the local aggregation feature of node i, h is the number of heads of multi-head attention; α ij is the normalized attention weight; is the neighbor set of node i; represents the feature vector of node j after mapping

[0098] D3: Obtain the LSTM time sequence dynamic branch, including steps D31-D33:

[0099] D31: Synchronize the basic data according to the preset timestamp to obtain the historical time sequence features.

[0100] Specifically, according to the pre-processed basic data according to step S100, the preset time stamp interval is determined according to the sampling frequency of the power distribution network, and the uniformity of the time dimension is ensured.

[0101] Further, the LSTM network configuration is performed, the input dimension F, the single-layer LSTM network with the hidden state dimension d = 64; the running data of the latest k + 1 continuous time stamps of each device node is intercepted to form the historical time sequence feature sequence of the node. For example, the historical time sequence feature sequence of node i is represented as {x i,t-k ,x i,t-k+1 ,…,x i,t}, wherein k = 9, and the specific form of the sequence-organized historical time sequence feature is:

[0102]

[0103] In the formula, is the sequence-organized historical time sequence feature, k is the time stamp, and F is the dimension.

[0104] D32: By analyzing the change rule between the historical time sequence feature and the corresponding device node, the time sequence dynamic feature of the device node is extracted.

[0105] Specifically, according to the forward propagation phenomenon of the historical time sequence feature, the time sequence dynamic features of time steps τ = t-k to t are extracted, and the specific form is:

[0106] (h i,τ ,c i,τ )=LSTMCell(x i,τ ,h i,τ-1 ,c i,τ-1 ).;

[0107] In the formula, x i,τ is the input feature vector of the device node i at the time step τ; h i,τ-1 is the hidden state of the device node i at the time step τ-1; c i,τ-1 is the cell state of the device node i at the time step τ-1; h i,τ is the updated hidden state of the device node i at the time step τ; and c i,τ is the updated cell state of the device node i at the time step τ.

[0108] D33: The time sequence dynamic features are integrated to obtain the LSTM time sequence dynamic branch.

[0109] Specifically, the sequences with a length of k + 1 of all N nodes are input into the LSTM interface to obtain the batch

[0110] In an implementation manner, the LSTM interface of PyTorch / TensorFlow can be called to integrate the time sequence dynamic feature to obtain the LSTM time sequence dynamic branch. In this scheme, the time sequence of N nodes with a length of T is first arranged, an LSTM layer is defined in the Keras framework, the arranged batch data is input into the LSTM layer, the time sequence dynamic feature output of each node is obtained through forward propagation calculation, and then all node features are summarized to complete the construction of the LSTM time sequence dynamic branch, which meets the batch processing and feature integration requirements of the original scheme.

[0111] In another implementation manner, the LSTM interface of PyTorch / TensorFlow can be called to integrate the time sequence dynamic feature to obtain the LSTM time sequence dynamic branch. In this scheme, the time sequence of N nodes with a length of T is first arranged, an LSTM layer is defined in the Keras framework, the arranged batch data is input into the LSTM layer, the time sequence dynamic feature output of each node is obtained through forward propagation calculation, and then all node features are summarized to complete the construction of the LSTM time sequence dynamic branch, which meets the batch processing and feature integration requirements of the original scheme.

[0112] In the embodiment of the application, in step S500, the first abnormal data feature, the second abnormal data feature and the third abnormal data feature of the digital twin power distribution network model are extracted according to the multi-branch sub-model, which includes the following steps E1-E3:

[0113] E1: The first abnormal data feature is extracted from the GCN global topology branch, including the following steps E11-E12:

[0114] E11: The feature matrix is projected through a preset learning matrix to obtain a projection result.

[0115] Specifically, the original feature X is projected into a d-dimensional space by using the enhanced adjacency matrix A' to obtain a global topology perception representation H. The reconstruction error ||H i -X i W gcn ||Judge which nodes deviate significantly in the global network structure, and the threshold τ gcn According to the average error, the depth feature representation of the topological abnormal node is collected as the abnormal feature output of the GCN branch.

[0116] Further, the adjacency matrix A is added with a self-loop and parallel normalized to obtain A same-size Gaussian noise matrix E is generated, and the elements thereof are subject to (recommended σ = 0.05) is superimposed to the original graph with a proportion α = 0.05 The enhanced graph matrix is obtained, and the specific form is as follows:

[0117]

[0118] In the formula, A' is the enhanced graph matrix, A is the preprocessed adjacency matrix; α is a preset proportion, which is set to 0.05 in the scheme; E is a Gaussian noise matrix with the same size as A.

