A power grid enterprise architecture operation deviation quantification evaluation method

Abstract

The present application relates to a kind of power grid enterprise architecture operation deviation quantification evaluation method, to solve the node semantic deviation identification and the problem of interpretable feedback under the multi-dimensional governance perspective in power grid enterprise architecture blueprint.The core scheme includes structured analysis blueprint model, generates the semantic anchor point set carrying governance metadata, utilizes anchor point perception subgraph sampling and semantic projection consistency constraint, trains the architecture embedding vector embedded with multi-perspective feature, and realizes deviation type identification and natural language backtracking attribution by decoupling significance analysis, finally automatically synchronizes alarm and triggers active governance process.The method realizes the accurate analysis and abnormal attribution of node role under the dimensions of business, application, data, technology and other dimensions in complex architecture, improves the transparency, automation degree and accuracy of architecture governance.

Description

Technical Field

[0001] This invention relates to the field of interpretable graph representation learning and architecture governance technology for power grid enterprise architecture, and in particular to a method for quantifying and evaluating deviations in the operation of power grid enterprise architecture. Background Technology

[0002] With the deepening of digital transformation, the complexity and heterogeneity of power grid enterprise architectures are becoming increasingly prominent. For digital modeling and governance of enterprise architectures, current mainstream technologies mostly employ graph embedding-based methods, mapping the relationships between business components, application systems, data, and technical components in the architecture blueprint into low-dimensional vectors to achieve automatic deviation detection, dependency path analysis, and structural optimization. These graph embedding technologies mainly include node2vec, heterogeneous graph convolutional neural networks (GCN), and multi-view embedding models that combine structural and attribute features. In recent years, with the development of deep learning, the structural representation capabilities of graph embeddings have been significantly improved, further supporting intelligent monitoring and operation and maintenance of enterprise architectures.

[0003] However, existing graph embedding methods generally suffer from insufficient interpretability. First, mainstream embedding models often operate as black boxes during the modeling process, resulting in low-dimensional vectors that lack a direct correspondence with governance semantics such as business, applications, data, and technology within the power grid enterprise architecture blueprint. This makes it difficult to clearly inform architects "which dimensions represent which governance elements" in practical governance scenarios, leading to deviation detection results that are hard for architecture governance personnel to understand and trust. Second, current embedding interpretations typically rely on external knowledge injection, post-interpretation models, or general attribution tools (such as SHAP / LIME). These external interpretation mechanisms are disconnected from the hierarchy, responsibilities, and compliance requirements of the digital enterprise architecture itself, making it impossible to trace deviation location results back to specific governance constraints, which is detrimental to standardized and traceable architecture remediation. Furthermore, existing methods often focus on overall performance and lack the ability to form a distinct coordinate system for different types of semantic units (such as business components, data entities, and technical components) in the embedding space. This makes it difficult to support the qualitative classification and location of node deviation causes during architecture governance, affecting the automation and transparency of actual governance. Summary of the Invention

[0004] This application provides a quantitative assessment method for the deviation of power grid enterprise architecture operation, aiming to solve the problem mentioned in the background art of the difficulty in qualitatively classifying and locating the deviation of power grid enterprise architecture operation.

[0005] This application provides a quantitative assessment method for measuring operational deviations of power grid enterprise architecture, specifically including: S1: Structure and analyze the core semantic units from the power grid enterprise architecture blueprint to generate a set of semantic anchor points carrying governance metadata; S2: Based on the set of semantic anchor points, perform anchor-aware heterogeneous neighborhood sampling operations to generate a sequence of local topological subgraphs with semantic traction characteristics; S3: Calculate the weighted semantic similarity index based on the local topological subgraph sequence, and construct the anchor point projection consistency constraint parameter set for constraining the vector space distribution; S4: Use the local topological subgraph sequence as the input of the graph structure fidelity channel, and combine it with the anchor point projection consistency constraint parameter set to construct a semantic regularization channel, and generate the original architecture embedding vector based on the semantic regularization channel; S5: Perform dimensionality partitioning on the original architecture embedding vector to obtain the decoupled architecture feature vector; S6: Based on the comparison results between the decoupled architecture feature vector and the preset requirements of the power grid enterprise architecture blueprint, identify abnormal patterns such as score missing, score redundancy and dimension mismatch, and generate architecture deviation type labels classified as missing, redundant and mismatched. S7: For abnormal nodes marked with architecture deviation type labels, reverse the search for the nearest neighbor semantic anchors of the same type, generate an interpretable backtracking chain of architecture deviation, and complete the quantitative assessment of the deviation degree of the power grid enterprise architecture operation based on the interpretable backtracking chain of architecture deviation.

[0006] This invention also provides a device for quantitatively assessing operational deviations of power grid enterprise architecture, employing the aforementioned method for quantitatively assessing operational deviations of power grid enterprise architecture, comprising: The semantic anchor generation module is used to structurally parse core semantic units from the power grid enterprise architecture blueprint and generate a set of semantic anchors carrying governance metadata. The local topological subgraph sequence generation module is used to perform anchor-aware heterogeneous neighborhood sampling operations based on the semantic anchor set to generate a local topological subgraph sequence with semantic traction characteristics. The constraint parameter set construction module is used to calculate a weighted semantic similarity index based on the local topological subgraph sequence and construct an anchor point projection consistency constraint parameter set for constraining the distribution of the constraint vector space. The original architecture embedding vector generation module is used to use the local topological subgraph sequence as input to the graph structure fidelity channel, and to construct a semantic regularization channel in combination with the anchor projection consistency constraint parameter set, and generate the original architecture embedding vector based on the semantic regularization channel. The decoupled architecture feature extraction module is used to perform dimensionality partitioning on the original architecture embedding vector to obtain the decoupled architecture feature vector. The deviation type identification module is used to identify abnormal patterns such as score missing, score redundancy and dimension mismatch based on the comparison results between the feature vector of the decoupled architecture and the preset requirements of the power grid enterprise architecture blueprint, and generate architecture deviation type labels classified as missing, redundant and mismatched. The interpretable backtrace chain generation module is used to reverse-search for the nearest neighbor semantic anchors of the same type for abnormal nodes marked as architectural deviation type labels, generate an interpretable backtrace chain of architectural deviation, and complete the quantitative assessment of the deviation degree of the power grid enterprise's architecture operation based on the interpretable backtrace chain of architectural deviation.

[0007] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the power grid enterprise architecture operation deviation quantification assessment method.

[0008] The present invention also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the steps of the aforementioned power grid enterprise architecture operation deviation quantification assessment method.

[0009] This application provides a quantitative assessment method for measuring operational deviations in power grid enterprise architecture, which has the following beneficial effects: This application significantly improves the accuracy and reliability of architecture consistency analysis by constructing a semantic anchor-guided graph embedding framework. It innovatively extracts four types of semantic anchors—business components, application services, data entities, and technical components—from the power grid enterprise architecture blueprint as the "origin of the coordinate space." Furthermore, it introduces an anchor-aware heterogeneous sampling strategy during the neighborhood sampling phase, ensuring that each type of node is always semantically guided by its corresponding governance dimension during the aggregation process, effectively avoiding cross-domain information confusion and interference from irrelevant paths. Further, by designing a dual-channel loss function, while preserving graph structure fidelity, it adds an "anchor projection consistency constraint," forcing the distribution intensity of node embedding vectors in a preset semantic subspace to be positively correlated with governance indicators such as the degree of responsibility matching, interface contract compliance, and lifecycle coupling to similar anchors. This makes the embedding process itself carry clear management logic, thereby achieving end-to-end alignment from the original topology to governance semantics, significantly improving the relevance and stability of deviation identification. Attached Figure Description

[0010] Figure 1 This is the main flowchart of a quantitative assessment method for operational deviations of power grid enterprise architecture.

[0011] Figure 2 This is a sub-flowchart of a method for quantifying and evaluating deviations in the operational structure of a power grid enterprise.

[0012] Figure 3 This is another sub-flowchart of a method for quantifying deviations in the operational structure of a power grid enterprise.

[0013] Figure 4 This is a schematic diagram of a device for quantifying and evaluating operational deviations from the architecture of a power grid enterprise.