[0119] Further, the weight matrix W is used to calculate the global of each node, and the specific form is as follows:

[0120]

[0121] In the formula, h i is the global topological correlation feature vector output by the i-th device node in the power distribution network after graph convolution and activation; A' ij is the element in the i-th row and the j-th column of the enhanced graph matrix; X j is the original feature vector of the j-th device node; W gcn is a preset learning matrix of GCN.

[0122] E12: Calculate the first difference value between the projection result and the global feature, and distinguish the first difference value from the first error threshold to obtain the first abnormal data feature.

[0123] Specifically, the form of calculating the first error is as follows:

[0124] e i = ||h i -(X i W gcn )||2;

[0125] In the formula, e i is the first error of the i-th device node; h i is the global topological correlation feature vector output by the i-th device node in the power distribution network after graph convolution and activation; X i is the original running feature vector of the i-th device node; W gcn is a preset learning matrix of GCN; ||·||2 is the square of the Euclidean norm.

[0126] Further, if e i exceeds the threshold τ gcn , it is marked as “topology anomaly”, and h i of the node is included in the abnormal feature set of the GCN branch. ​​

[0127] Collecting the abnormal feature representation H of all marked device nodes gcn , as the first abnormal feature.

[0128] It should be noted that τ gcn is set to twice the average error of the current device node.

[0129] E2: Extract the second abnormal data feature from the GAT local association branch, including steps E21-E22:

[0130] E21: Calculate the attention score of each device node.

[0131] Specifically, the GAT branch assigns a weight α ij to each neighbor j through an attention mechanism, reflecting the influence of the neighbor on node i; if the attention value α ij of the edge exceeds the threshold τ gat , it means that the neighbor abnormally dominates node i and is judged as abnormal.

[0132] Further, by linearly mapping each node feature to a d-dimensional space with the matrix , the attention score of each edge is calculated, and the specific form of calculating the attention score is:

[0133] e ij = LeakyReLU(a T [W gat x i ||W gat x j ]);

[0134] In the formula, e ij is the original attention score between device node i and node j; a T is a learnable vector of the attention mechanism; W gat is a learnable weight matrix of the GAT branch; x i and x j are the original feature vectors of device node i; || is the vector concatenation operation.

[0135] Further, the original attention score is processed to obtain the attention score, and the specific form is:

[0136]

[0137] In the formula, α ij is the normalized attention score of device node i to its neighbor node j; e ij is the original attention score of node i and node j; e ik is the original attention score of node i and its neighbor node k.

[0138] E21: obtaining the second abnormal data feature by distinguishing the attention score and the second error threshold.

[0139] Specifically, the second error threshold is set to τ gat = 0.8, and when α ij is greater than τ gat , it indicates that the neighbor "dominates" the node state, that is, the node i appears "local attention anomaly", and the corresponding h i is marked as H gat and is taken as the second abnormal data feature.

[0140] E3: extracting the abnormal data feature from the LSTM time sequence dynamic branch, including steps E31-E32:

[0141] E31: calculating the third difference value of the historical time sequence feature and the current time sequence feature.

[0142] Specifically, the specific form of calculating the third difference value is as follows:

[0143]

[0144] In the formula, ε i is the third error threshold of the ith equipment node at time t; x i,t is the actual running feature vector of the ith equipment node at time t; is the LSTM network predicted feature vector of the ith equipment node at time t; and ||·||2 is the Euclidean norm.

[0145] E32: obtaining the third abnormal data feature by distinguishing the third difference value and the third error threshold.

[0146] Specifically, the third error threshold is set to 1.5 times of the average error of all nodes, and when ε i > τ lstm , the corresponding h i of the equipment node is marked as H lstm and is taken as the third abnormal data feature.

[0147] In the embodiment of the application, in step S600, the first abnormal data feature, the second abnormal data feature and the third abnormal data feature are fused to obtain the abnormal data feature, including steps F1-F2:

[0148] F1: unifying the first abnormal data feature, the second abnormal data feature and the third abnormal data feature to a unified dimension.

[0149] Specifically, the first abnormal data feature, the second abnormal data feature and the third abnormal data feature are arranged side by side in the feature dimension, and the feature values of non-abnormal nodes are supplemented for each abnormal data feature based on the total equipment node number N in the power grid physical structure framework, so as to unify the node coverage range; then a learning matrix is introduced to calibrate the feature dimension.

[0150] It should be noted that in the preceding step, the dimension unification of the first abnormal data feature, the second abnormal data feature and the third abnormal data feature has been completed by the preset learning matrix.

[0151] F2: obtaining the abnormal data feature by weighted fusion of the first abnormal data feature, the second abnormal data feature and the third abnormal data feature.

[0152] Specifically, h gcn , h gat , h lstm are multiplied by vectors , and the attention weights {ω1, ω2, ω3} are obtained by point multiplication of each output and softmax normalization; the abnormal data feature is obtained by weighted summation, and the specific form is:

[0153] h fused = ω1h gcn + ω2h gat + ω3h lstm .

[0154] In the formula, ω1, ω2, ω3 are weights of the first abnormal data feature, the second abnormal data feature and the third abnormal data feature respectively, h gcn , h gat , h lstm are the first abnormal data feature, the second abnormal data feature and the third abnormal data feature respectively.

[0155] In the embodiment of the application, in step S700, the safety situation is evaluated by referring to the safety situation level and using the fusion feature, which includes the following steps G1-G3:

[0156] G1: inputting the abnormal data feature pair into the full connection network, calculating each abnormal data feature pair corresponding to the equipment node, and obtaining the non-normalized score, which includes the following steps G11-G12:

[0157] G11: calculating the fusion feature.

[0158] Specifically, the feature dimension unification is performed according to the F1 step in step S600, and the abnormal data feature matrix H stack is obtained, and the specific form is:

[0159] H stack = [hgcn ||h gat ||h lstm ];

[0160] wherein H stack is an abnormal data feature matrix; h gcn is a first abnormal data feature; h gat is a second abnormal data feature; and h lstm is a third abnormal data feature.

[0161] Further, by introducing a trainable vector, the normalized score of each abnormal data feature is calculated, and the specific form is:

[0162] s = H stack w;

[0163] wherein s is the normalized score of all nodes; H stack is an abnormal data feature matrix; and w is a trainable vector,

[0164] Further, the original score of the feature of each node is calculated, and for each path i∈{gcn, gat, lstm}, the dot product of each node feature h i and w is calculated in the batch dimension, and the specific form is:

[0165]

[0166] wherein s i is the average comprehensive score of the device node i; s gcn is the normalized factor of the average operation; N is the total number of samples participating in the average calculation; w gat is the transpose of the trainable weight vector; and h lstm is the feature vector of the nth node in the ith path, and the average value is taken to keep the score stable when the number of nodes changes.

[0167] Further, the normalization is the attention weight, and the three original scores {s i , s i , s i} are input into the activation function, and the specific form is:

[0168]

[0169] wherein ω i is the normalized weight of the target branch, satisfying ∑ gcn ω gat = 1; s lstm is the comprehensive score of the current branch; s (n) , s gcn , and s gat .is the comprehensive score of GCN global topology branch, GAT local correlation branch and LSTM time series dynamic branch respectively; τ is a temperature parameter, which controls the sharpness of the distribution, and is usually set to 0.5.

[0170] Further, for each node n, the three-way feature is weighted by weight to obtain the fusion feature:

[0171]

[0172] In the formula, H (n) is the fusion feature of the nth sample; and is the feature vector of the nth sample in the GCN global topology branch, the feature vector of the nth sample in the GAT local correlation branch and the feature vector of the nth sample in the LSTM time series dynamic branch respectively; ω gcn , ω gat and ω lstm are the weight coefficients of the GCN global topology branch feature, the weight coefficients of the GAT local correlation branch feature and the weight coefficients of the LSTM time series dynamic branch feature respectively.

[0173] It should be noted that the final fusion feature matrix H is obtained In the scenario of sudden failure, the GAT local attention branch is dominant, the local topology changes significantly, and the weight ω gat is automatically increased to enhance the sensitivity to local anomalies;

[0174] In the slow drift scenario, the LSTM time series dynamic branch is dominant, and ω lstm is increased to capture gradual risks;

[0175] When the global state is stable, the GCN global topology branch is dominant, and the overall network connectivity is evaluated.

[0176] G12: Calculate the unnormalized score according to the fusion feature.

[0177] Specifically, through forward inference, for each node n, the unnormalized score is calculated:

[0178]

[0179] In the formula, z (n) is the unnormalized score; H (n) is the fusion feature vector of the nth sample; W c is a trainable weight matrix; and b c is a trainable bias vector.

[0180] G2: Convert the unnormalized score to a probability distribution through an activation function.