[0014] Figure 5 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0015] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0017] like Figure 1 As shown, this application provides a method for quantifying and evaluating operational deviations of a power grid enterprise architecture, specifically including: S1: Structure and analyze the core semantic units from the power grid enterprise architecture blueprint to generate a set of semantic anchor points carrying governance metadata; S2: Based on the set of semantic anchor points, perform anchor-aware heterogeneous neighborhood sampling operations to generate a sequence of local topological subgraphs with semantic traction characteristics; S3: Calculate the weighted semantic similarity index based on the local topological subgraph sequence, and construct the anchor point projection consistency constraint parameter set for constraining the vector space distribution; S4: Use the local topological subgraph sequence as the input of the graph structure fidelity channel, and combine it with the anchor point projection consistency constraint parameter set to construct a semantic regularization channel, and generate the original architecture embedding vector based on the semantic regularization channel; S5: Perform dimensionality partitioning on the original architecture embedding vector to obtain the decoupled architecture feature vector; S6: Based on the comparison results between the decoupled architecture feature vector and the preset requirements of the power grid enterprise architecture blueprint, identify abnormal patterns such as score missing, score redundancy and dimension mismatch, and generate architecture deviation type labels classified as missing, redundant and mismatched. S7: For abnormal nodes marked with architecture deviation type labels, reverse the search for the nearest neighbor semantic anchors of the same type, generate an interpretable backtracking chain of architecture deviation, and complete the quantitative assessment of the deviation degree of the power grid enterprise architecture operation based on the interpretable backtracking chain of architecture deviation.

[0018] Step S1: Structured parsing of core semantic units from the power grid enterprise architecture blueprint to generate a set of semantic anchor points carrying governance metadata. Specifically, this includes: S1.1: Standardize and parse the original model file in the power grid enterprise architecture blueprint to identify and separate four types of basic architecture elements: business component nodes, application service nodes, data entity nodes, and technical component nodes, and generate an initial architecture node set containing unique node identifiers and type labels.

[0019] Obtain the original model file of the power grid enterprise architecture blueprint. The file is stored in XML, JSON or YAML format and contains the topological connection relationship and attribute description information of business components, application services, data entities and technical components.

[0020] The original model file is subjected to format verification and encoding standardization processing. The integrity and legality of the file structure are checked by using the schema verification mechanism. Text with different encoding formats is uniformly converted to UTF-8 encoding to eliminate parsing anomalies caused by character set differences and generate a standardized architecture description data stream.

[0021] Based on a predefined architecture meta-model ontology library, a multi-mode parsing engine is built. For XML format, a DOM tree traversal method is used to extract the node hierarchy structure. For JSON format, a recursive descent parser is used to extract the key-value pair mapping relationship. For YAML format, an indentation-aware parser is used to restore the nested object structure, thereby achieving a unified abstract representation of heterogeneous data sources.

[0022] During the parsing process, four types of core semantic units are identified and separated based on the node label characteristics. By matching key fields in the node name using regular expressions, nodes containing the identifiers "business", "process", and "capability" are classified as business component nodes; nodes containing the identifiers "API", "service", and "interface" are classified as application service nodes; nodes containing the identifiers "table", "view", and "entity" are classified as data entity nodes; and nodes containing the identifiers "server", "container", and "middleware" are classified as technical component nodes.

[0023] Each identified node type is assigned a globally unique identifier. A hash algorithm is used to combine the node type prefix with the original ID to generate a unique primary key that is not prone to conflict, ensuring the consistency of node references across systems and levels. At the same time, the original ID is retained as a traceability index to generate a basic node list with a unique identifier.

[0024] Each basic node is labeled with a type tag, which is "Business_Component", "Application_Service", "Data_Entity" or "Technical_Component" according to the aforementioned classification results. The type tag is used as the key index field for subsequent hierarchical mapping and semantic anchor construction to form the initial set of architecture nodes.

[0025] By using standardized parsing methods, the original model files of heterogeneous power grid enterprise architecture blueprints are transformed into a unified, uniquely identified, and clearly categorized set of initial architecture nodes. This achieves the transformation from unstructured or semi-structured documents to standardized graph data nodes, providing a clean and standardized data input foundation for subsequent hierarchical mapping.

[0026] S1.2: Based on the node type labels in the initial architecture node set, perform a hierarchical affiliation mapping operation, and use predefined architecture metamodel hierarchical rules to associate each node with the corresponding governance level dimension, generating a hierarchical architecture node sequence carrying hierarchical path information.

[0027] Receive the initial set of architecture nodes generated in step S1.1, which includes unique node identifiers and type labels, as the input data source for the hierarchy affiliation mapping operation.

[0028] Load the predefined hierarchical rule base of the power grid enterprise architecture metamodel. This rule base uses a directed acyclic graph structure to store the vertical dependencies and horizontal collaboration boundaries between business domains, application domains, data domains and technology domains, and defines the complete governance path from top-level strategic capabilities to bottom-level physical infrastructure.

[0029] Iterate through each node in the initial architecture node set, read its type label, and retrieve the corresponding root node path template in the metamodel hierarchical rule base according to the type label to determine the main governance dimension to which the node belongs. For example, map the node of type "business component" to the business capability view branch, and map the node of type "technical component" to the infrastructure view branch.

[0030] For each node, execute the upward tracing method, recursively searching for its parent node along the "belongs to", "supports to", or "implements as" relation edges defined in the metamodel until the root node is reached, recording all intermediate node identifiers along the way, and constructing a complete hierarchical path string from the root node to the current node.

[0031] For complex nodes with multiple inheritance relationships, the principle of maximum information gain is used to select the primary level path. The semantic overlap between the parent node's responsibility description and the current node's responsibility description in each candidate path is calculated. The path with the highest overlap is selected as the primary level, and the remaining paths are retained as auxiliary association attributes to ensure the uniqueness and determinism of the level division.

[0032] The generated hierarchical path string is converted into a hierarchical encoding vector. Each node in the path is digitally represented using one-hot encoding. If the hierarchical depth is less than the preset maximum value, zeros are padded at the end of the vector to align the dimensions, forming a fixed-length hierarchical position feature vector.

[0033] The hierarchical location feature vector is concatenated and fused with the original node attributes to generate a hierarchical architecture node sequence carrying hierarchical path information. Each node object not only contains the original identifier and type, but also embeds its precise coordinate position in the overall architecture blueprint.

[0034] Through the hierarchical attribution mapping process described above, the flattened node set obtained in the previous step is transformed into a hierarchical node sequence with a three-dimensional topology, realizing the precise positioning of architectural elements in the governance space and providing a clear structured context for subsequent responsibility extraction and compliance verification based on hierarchical constraints.

[0035] S1.3: Perform responsibility definition extraction processing on each node in the hierarchical architecture node sequence, parse the functional verbs and object nouns in the node description text using natural language processing technology, construct a standardized responsibility description vector, and generate a responsibility-enhanced node list containing semantic responsibility features.

[0036] S1.4: Based on the node attributes in the enhanced responsibility node list, perform a compliance constraint tag matching operation, logically compare the node features with the constraint clauses in the architecture governance rule base, extract the applicable compliance restrictions, and generate a set of constraint-tagged nodes with compliance constraint tags.

[0037] The architecture governance rule base consists of: The architecture governance rule base consists of the following four core sub-bases, corresponding to the four semantic units extracted in step S1.1 (business components, application services, data entities, and technical components): 1. Business Compliance Rules Sub-library: Stores laws, regulations, industry standards, and internal control requirements related to power grid business activities. For example: The “Load Forecasting Service Capability” node must update the forecasting model every 15 minutes (timeliness constraint). "Electricity billing service" must comply with the "Regulations on Electricity Supply and Use" regarding billing cycle and accuracy requirements (accuracy constraints); Business components must not directly access user privacy fields in the production database (access restrictions).

[0038] 2. Application Architecture Rules Sub-library: Focuses on application service interface specifications, communication protocols, security authentication, and performance metrics. For example: "API Gateway" type services must support OAuth2.0 token authentication (security constraint). The response time for inter-service calls must not exceed 200ms (performance constraint). All exposed RESTful APIs must adhere to the naming prefix / api / v1 / (naming convention constraint).

[0039] 3. Data Governance Rules Sub-library: Covers the lifecycle, quality, lineage, and security requirements of data entities. For example: The "User Electricity Consumption Record Table" must include a data creation timestamp and a last modified person field (integrity constraint). If a data entity contains sensitive information (such as contact information), it must be marked with a desensitization level of "high" (privacy constraint). Foreign key associations between entities must not form circular dependencies (structural constraints).

[0040] 4. Technical Infrastructure Rules Sub-library: This sub-library addresses the deployment environment, version baseline, disaster recovery strategy, and resource quotas for technical components. For example: The containerized Kubernetes cluster must run in a Kubernetes environment of version 1.20 or higher (version baseline constraint). All "server" nodes must be configured with a CPU limit of no more than 8 cores and a memory limit of no more than 32GB (resource quota constraints). Critical middleware (such as message queues) must have persistence and replication mechanisms enabled (high availability constraints).

[0041] S1.5: Perform structured encapsulation processing on the set of constraint marker nodes, integrate the node's unique identifier, hierarchical path information, semantic responsibility features, and compliance constraint tag set, construct a semantic anchor object with complete governance metadata description capabilities, and finally generate a set of semantic anchors carrying governance metadata.

[0042] Step S2: Based on the semantic anchor set, perform anchor-aware heterogeneous neighborhood sampling to generate a sequence of local topological subgraphs with semantic traction characteristics. Specifically, this includes: S2.1: Parse and process the hierarchical attribution tags and responsibility definition tags in the semantic anchor set to extract the strong governance association types between the four core semantic units: business components, application services, data entities, and technical components, and generate a heterogeneous edge type mapping table containing dependency direction and weight coefficient.

[0043] S2.2: Based on the heterogeneous edge type mapping table, the original neighbor nodes of the node to be embedded are filtered. The dynamic depth adjustment method is used to limit the number of sampling hops according to the semantic type of the node, and a set of candidate neighbor nodes that have strong dependencies on the same semantic anchor point is generated.

[0044] The process receives a heterogeneous edge type mapping table and the original set of neighbor nodes of the node to be embedded, serving as the input data source for dynamic depth adjustment and filtering. This step aims to address the context drift problem caused by ignoring the semantic hierarchy of the power grid architecture in traditional random walk sampling. By introducing an adaptive hop count constraint mechanism based on node type, it ensures that the sampling path is always anchored within a semantic neighborhood with strong governance relevance.

[0045] The semantic type labels of the nodes to be embedded are parsed to determine their specific governance dimensions from four categories: business components, application services, data entities, and technical components. A predefined hierarchical constraint rule base is then queried to obtain the maximum allowable sampling depth baseline value for each node type. For business component nodes, due to their top-level architecture and long dependency chains, the initial maximum hop count is set to 3; for application service nodes, the initial maximum hop count is set to 2; and for data entity and technical component nodes, due to their bottom-level architecture and dense connections, the initial maximum hop count is set to 1. This forms the basis for constructing the depth constraint vector.

[0046] The system iterates through the list of direct neighbors of the node to be embedded, identifies the governance association type of each connection edge based on the heterogeneous edge type mapping table, and distinguishes between strong dependencies (such as "deployed on", "hosted on", "belong to") and weak references (such as "logging" and "temporary call"). Only neighbor nodes with strong dependencies are retained in the candidate set, while all neighbor nodes with only weak references are removed. This process initially filters noisy connections at the physical topology level, ensuring the semantic purity of subsequent sampling paths.

[0047] For each neighbor node in the candidate set, recursive deep probing is performed to calculate the shortest path length from the node to be embedded to that neighbor node. The calculated path length is compared with the corresponding threshold in the aforementioned basic depth constraint vector. If the path length exceeds the maximum number of hops allowed for the semantic type of the neighbor node, the neighbor node and its subsequent extended subtrees are removed from the candidate set. This process ensures that the sampling range of nodes at different semantic levels matches its governance radius in the architecture blueprint, avoiding the excessive diffusion of underlying technical components to irrelevant business domains.

[0048] A dynamic depth adjustment factor is introduced to adjust the retention priority of remaining candidate nodes based on the historical interaction frequency between neighboring nodes and similar semantic anchors. A depth decay function is constructed to impose an exponential weight penalty on nodes at the critical hop count, so that core neighbors close to the node to be embedded have a higher retention probability, while edge neighbors face the risk of being truncated even if they have not exceeded the maximum hop count, thereby achieving elastic shrinkage of the sampling range.

[0049] The dynamically adjusted effective sampling depth threshold is calculated using the following formula: in, The effective sampling depth threshold is dynamically adjusted. The initial maximum number of hops is found in the base depth constraint vector. This is a depth adjustment coefficient, ranging from 0.1 to 0.5, used to control the influence of anchor point density on sampling depth. This represents the number of similar semantic anchor points detected within the current local neighborhood. This represents the total number of nodes in the current local neighborhood. When the anchor point density in the neighborhood is high, the calculated result increases, allowing for deeper sampling to capture fine-grained semantics; conversely, it decreases, forcing sampling to focus on the core structure.

[0050] The remaining node set after depth truncation and weight filtering is encapsulated into a structured data object. Each object contains a unique node identifier, hop distance relative to the root node, edge type label, and dynamic weight value. Deduplication is performed on this set to eliminate duplicate node references caused by multi-path reachability, ensuring the uniqueness and topological accuracy of the candidate neighbor node set.

[0051] By using the above-mentioned dynamic depth adjustment and strong dependency filtering method based on semantic type awareness, the heterogeneous edge type mapping table generated in the previous step is transformed into a set of candidate neighbor nodes that retain only those with strong dependencies on the same semantic anchor points. This achieves the convergence of the sampling space from global disorder to local semantic traction, significantly reducing semantic noise interference in the graph embedding process and laying a data foundation for the subsequent construction of high-fidelity local topological subgraph sequences.

[0052] S2.3: Perform probability weighted calculation on the connection relationships in the candidate neighbor node set, and construct a non-uniform transition probability matrix using interface contract compliance degree and life cycle coupling degree as weight factors to generate an anchor point-aware random walk path sequence that reflects the semantic traction strength.

[0053] S2.4: Perform subgraph reconstruction operation based on the anchor point-aware random walk path sequence, extract local topology structure by utilizing the occurrence frequency and order relationship of nodes in the path, and generate a local topology subgraph sequence carrying semantic coordinate origin set information.

[0054] S2.5: Perform structured encapsulation processing on the local topological subgraph sequence, bind and map the subgraph node feature vectors with edge type identifiers, and output a local topological subgraph sequence with semantic traction characteristics as standardized input data for the graph structure fidelity channel.

[0055] like Figure 2 As shown, step S3 involves calculating a weighted semantic similarity index based on the local topological subgraph sequence, and constructing a set of anchor point projection consistency constraint parameters to constrain the vector space distribution. Specifically, this includes: S3.1: Parse the nodes to be embedded and their associated semantic anchors in the local topology subgraph sequence to extract the original governance metadata set containing responsibility definition text, interface contract description and lifecycle state markers, so as to generate a standardized multidimensional governance feature input vector.

[0056] S3.2: Based on the standardized multidimensional governance feature input vector, the semantic matching metric method, the contract compliance verification method, and the coupling degree time series analysis method are executed respectively to independently calculate the responsibility matching degree value reflecting the consistency of business logic, the interface contract compliance degree value reflecting the consistency of interface specifications, and the life cycle coupling degree value reflecting the consistency of the operation cycle, so as to generate a three-dimensional basic similarity index set.

[0057] It receives a standardized multidimensional governance feature input vector, which contains structured data of the node to be embedded and similar semantic anchors in three dimensions: responsibility definition text, interface contract description, and lifecycle state markers.

[0058] For the responsibility definition text dimension, a domain-adaptive pre-trained language model is used to extract deep semantic representations, mapping unstructured functional verbs and object nouns into semantic points in a high-dimensional dense vector space.

[0059] Calculate the cosine similarity between the responsibility vector of the node to be embedded and the responsibility vector of the semantic anchor point, quantify the degree of alignment between the two in terms of business logic intent, and generate a responsibility matching degree value that reflects the consistency of business logic.

[0060] For the interface contract description dimension, the API interface definition file of the node to be embedded and the semantic anchor point is parsed to extract the structured contract elements, including request method, parameter type, return value structure and error code definition.

[0061] Construct an interface contract compliance verification matrix, compare the compatibility of the node to be embedded with the anchor point at the interface signature level, assign full weight to the interface elements that match completely, assign partial weight to the elements that are type compatible but have different names, and assign zero weight to the incompatible elements.

[0062] The weighted average of all interface elements in the verification matrix is ​​calculated to generate an interface contract compliance value that reflects the consistency of interface specifications, ensuring that the technical implementation strictly follows the interaction standards defined in the architecture blueprint.

[0063] Based on the lifecycle state labeling dimension, obtain the version iteration records, launch time, and expected decommissioning time series data of the nodes to be embedded and semantic anchors during the architecture evolution process.

[0064] The coupling degree time series analysis method is executed to calculate the proportion of the overlapping interval between the node to be embedded and the anchor point on the time axis, and the transitivity constraint of the version dependency relationship is introduced to evaluate the synchronization and stability of the two in the running cycle.

[0065] If the lifecycle of the node to be embedded is completely contained within or highly overlaps with the lifecycle of the anchor point, it is determined to be strongly coupled; if there is a significant time misalignment or version gap, it is determined to be weakly coupled. Based on this, a lifecycle coupling degree value reflecting the consistency of the running cycle is generated.

[0066] in, This represents the lifecycle coupling degree value. The set of valid runtimes for the nodes to be embedded. The set of effective runtimes for semantic anchors. This is the version dependency consistency coefficient (ranging from 0 to 1).