[0181] Specifically, the unnormalized score is generated into each level probability distribution by Softmax, and the specific form is:

[0182]

[0183] In the formula, is the probability distribution of each level, z (n) is the unnormalized score.

[0184] G3: The probability distribution is compared with the four levels to determine the security situation,

[0185] Specifically, according to the safe operation requirement of the power distribution network, a full connection network including four levels of normal, early warning, serious, and failure is set up.

[0186] The four levels of normal, early warning, serious, and failure are compared and judged to obtain the security situation evaluation result.

[0187] In summary, the application guarantees data quality through sensor sampling and edge gateway preprocessing, constructs a digital twin power distribution network model combined with GIS topology, realizes mapping and simulation of sudden scenarios, captures global topology, local correlation, and time sequence dynamic abnormal conditions through multi-branch sub-models of GCN global topology branch, GAT local correlation branch, and LSTM time sequence dynamic branch, improves the integrity and sensitivity of abnormal feature extraction through synergistic effect, effectively reduces false positives and false negatives; through standardized unified feature dimension, weighted fusion of high-risk features, ensure the stable convergence of the digital twin power distribution network model, and improve the accuracy of situation judgment with the full connection network evaluation.

[0188] Embodiment 3 is a third embodiment of the application, which provides a situation evaluation system based on a digital twin platform, comprising.

[0189] The data acquisition module acquires GIS topology information and operation data of the power distribution network, preprocesses the operation data, and obtains basic data.

[0190] The graph model matrix construction module constructs a graph model matrix using GIS topology information.

[0191] The digital twin power distribution network model construction module combines the basic data and the physical structure framework of the GIS topology information to construct a digital twin power distribution network model.

[0192] The multi-branch sub-model construction module combines the basic data and the graph model matrix to construct a multi-branch sub-model.

[0193] The feature extraction module extracts the first, second, and third abnormal data features of the digital twin power distribution network model according to the multi-branch sub-model.​

[0194] The abnormal data feature acquisition module fuses the first abnormal data feature, the second abnormal data feature and the third abnormal data feature to acquire an abnormal data feature.

[0195] The security evaluation module refers to the security posture level and uses the fused feature to evaluate the security posture.

[0196] Embodiment 4 is the fourth embodiment of the present application, which is different from the first three embodiments in that: if the function is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solution of the present application or the part that contributes to the prior art or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0197] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logic functions, and can be specifically embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, apparatus, or device and execute the instructions, or in conjunction with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device, or in conjunction with these instruction execution systems, apparatuses, or devices.

[0198] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer readable medium can be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example via an optical scanner, then compiled, interpreted or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0199] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following techniques can be used to implement the hardware: discrete logic circuits having logic gates for implementing logic functions upon data signals, application specific integrated circuits having logic gates for logic functions, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and so on.

[0200] It should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, but not limit the present application, although the present application is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or equivalent replaced, without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of claims of the present application.

Claims

1. A situation assessment method based on a digital twin platform, characterized in that: include, Collect GIS topology information and operation data of the power distribution network, preprocess the operation data, and obtain basic data; Construct a graph model matrix using the GIS topology information; By combining the basic data with the physical structure framework of the GIS topology information, a digital twin power distribution network model is constructed. The basic data is combined with the graph model matrix to construct a multi-branch sub-model; The first, second, and third abnormal data features of the digital twin distribution network model are extracted based on the multi-branch sub-model. The first abnormal data feature, the second abnormal data feature, and the third abnormal data feature are fused together to obtain abnormal data features; Security situation assessment is performed using the fused features, with reference to the security situation level.

2. The situation assessment method based on a digital twin platform as described in claim 1, characterized in that: The steps for constructing a graph model matrix using the GIS topology information include: Create an adjacency matrix based on the physical structure framework; The adjacency matrices are then stacked after row vector normalization. The processed adjacency matrix is ​​pruned and then normalized using the row vectors to obtain the graph model matrix.

3. The situation assessment method based on a digital twin platform as described in claim 2, characterized in that: The steps for constructing a digital twin distribution network model include: The physical structure framework is determined by defining the relationships between device nodes and nodes in the GIS topology information; A mapping relationship is established between the basic data and the physical structure framework, and the basic data and the physical structure framework are combined to obtain the digital twin distribution network model.