[0067] The calculated responsibility matching degree, interface contract compliance degree, and lifecycle coupling degree are encapsulated into a triple structure to form a three-dimensional basic similarity index set, which serves as the standard input data for subsequent weighted fusion operations.

[0068] Through the independent calculation process of semantic matching quantification, contract compliance verification, and coupling time-series analysis, the multi-dimensional governance feature input vector of the previous step is transformed into a quantified three-dimensional basic similarity index set. This enables refined similarity measurement between the embedded node and the semantic anchor point from three key governance perspectives: business logic, technical specifications, and operational cycle. This provides accurate data support for constructing an interpretable anchor point projection consistency constraint parameter set.

[0069] S3.3: Based on the governance importance weight configuration strategy of each indicator in the three-dimensional basic similarity index set, perform weighted fusion operation on the responsibility matching degree value, interface contract compliance degree value and life cycle coupling degree value to generate a comprehensive weighted semantic similarity scalar that represents the overall semantic proximity of the node and the anchor point.

[0070] The system receives a three-dimensional basic similarity index set from the output of the preceding step S3.2. This set includes the responsibility matching degree, interface contract compliance degree, and lifecycle coupling degree calculated for the node to be embedded and similar semantic anchors. These three types of values ​​represent the local semantic proximity features of the node in the three dimensions of business logic consistency, technical specification consistency, and operation cycle consistency, respectively, and constitute the original input data for the weighted fusion operation.

[0071] A dynamic weight configuration strategy matrix based on the governance priorities of the power grid architecture is constructed. This strategy matrix assigns initial weight coefficients to similarity indicators from different governance perspectives according to the business-driven principles and technology support principles defined in the architecture metamodel. For business components and application service nodes, basic weights with higher responsibility matching are assigned to strengthen the dominance of business intent in the vector space. For data entity and technology component nodes, basic weights with higher interface contract compliance and lifecycle coupling are assigned to ensure the standardization of data flow and the stability of technical implementation. The setting of weight coefficients follows normalization constraints to ensure that the sum of the weights of each dimension is always equal to 1, thereby maintaining the dimensional consistency of the semantic similarity scalar.

[0072] An adaptive weighted fusion operation of multidimensional governance indicators is performed, whereby the three-dimensional basic similarity indicators are multiplied by their corresponding dynamic weight coefficients to generate a preliminary comprehensive semantic similarity value. During this process, a non-linear activation function is introduced to smooth the weighted indicators, suppressing overall similarity distortion caused by abnormal fluctuations in a single indicator. Specifically, the Sigmoid function is used to map the weighted responsibility matching degree, interface contract compliance degree, and lifecycle coupling degree, ensuring that the fused similarity value strictly converges within the [0,1] interval, enhancing the robustness of the numerical distribution.

[0073] A semantic consistency regularization correction term is introduced to perform a secondary calibration on the preliminary comprehensive semantic similarity value. This correction term is calculated based on the centrality index of nodes in the local topological subgraph. If a node is on a critical dependency path, its confidence weight in the similarity value is increased; if a node is in an isolated marginal position, its contribution to the similarity value is decreased.

[0074] The generated comprehensive weighted semantic similarity scalar undergoes boundary validity verification and extreme value truncation to eliminate non-physically meaningful values ​​caused by data noise or computational errors. If the similarity scalar exceeds a preset effective threshold range, it is forcibly mapped to the nearest valid boundary value to ensure the numerical stability of subsequent vector space projection transformations. The final output is a calibrated comprehensive weighted semantic similarity scalar, which accurately represents the overall semantic proximity between the node to be embedded and similar semantic anchors from a multi-dimensional governance perspective.

[0075] Through the above adaptive weighted fusion and nonlinear regularization correction methods, the discretized three-dimensional basic similarity index obtained in the previous step is transformed into a continuous and comprehensive weighted semantic similarity scalar with governance semantic interpretability. This realizes the mapping from multi-source heterogeneous governance features to a unified semantic measurement space, providing a high-precision quantitative basis for constructing the anchor point projection consistency constraint parameter set, and effectively solving the problem of spatial distribution disorder caused by semantic drift in traditional graph embedding.

[0076] S3.4: Using the comprehensive weighted semantic similarity scalar as the mapping coefficient, combined with the preset four-dimensional governance perspective space basis vector, perform a vector space projection transformation operation to convert the scalar similarity into an anchor point projection consistency constraint tensor with directionality and hierarchy, so as to generate preliminary vector space distribution constraint parameters.

[0077] Obtain the comprehensive weighted semantic similarity scalar output from step S3.3. This scalar represents the overall proximity between the node to be embedded and the four types of semantic anchors in the dimensions of responsibility, contract, and lifecycle.

[0078] Read the preset set of basis vectors in the four-dimensional governance perspective space. This set contains the reference direction vectors corresponding to the four orthogonal subspaces of business components, application services, data entities and technical components, respectively. The dimension of each basis vector is consistent with the dimension of the target embedding space.

[0079] For each type of semantic anchor, a mapping function from scalar similarity to vector space projection intensity is constructed. A nonlinear activation function is used to convert the comprehensive weighted semantic similarity scalar into projection coefficients with physical meaning, ensuring that the projection intensity is non-negative and has gradient differentiability.

[0080] The obtained projection coefficients are multiplied element-wise with the corresponding four-dimensional governance perspective space basis vectors to generate the initial projection vector component in the direction of the anchor point. This component reflects the semantic contribution strength of the node under a specific governance perspective.

[0081] Vector superposition is performed on the initial projection vector components from four governance perspectives to construct a preliminary anchor point projection consistency constraint tensor. This tensor forms a local coordinate system constraint field in the vector space with the semantic anchor point as the origin.

[0082] By introducing a hierarchical awareness correction factor, the components of different dimensions in the constraint tensor are weighted and adjusted according to the node hierarchy depth defined in the architecture blueprint, so that the constraint force of high-level business nodes occupies a dominant position in the vector space, while the constraint force of low-level technical nodes serves as auxiliary support.

[0083] S3.5: Normalize and smooth the preliminary vector space distribution constraint parameters and verify the boundary validity to eliminate abnormal extreme values ​​caused by data noise. Finally, construct the anchor point projection consistency constraint parameter set to guide the training of the graph embedding model, so as to achieve strict alignment between the vector space distribution and the semantics of the architecture blueprint.

[0084] like Figure 3 As shown, step S4 involves using the local topological subgraph sequence as input to the graph structure fidelity channel, and constructing a semantic regularization channel by combining it with the anchor projection consistency constraint parameter set. The original architecture embedding vector is then generated based on the semantic regularization channel. Specifically, this includes: S4.1: Based on the node connection relationships and edge weight information in the local topological subgraph sequence, perform graph convolution aggregation operation to extract neighborhood structure features and generate an initial node hidden state representation containing topological dependencies.

[0085] S4.2: Using the initial node hidden state representation and the responsibility matching degree and interface contract compliance degree index defined in the anchor point projection consistency constraint parameter set, perform four-dimensional semantic space projection transformation processing to generate an anchor point projection feature vector that reflects the distribution intensity of nodes from four governance perspectives: business, application, data, and technology.

[0086] S4.3: Based on the distance deviation between the anchor point projection feature vector and the preset standard semantic origin, and combined with the life cycle coupling degree weight, calculate the semantic regularization loss term to generate a semantic alignment constraint gradient signal used to punish semantic drift.

[0087] Receive the anchor point projection feature vector output from step S4.2. This vector represents the distribution strength of the node to be embedded from four governance perspectives: business, application, data, and technology. At the same time, obtain the lifecycle coupling weight coefficients contained in the anchor point projection consistency constraint parameter set constructed in step S3.

[0088] Extract a set of predefined standard semantic origins. This set consists of the ideal coordinates of the four types of core semantic anchors defined in the architecture blueprint in the low-dimensional embedding space. These coordinates correspond to the origins of business components, application services, data entities, and technical components, respectively, and serve as a benchmark reference system for measuring semantic drift.

[0089] Calculate the Euclidean distance deviation between each component of the anchor point projection feature vector and the standard semantic origin of the corresponding category. For the business dimension, calculate the geometric distance between the component of the projection vector in the business subspace and the business anchor origin; for the application dimension, calculate the geometric distance between the component of the projection vector in the application subspace and the application anchor origin; and so on, calculate the distance deviation for the data dimension and the technical dimension to generate a four-dimensional distance deviation vector.

[0090] A lifecycle coupling degree weight coefficient is introduced to weight and correct the four-dimensional distance deviation vector. This weight coefficient reflects the consistency between the node's running state and the blueprint planning state in the time dimension. High coupling degree means that the node should be more strictly close to the semantic origin of its type, while low coupling degree allows for a certain degree of semantic flexibility.

[0091] The weighted semantic regularization loss term is calculated using the following formula: in, The semantic regularization loss term is represented by k, which represents the governance perspective dimension index, with a value range of 1 to 4, corresponding to the business, application, data, and technology dimensions, respectively. The k-th dimension represents the lifecycle coupling weight coefficient, which is obtained from step S3; This represents the component of the anchor point projection eigenvector in the k-th dimensional subspace; The coordinate vector represents the origin of the k-th standard semantic class; ||·|| represents the Euclidean norm operation.

[0092] Automatic differentiation is performed on the calculated semantic regularization loss term. The gradient of the loss function with respect to the anchor projection feature vector is calculated by backpropagation using the chain rule, generating a semantic alignment constraint gradient signal. This gradient signal indicates the direction and magnitude of the current embedding vector's deviation from the ideal origin in the semantic space.

[0093] The semantic alignment constraint gradient signal is passed to the optimizer to update the neural network parameters of the graph embedding model. This forces the embedding vectors generated in subsequent iterations to move closer to the standard origin of their respective semantic types while maintaining the topological structure, thereby suppressing semantic drift caused by random initialization or noise interference.

[0094] By using the weighted distance deviation calculation and gradient backpropagation processing described above, the anchor point projection feature vector from the previous step is transformed into a semantic alignment constraint gradient signal used to penalize semantic inconsistencies. This achieves strict alignment between the embedding space distribution and the semantic standards of the architecture blueprint, ensuring that the deviation evaluation results have traceable semantic basis.

[0095] S4.4: Calculate the graph structure fidelity loss term based on the reconstruction error of the initial node hidden state representation, and then perform weighted fusion of the graph structure fidelity loss term and the semantic alignment constraint gradient signal to generate a dual-channel comprehensive optimization loss function value that includes the topology preservation objective and the semantic alignment objective.

[0096] The initial hidden state representation of nodes is obtained from the graph structure fidelity channel output, as well as the node connection adjacency matrix and edge weight information defined in the local topological subgraph sequence, which serve as the basic input data for calculating the graph structure reconstruction error.

[0097] Based on the initial node hidden state representation, a decoder mapping layer is constructed to project the high-dimensional hidden state vector back into the node feature space of the original architecture graph, generating a predicted node feature vector for evaluating the reconstruction quality.

[0098] The mean squared error loss function is used to calculate the sum of squared Euclidean distances between the predicted node feature vector and the original node true feature vector, thereby quantifying the degree of reconstruction deviation of a single node in the feature space.

[0099] A negative sampling mechanism is introduced, which selects non-directly connected node pairs from the local topological subgraph sequence as negative samples, and calculates the dot product similarity score between positive sample node pairs and negative sample node pairs in the embedding space.

[0100] We use the binary cross-entropy loss function to measure the difference between the predicted connection probability of positive and negative sample pairs and the actual labels, and construct a graph structure fidelity loss term that reflects the ability to preserve the topology.

[0101] Obtain the semantic alignment constraint gradient signal output by the semantic regularization channel. This signal represents the penalty for the node embedding vector deviating from the preset four-dimensional governance perspective semantic origin.

[0102] Set topology preservation weight coefficient and semantic alignment weight coefficient, and dynamically adjust the relative importance of the two loss objectives based on the different emphasis requirements of business compliance and technical stability in the governance of power grid enterprise architecture.

[0103] The graph structure fidelity loss term and the semantic regularization loss term corresponding to the semantic alignment constraint gradient signal are linearly weighted and summed to construct a dual-channel comprehensive optimization loss function.

[0104] in, To optimize the loss function value for dual-channel synthesis, To preserve the weight coefficients in the topology, For the graph structure fidelity loss term, For semantic alignment weight coefficients, This is the semantic regularization loss term.

[0105] Automatic differentiation is performed on the dual-channel integrated optimization loss function value to calculate the gradient vector of the loss function relative to the network parameters of each layer of the graph embedding model.

[0106] The gradient signal is backpropagated to the graph convolutional aggregation layer and the semantic space projection layer using the chain rule, updating the node embedding vector and model weight parameters to minimize the overall loss function value.

[0107] By using the above dual-channel loss fusion and backpropagation optimization method, the initial node hidden state representation and semantic alignment constraint gradient signal generated in the previous step are transformed into a dual-channel comprehensive optimization loss function value containing the topology preservation objective and the semantic alignment objective, thereby achieving the expected technical effect of synergistic optimization of graph structure integrity and governance semantic consistency.

[0108] S4.5: Utilize the dual-channel integrated optimization loss function value to update the network parameters of the graph embedding model through backpropagation, iterate and optimize until convergence, and output the original architecture embedding vector with embedded four-dimensional governance perspective features and topological consistency.

[0109] The dual-channel integrated optimization loss function value is received. This value is formed by weighted fusion of the graph structure fidelity loss term and the semantic alignment constraint gradient signal, and serves as the starting error signal for the backpropagation method.

[0110] Based on the chain rule, the partial derivatives of the dual-channel integrated optimization loss function with respect to the network parameters of each layer in the graph embedding model are calculated to generate a composite gradient vector containing the gradient of topology reconstruction error and the gradient of semantic drift penalty.

[0111] For the weight matrix in the graph convolutional aggregation layer, gradient backpropagation is performed using the composite gradient vector to calculate the contribution of each neuron in the total loss function, forming a progressively decreasing error sensitivity distribution map.

[0112] An adaptive moment estimation optimizer is introduced to perform exponentially weighted moving average processing of the first and second moments on the calculated original gradient, thereby eliminating gradient oscillation caused by the sparsity of power grid architecture data.

[0113] Based on the preset learning rate decay strategy, the parameter update step size is dynamically adjusted, and the gradient value after moment estimation is applied to the iterative update of the model weights to ensure that the optimal solution is quickly approached in the early stage of convergence and then finely adjusted in the later stage.

[0114] Monitor the rate of change of the loss function within consecutive iterations. When the difference between the loss values ​​of two adjacent epochs is lower than the preset convergence threshold, determine that the model training has reached a stable state and stop the parameter update process.

[0115] Extract the node hidden layer representation in the final converged state, map it to a low-dimensional vector space, and generate the original architecture embedding vector with embedded four-dimensional governance perspective features and topological consistency.

[0116] By combining backpropagation with adaptive optimization strategies, the dual-channel loss function value is transformed into the optimal configuration of model parameters, achieving the expected technical effect that the architecture instance maintains both topological adjacency and semantic anchor constraints in the vector space.

[0117] Step S5: The original architecture embedding vector is subjected to dimensionality partitioning to obtain the decoupled architecture feature vector. Specifically, this includes: S5.1: Obtain the original architecture embedding vector with embedded four-dimensional governance perspective features, and perform fully connected layer mapping processing on the original architecture embedding vector based on the preset linear projection transformation matrix to generate intermediate hidden layer feature representations carrying global semantic hybrid information.

[0118] The original architecture embedding vector, which is generated by optimization training of the dual-channel loss function and contains embedded four-dimensional governance perspective features, is received. The vector dimension is set to 128 dimensions, which carries the hybrid feature information of the power grid architecture node under the dual effects of topological structure and semantic constraints.

[0119] A pre-defined linear projection transformation matrix is ​​constructed, with a dimension of 128 rows and 64 columns. This matrix serves as the set of weight parameters for mapping from the original high-dimensional embedding space to the intermediate hidden layer feature space. The initial values ​​of the matrix are initialized using the Xavier initialization strategy to ensure the stability of gradient propagation.

[0120] The fully connected layer mapping process is performed, and the original architecture embedding vector and linear projection transformation matrix are multiplied to achieve linear transformation and dimensionality compression of the feature space, thereby extracting key latent factors from the global semantic mixture information.

[0121] A bias vector term is introduced, which has a dimension of 64, to adjust the feature distribution center after linear transformation, compensate for the information offset that may be caused by dimension compression, and enhance the model's ability to fit nonlinear semantic relationships.

[0122] By applying linear activation functions to maintain the continuity and differentiability of feature values ​​and avoid information loss caused by nonlinear truncation, an intermediate hidden layer feature representation carrying global semantic hybrid information is generated. This representation preserves the coupling relationship of the four-dimensional perspectives of business, application, data and technology, and provides a standardized input basis for subsequent semantic slot decoupling.

[0123] By using linear projection transformation, the original architecture embedding vector generated in the previous step is transformed into intermediate hidden layer feature representation data carrying global semantic hybrid information, realizing the initial mapping from high-dimensional hybrid features to low-dimensional decoupled features, laying the data structure foundation for the channel segmentation operation of the subsequent multi-dimensional mask segmentation operator.