4. The situation assessment method based on a digital twin platform as described in claim 3, characterized in that: The steps to construct a multi-branch sub-model include: The multi-branch sub-model includes the GCN global topology branch, the GAT local correlation branch, and the LSTM temporal dynamic branch; The steps for obtaining the GCN global topology branch are as follows: Based on the operating status of the device nodes in the physical structure framework, a feature matrix is ​​constructed; The graph model matrix is ​​multiplied by the feature matrix to obtain the multiplication result. The multiplication result is then unified to the same dimension through a preset learning matrix to obtain the global features. The global topology branch of the GCN is obtained by sampling the global features; The steps for obtaining the local association branch of GAT are as follows: The device nodes are mapped according to the graph model matrix and the feature matrix respectively to obtain the corresponding mapping parameter features; The attention score of each device node is obtained by weighted fusion of the corresponding mapping parameter features. The attention scores are concatenated to obtain the local correlation branch of GAT; The steps for obtaining the LSTM timing dynamic branch are as follows: The basic data is synchronized according to a preset timestamp to obtain historical time series characteristics; By analyzing the historical time-series characteristics and the corresponding changes between the device nodes, the time-series dynamic characteristics of the device nodes are extracted. The temporal dynamic features are integrated to obtain the LSTM temporal dynamic branch.

5. The situation assessment method based on a digital twin platform as described in claim 4, characterized in that: The steps for extracting the first, second, and third anomalous data features include: The first, second, and third abnormal data features are extracted through the GCN global topology branch, the GAT local correlation branch, and the LSTM temporal dynamic branch in the multi-branch sub-model, respectively. The step of extracting the first abnormal data feature from the GCN global topology branch is as follows: The feature matrix is ​​projected through the preset learning matrix to obtain the projection result; Calculate the first difference between the projection result and the global feature, and compare the first difference with a first error threshold to obtain the first abnormal data feature; When the first difference is greater than the first error threshold, the global matrix is ​​determined to be the first abnormal data feature; When the first difference is not greater than the first error threshold, the global matrix is ​​removed. The step of extracting the second abnormal data feature from the local correlation branch of GAT is as follows: The attention score is compared with a second error threshold to obtain the second abnormal data feature; When the attention score is greater than the second error threshold, the mapping parameter feature corresponding to the attention score is determined as the second abnormal data feature; When the attention score is not greater than the second error threshold, the mapping parameter feature corresponding to the attention score is removed. The step of extracting anomalous data features from the third time-series dynamic branch of the LSTM is as follows: Calculate the third difference between the historical time series features and the current time series features; The third difference is compared with the third error threshold to obtain the third abnormal data feature; When the third difference is greater than the third error threshold, the time-series dynamic feature corresponding to the historical time-series feature is determined as the third abnormal data feature; When the third difference is not greater than the third error threshold, the historical time series features are removed.

6. The situation assessment method based on a digital twin platform as described in claim 5, characterized in that: The steps for obtaining characteristics of abnormal data include: The first abnormal data feature, the second abnormal data feature, and the third abnormal data feature are standardized to the same dimension. The processed first abnormal data feature, the second abnormal data feature, and the third abnormal data feature are weighted and fused to obtain the abnormal data feature.

7. The situation assessment method based on a digital twin platform as described in claim 6, characterized in that: The steps involved in a security situation assessment include: Based on the requirements for safe operation of the distribution network, L levels of fully connected networks are established; The abnormal data features are input into the fully connected network, and the unnormalized score is calculated for each device node corresponding to the abnormal data feature. The unnormalized score is converted into a probability distribution using an activation function; The probability distribution is compared with the L levels to determine the security situation.

8. A situation assessment system based on a digital twin platform, employing the situation assessment method based on a digital twin platform as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module collects GIS topology information and operational data of the power distribution network, preprocesses the operational data, and obtains basic data. The graph model matrix construction module uses the GIS topology information to construct a graph model matrix. The digital twin distribution network model construction module combines the basic data with the physical structure framework of the GIS topology information to construct a digital twin distribution network model. The multi-branch sub-model construction module combines the basic data with the graph model matrix to construct a multi-branch sub-model. The feature extraction module extracts the first abnormal data feature, the second abnormal data feature, and the third abnormal data feature of the digital twin distribution network model based on the multi-branch sub-model. The abnormal data feature acquisition module fuses the first abnormal data feature, the second abnormal data feature, and the third abnormal data feature to obtain abnormal data features; The security assessment module performs a security situation assessment using the fused features, with reference to the security situation level.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the situation assessment method based on a digital twin platform as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the situation assessment method based on the digital twin platform as described in any one of claims 1 to 7.

Images