[0124] S5.2: Utilizing the intermediate hidden layer feature representation, based on the distribution patterns of the four types of semantic anchor points defined in the architecture blueprint (business components, application services, data entities, and technical components), construct a multi-dimensional mask segmentation operator and perform channel segmentation on the intermediate hidden layer feature representation to generate four mutually orthogonal and dimension-independent initial semantic slot vector sequences.

[0125] S5.3: For each initial semantic slot vector in the initial semantic slot vector sequence, a nonlinear feature enhancement process is performed using an activation function with sparse constraints to suppress noise responses in irrelevant dimensions and highlight key features from a specific governance perspective, thereby generating a pure semantic slot vector with sparse activation characteristics.

[0126] The four mutually orthogonal and dimension-independent initial semantic slot vector sequences output from step S5.2 are received as input data objects for nonlinear feature enhancement processing.

[0127] For each initial semantic slot vector, a sparse activation mapping mechanism based on L1 regularization constraints is constructed to suppress background noise response of irrelevant dimensions and highlight key semantic features from a specific governance perspective.

[0128] A nonlinear activation function with sparsity constraints is defined. This function retains the rectification characteristics of ReLU and introduces an L1 norm penalty term to force the sparsity of vector elements. A weighted linear transformation operation is performed on the input vector to project the high-dimensional mixed features to a feature subspace that matches the governance perspective, generating a preliminary linear activation precursor vector.

[0129] Calculate the L1 norm of the linear activation predecessor vector to quantify the overall strength of non-zero elements in the current vector, serving as an intermediate indicator for measuring semantic redundancy.

[0130] Based on the preset sparsity control coefficient λ, a soft threshold shrinkage process is performed on each element in the linear activation predecessor vector, forcing small noise components with absolute values ​​less than the threshold to zero, while retaining the core feature components with higher significance.

[0131] A modified linear unit activation function is applied to the vector after soft thresholding to remove all negative components, ensuring that the output features are non-negative and monotonic in semantic interpretation.

[0132] The sparsity control coefficient λ is dynamically adjusted by an iterative optimization method to maintain the proportion of non-zero elements in the output vector within a preset sparsity range, thus balancing the feature retention rate and the noise suppression rate.

[0133] The processed vector is normalized and rescaled to restore the magnitude information lost due to sparsification and generate a pure semantic slot vector with sparse activation characteristics.

[0134] By using a nonlinear activation function with sparse constraints, the initial semantic slot vector from the previous step is transformed into a pure semantic slot vector with sparse activation characteristics. This enables the accurate extraction of key features from four governance perspectives—business, application, data, and technology—and the effective suppression of irrelevant noise, significantly improving the signal-to-noise ratio and interpretability of subsequent deviation assessments.

[0135] S5.4: Based on the pure semantic slot vector, the probability distribution mapping operation is performed by the normalized index calculation module to convert the absolute values ​​in the vector space into relative strength indicators, so as to quantify the existence confidence of the node from the corresponding governance perspective and generate the original saliency score scalar representing the saliency of the role in each dimension.

[0136] S5.5: Integrate the original saliency score scalars generated from the four governance perspectives, and perform vector concatenation and structured encapsulation processing in a fixed order of business, application, data and technology to form a final output data structure in a unified format, generating a decoupled architecture feature vector containing the saliency scores of each role dimension.

[0137] Step S6: Based on the comparison results between the decoupled architecture feature vector and the preset requirements of the power grid enterprise architecture blueprint, identify abnormal patterns such as score missing, score redundancy, and dimension mismatch, and generate architecture deviation type labels classified as missing, redundant, and mismatched. Specifically, this includes: S6.1: Numerical extraction processing is performed on the business semantic slots, application semantic slots, data semantic slots and technical semantic slots in the feature vector of the decoupled architecture to obtain the original role saliency score set of the node to be evaluated from the perspective of four-dimensional governance, which serves as the benchmark input data for subsequent deviation calculation.

[0138] S6.2: Based on the compliance constraint labels defined in the architecture blueprint, extract the expected existence flag and threshold range parameters corresponding to each semantic unit, and use the logical mapping method to align and match the original role saliency score set with the expected existence flag to generate a standardized comparison matrix containing the actual score value and the theoretical expected value.

[0139] S6.3: Perform multi-dimensional state discrimination logic operation based on the standardized comparison matrix. Dimensions with actual scores below the zero threshold but theoretical expected values ​​are marked as potential missing states. Dimensions with actual scores above the dynamic redundancy threshold but theoretical expected values ​​are marked as potential redundant states. Dimensions with negative correlation deviations between business semantic slot scores and application semantic slot scores or data semantic slot scores are marked as potential mismatch states. Output a preliminary deviation pattern set containing various abnormal state identifiers.

[0140] The set of actual role saliency scores and theoretical expected value flags in the standardized alignment matrix are read as the input data source for the multidimensional state discrimination logic operation. For each semantic dimension slot in the alignment matrix, the corresponding actual score value and theoretical existence flag are extracted to construct a dimension-level state judgment triplet.

[0141] Perform a logical judgment operation for missing deviation states. Compare the actual score value with a preset zero threshold. If the actual score value is less than or equal to the zero threshold, and the theoretical existence flag indicates that the dimension is a mandatory existence state in the architecture blueprint, then the dimension is determined to be in a potential missing state.

[0142] The dimension identifiers identified as potentially missing and their corresponding confidence markers are written into the initial deviation pattern set, and the missing type code and associated semantic anchor index of the dimension are recorded.

[0143] Perform a logical judgment operation for redundant deviation states. Obtain the dynamic redundancy threshold of the business scenario to which the current node belongs. This threshold is dynamically calculated based on the statistical distribution of historical operating data. Compare the actual score with the dynamic redundancy threshold. If the actual score is higher than the dynamic redundancy threshold, and the theoretical existence flag indicates that the dimension is unnecessary or prohibited in the architecture blueprint, then the dimension is determined to be in a potential redundancy state.

[0144] The dimension identifiers identified as potentially redundant and their magnitude indicators exceeding the threshold are written into the initial deviation pattern set, and the redundancy type code and associated semantic anchor index of the dimension are recorded.

[0145] Perform logical judgment operations for mismatched deviation states. Extract the role saliency score of the business semantic slot as a benchmark reference value. Calculate the Pearson correlation coefficient or cosine similarity deviation value between the business semantic slot score and the application semantic slot score, and between the business semantic slot score and the data semantic slot score, respectively.

[0146] If the calculated negative correlation deviation index exceeds the preset mismatch judgment threshold, it indicates that there is a significant semantic gap between the business function and its underlying application support or data carrier, and the combination of dimensions is judged to be in a potential mismatch state.

[0147] The identifiers and deviation indices of the dimension pairs identified as potential mismatches are written into the initial deviation pattern set, and the mismatch type code of the dimension and the two semantic slot indices involved are recorded.

[0148] Perform integrity checks on the initial set of deviation patterns to ensure that all states marked as abnormal have clear dimensional orientation and type classification criteria.

[0149] Through the above multi-dimensional state discrimination logic operation, the numerical differences in the standardized comparison matrix are transformed into a preliminary deviation pattern set containing three types of potential abnormal state identifiers: missing, redundant, and mismatched. This enables refined anomaly capture of architecture running instances from a four-dimensional governance perspective, providing structured intermediate results for subsequent confidence correction and final classification.

[0150] S6.4: Use a weighted fusion strategy to perform cross-dimensional correlation verification on single-dimensional anomaly identifiers in the initial deviation pattern set, calculate the comprehensive deviation confidence index by combining the responsibility definition label, filter out false alarm states with confidence scores below the preset noise threshold, and generate a deterministic architecture deviation pattern sequence after confidence score correction.

[0151] Obtain the preliminary deviation pattern set output from step S6.3. This set contains potential anomaly status identifiers and corresponding original role salience scores for the nodes to be evaluated across four dimensions: business, application, data, and technology. For each node marked as potentially anomaly in the set, extract the four-dimensional salience score vector from its decoupling architecture feature vector and the responsibility definition label metadata generated in step S1, which serve as the input basis for confidence calculation.

[0152] A cross-dimensional correlation verification matrix is ​​constructed, and the logical coupling weights between various governance perspectives are determined based on the predefined hierarchical dependency rules in the power grid enterprise architecture blueprint. Specifically, business components have strong dependencies on application services, application services have strong references to data entities, and technical components have a carrying relationship with application services. Based on this topology, a 4x4 dimensional correlation weight matrix is ​​generated, where diagonal elements are set to 1.0, off-diagonal elements are assigned values ​​of 0.8 (strong dependency) or 0.5 (weak dependency) according to the dependency strength, and elements with no direct correlation are assigned a value of 0.1.

[0153] Calculate the local confidence score for single-dimensional anomalies. For each potential anomaly dimension in the initial deviation pattern, combine the deviation magnitude of the role saliency score of that dimension with a preset threshold, and normalize using the Sigmoid function to map the deviation magnitude to the [0, 1] interval to obtain the initial anomaly strength value. Simultaneously, introduce a semantic matching factor for responsibility definition tags. If the overlap between the actual responsibility description of a node and the responsibility keywords defined in the blueprint is higher than 0.7, assign a high confidence coefficient of 1.2; otherwise, assign a low confidence coefficient of 0.8.

[0154] Perform cross-dimensional consistency checks and use a weighted fusion strategy to correct local confidence levels. For dimensions marked as "potentially missing," check whether their upstream dependent dimensions also have low scores. For example, if the "Application Service" dimension has an extremely low score, it is necessary to check whether its downstream "Data Entity" dimension also has a score below the threshold. If both upstream and downstream dimensions show a low score trend, it is determined to be a systemic missing dimension, and the overall confidence level of that dimension is increased; if only a single dimension has a low score while related dimensions are normal, it is determined to be local noise, and its overall confidence level is reduced.

[0155] A dynamic noise filtering threshold is set, which is dynamically adjusted based on the historical operational stability of the nodes. For core critical nodes, the threshold is set to 0.85 to ensure high sensitivity; for edge auxiliary nodes, the threshold is set to 0.65 to tolerate some fluctuations. The calculated comprehensive deviation confidence index C is compared with the dynamic noise filtering threshold.

[0156] If the overall deviation confidence index C is greater than the dynamic noise filtering threshold, the anomaly marker for that dimension is retained, and its status is updated from "potential" to "deterministic". If the overall deviation confidence index C is less than or equal to the dynamic noise filtering threshold, the anomaly is determined to be a false alarm or noise interference, the anomaly marker for that dimension is cleared, and the status of that dimension is reset to "normal".

[0157] Iterate through all nodes and dimensions in the initial deviation pattern set, repeating the confidence calculation and filtering operations described above. Aggregate the retained deterministic outlier states after filtering according to node ID to form a structured data sequence. This sequence contains a unique node identifier, a list of deterministic outlier dimensions, the corresponding comprehensive deviation confidence score, and specific score evidence that triggered the outlier.

[0158] By employing a weighted fusion strategy and cross-dimensional correlation verification, the fuzzy potential anomaly identifiers present in the previous step are transformed into a highly reliable deterministic architectural deviation pattern sequence. This achieves noise suppression and accuracy enhancement of the architectural deviation detection results, ensuring that subsequent classification mapping is based solely on conclusive semantic deviation evidence.

[0159] S6.5: Perform classification encoding mapping operation based on deterministic architecture deviation pattern sequence, map potential missing states to missing architecture deviation type labels, map potential redundant states to redundant architecture deviation type labels, map potential mismatch states to mismatch architecture deviation type labels, and finally output a set of architecture deviation type labels carrying clear classification results.

[0160] Receive a deterministic architecture deviation pattern sequence after confidence correction. This sequence contains the abnormal state identifiers of the node to be evaluated in each semantic slot and the corresponding confidence scores, which serve as input data for the classification coding mapping.

[0161] A classification coding mapping table based on governance intent is constructed, and three deviation types, namely missing, redundant, and mismatched, are mapped to unique integer coding identifiers. Specifically, missing is mapped to code 1, redundant to code 2, mismatched to code 3, and normal state to code 0. A one-way mapping relationship is established from deviation pattern to standardized label space.

[0162] It iterates through each node record in the deterministic architecture deviation pattern sequence, extracts its associated dominant deviation state identifier, and handles the coexistence of multiple states according to the preset priority rules. When a node has multiple potential deviation states at the same time, the state with the highest confidence score is selected as the final judgment basis to ensure the uniqueness and determinism of the classification results.

[0163] For nodes identified as potentially missing, verify whether the role salience score of any dimension of its business, application, data, or technology is lower than the zero threshold and whether the blueprint requires that dimension to exist. If the condition is met, assign the deviation type field of the node to the missing type code 1 and mark the node as an architectural element missing anomaly.

[0164] For nodes identified as potentially redundant, verify whether their role salience score in any dimension is higher than the dynamic redundancy threshold and whether the blueprint does not define the necessity of that dimension. If the condition is met, assign the deviation type field of the node to redundancy code 2 and mark the node as an architectural element redundancy anomaly.

[0165] For nodes identified as potentially mismatched, check whether there is a significant negative correlation between their business semantic slot score and their data or technical semantic slot score, i.e., a high business score and a low support dimension score. If the condition is met, assign the deviation type field of the node to mismatch type code 3 and mark the node as an architectural element mismatch anomaly.

[0166] For nodes that do not trigger any abnormal judgment conditions, their deviation type field is kept at the normal state code 0, indicating that the node meets the expected constraints of the architecture blueprint from the perspective of four-dimensional governance and no governance intervention is required.

[0167] The deviation type codes of all nodes are integrated into a unified data structure to form an architecture deviation type label set with clear classification results. This label set includes a unique node identifier, deviation type code, dominant deviation dimension index, and confidence score field.

[0168] Step S7: For anomalous nodes marked with an architecture deviation type label, reverse the search for the nearest neighbor semantic anchors of the same type to generate an interpretable backtracking chain of architecture deviation. Based on the interpretable backtracking chain of architecture deviation, a quantitative assessment of the deviation degree of the power grid enterprise's architecture operation is completed. Specifically, this includes: S7.1: Perform numerical analysis on the decoupled architecture feature vectors corresponding to the abnormal nodes marked as architecture deviation type labels to extract the role saliency score set from the four governance perspectives of business, application, data and technology, and determine the target semantic slot identifier with the highest confidence through the maximum value comparison method to generate the dominant dimension index signal to guide the backtracking direction.

[0169] S7.2: Based on the dominant dimension index signal, perform nearest neighbor retrieval operation in the original architecture embedding vector space, use the cosine similarity measurement method to calculate the spatial distance between the abnormal node and the same type of anchor in the semantic anchor set, filter out the reference object with the closest topological structure and semantic features, and generate the nearest neighbor semantic anchor object with unique identification information.

[0170] Receive the dominant dimension index signal output by step S7.1. This signal clearly identifies the semantic slot with the highest degree of deviation of the abnormal node in the four governance perspectives of business, application, data or technology, and serves as the dimension constraint condition for subsequent retrieval operations.

[0171] From the pre-built set of semantic anchors, a subset of candidate anchors that are consistent with the dominant dimension index signal type are selected to ensure that the retrieval scope is strictly limited to the reference object space with the same semantic level and governance attributes as the abnormal node, thus eliminating cross-domain semantic interference.

[0172] The full-dimensional feature vector of the abnormal node in the original architecture embedding vector space is extracted. This vector retains the comprehensive distribution information of the node under the dual effects of graph structure topology and semantic constraints, and serves as the query benchmark for similarity calculation.

[0173] For each semantic anchor in the candidate anchor subset, its corresponding standard anchor embedding vector is extracted. This vector is a fixed reference coordinate that is forcibly aligned to the origin of the four-dimensional governance space by a semantic consistency regularization term during the model training phase.

[0174] The cosine similarity metric is used to calculate the cosine of the angle between the embedding vector of the abnormal node and the embedding vector of each candidate anchor point, quantifying the directional consistency between the two in the semantic space. The cosine similarity set is sorted in descending order to identify the candidate anchor point with the highest similarity value, forming a preliminary nearest neighbor candidate list.

[0175] Euclidean distance is introduced as an auxiliary verification index to calculate the straight-line distance between the abnormal node and each anchor point in the preliminary nearest neighbor candidate list in the vector space, so as to correct the magnitude deviation problem that may be caused by relying solely on directional consistency.

[0176] A comprehensive distance scoring function is constructed, which converts cosine similarity into a distance penalty term and fuses it with Euclidean distance in a weighted manner to generate a comprehensive matching score that reflects the degree of proximity in both topological structure and semantic features.

[0177] The candidate list is sorted a second time based on the comprehensive matching score, and the single semantic anchor with the highest score that exceeds the preset confidence threshold is selected as the final target reference object to ensure the uniqueness and authority of the backtracking basis.

[0178] The unique identifier, semantic type label, and coordinate information in the vector space of the target reference object are encapsulated to generate the nearest neighbor semantic anchor object of the same type carrying unique identification information, which serves as the direct input for extracting governance constraint clauses in step S73.

[0179] S7.3: Perform in-depth mining of the governance metadata encapsulated in the nearest neighbor semantic anchor object, extract the set of compliance restrictions strongly associated with the anchor based on the predefined constraint clause mapping table, including responsibility definition matching rules, interface contract specifications and lifecycle coupling requirements, and generate a structured list of governance constraint clauses.

[0180] S7.4: Logically compare and analyze the governance constraint clause list with the actual architecture operation instance data of the abnormal node, and use the template filling engine to fill the deviation type label, dominant dimension index signal and the number of the specific constraint clause violated into the preset natural language attribution template to generate a preliminary natural language attribution statement fragment describing the cause and basis of the deviation.

[0181] S7.5: Perform semantic coherence verification and context completion processing on the preliminary natural language attribution statement fragments, construct a complete causal inference chain by combining the hierarchical path information of the architecture blueprint, and output a complete interpretable backtracking chain of architecture deviations that includes abnormal node identifiers, deviation pattern descriptions and specific violation clause references, thereby realizing a closed-loop mapping from vector space deviations to business semantic attributions.

[0182] Receive the initial natural language attribution statement fragments, unique identifiers of anomalous nodes, dominant dimension index signals, and complete governance metadata of nearest neighbor semantic anchor objects output from step S7.4.

[0183] Syntactic structural integrity checks are performed on preliminary natural language attribution sentence fragments. A dependency parser is used to identify the core subject-verb-object components in the sentence and detect grammatical defects such as missing predicates, unclear object references, or broken logical connectors.

[0184] For detected grammatical defects, a predefined semantic completion rule library is invoked, and the corresponding standard terminology dictionary is extracted from the metamodel definition of the architecture blueprint based on the governance perspective (business, application, data, or technology) determined by the dominant dimension index signal.

[0185] The extracted standard terminology dictionary is semantically matched with the missing parts of the sentence. The maximum likelihood estimation method is used to select the term with the highest confidence to fill the missing parts of the sentence, repair the semantic gaps caused by template filling, and generate the grammatically corrected attribution statement backbone.

[0186] Obtain the hierarchical path information encapsulated in the nearest neighbor semantic anchor object of the same type. This path information describes the subordinate relationship of the anchor in the power grid enterprise architecture blueprint, such as "provincial marketing system - load management module - real-time measurement service".

[0187] The node sequence in the hierarchical path information is parsed to construct the hierarchical context vector. The depth and breadth features of the hierarchy are encoded by a recursive neural network to quantify the positional sensitivity of the current abnormal node in the overall architecture topology.

[0188] The hierarchical context vector is mapped to natural language modifiers, and based on the preset hierarchical description template, context supplementary clauses containing spatial location descriptions such as "the current node is located under the parent node" or "as a dependent of the sibling node".

[0189] Inserting contextual supplement clauses after the subject of the attribution statement enhances the spatial specificity of the attribution statement, ensuring that architects can quickly pinpoint the specific architectural level at which the anomaly occurred.

[0190] Extract the list of governance constraint clauses corresponding to the nearest neighbor semantic anchor points of the same type, identify the logical antecedents and compliance requirement consequents in the clauses, and construct a causal logic graph.

[0191] Compare the actual test architecture running instance data of the abnormal node with the preconditions of the governance constraint clause to confirm the specific factual basis for triggering the constraint clause, such as "it was detected that the service has not registered any data source interface".

[0192] Based on the confirmed factual evidence, the corresponding description of the consequences of the violation is retrieved from the causal logic graph, that is, the consequent part of the constraint clause, such as "causing business data to be unable to be synchronized in real time, violating the data consistency principle".

[0193] By using causal connector templates, factual evidence and descriptions of the consequences of violations are linked together to generate logical explanatory chain paragraphs with clear causal relationships, thereby strengthening the logical persuasiveness of the attribution results.

[0194] The grammatically corrected attribution statement core, hierarchical context supplementary clauses, and logical explanation chain paragraphs are structurally assembled and the text is assembled in the order of "abnormal subject location - deviation phenomenon description - hierarchical background supplementation - violation causal inference - specific clause citation".

[0195] The coherence score is calculated for the assembled complete text. The semantic similarity score between adjacent sentences is calculated using the BERT model. If the score is lower than the preset threshold of 0.85, the conjunctions are adjusted or the sentence structure is reorganized until the coherence requirement is met.

[0196] The final output includes a complete interpretable backtracking chain of architectural deviations, containing anomaly node identifiers, deviation pattern descriptions, hierarchical background information, and specific violation clause references.

[0197] like Figure 4 As shown, the present invention also provides a quantification and assessment device for the operational deviation of a power grid enterprise architecture, employing the aforementioned quantification and assessment method for operational deviation of a power grid enterprise architecture, comprising: The semantic anchor generation module is used to structurally parse core semantic units from the power grid enterprise architecture blueprint and generate a set of semantic anchors carrying governance metadata. The local topological subgraph sequence generation module is used to perform anchor-aware heterogeneous neighborhood sampling operations based on the semantic anchor set to generate a local topological subgraph sequence with semantic traction characteristics. The constraint parameter set construction module is used to calculate a weighted semantic similarity index based on the local topological subgraph sequence and construct an anchor point projection consistency constraint parameter set for constraining the distribution of the constraint vector space. The original architecture embedding vector generation module is used to use the local topological subgraph sequence as input to the graph structure fidelity channel, and to construct a semantic regularization channel in combination with the anchor projection consistency constraint parameter set, and generate the original architecture embedding vector based on the semantic regularization channel. The decoupled architecture feature extraction module is used to perform dimensionality partitioning on the original architecture embedding vector to obtain the decoupled architecture feature vector. The deviation type identification module is used to identify abnormal patterns such as score missing, score redundancy and dimension mismatch based on the comparison results between the feature vector of the decoupled architecture and the preset requirements of the power grid enterprise architecture blueprint, and generate architecture deviation type labels classified as missing, redundant and mismatched. The interpretable backtrace chain generation module is used to reverse-search for the nearest neighbor semantic anchors of the same type for abnormal nodes marked as architectural deviation type labels, generate an interpretable backtrace chain of architectural deviation, and complete the quantitative assessment of the deviation degree of the power grid enterprise's architecture operation based on the interpretable backtrace chain of architectural deviation.

[0198] like Figure 5 As shown, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the power grid enterprise architecture operation deviation quantification assessment method.

[0199] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components 41, 42, and 43 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0200] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0201] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit and its external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions for deviations from quantification evaluation methods in the operation of power grid enterprise architecture. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0202] In some embodiments, the processor 42 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or to process data, for example, to execute computer-readable instructions that deviate from the quantification assessment method of the power grid enterprise architecture.

[0203] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0204] The present invention also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the steps of the aforementioned power grid enterprise architecture operation deviation quantification assessment method.

[0205] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0206] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0207] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for quantitatively evaluating operational deviations of power grid enterprise architecture, characterized in that, Specifically, it includes: S1: Structure and analyze the core semantic units from the power grid enterprise architecture blueprint to generate a set of semantic anchor points carrying governance metadata; S2: Based on the set of semantic anchor points, perform anchor-aware heterogeneous neighborhood sampling operations to generate a sequence of local topological subgraphs with semantic traction characteristics; S3: Calculate the weighted semantic similarity index based on the local topological subgraph sequence, and construct the anchor point projection consistency constraint parameter set for constraining the vector space distribution; S4: Use the local topological subgraph sequence as the input of the graph structure fidelity channel, and combine it with the anchor point projection consistency constraint parameter set to construct a semantic regularization channel, and generate the original architecture embedding vector based on the semantic regularization channel; S5: Perform dimensionality partitioning on the original architecture embedding vector to obtain the decoupled architecture feature vector; S6: Based on the comparison results between the decoupled architecture feature vector and the preset requirements of the power grid enterprise architecture blueprint, identify abnormal patterns such as score missing, score redundancy and dimension mismatch, and generate architecture deviation type labels classified as missing, redundant and mismatched. S7: For abnormal nodes marked with architecture deviation type labels, reverse the search for the nearest neighbor semantic anchors of the same type, generate an interpretable backtracking chain of architecture deviation, and complete the quantitative assessment of the deviation degree of the power grid enterprise architecture operation based on the interpretable backtracking chain of architecture deviation.

2. The method for quantifying and evaluating operational deviations of power grid enterprise architecture according to claim 1, characterized in that, The core semantic units include four categories: business components, application services, data entities, and technical components.

3. The method for quantifying and evaluating operational deviations of power grid enterprise architecture according to claim 1, characterized in that, The heterogeneous neighborhood sampling operation includes prioritizing the selection of neighboring subgraphs that have a strong dependency relationship with semantic anchors of the same type.

4. The method for quantifying and evaluating operational deviations of power grid enterprise architecture according to claim 1, characterized in that, The step of performing dimensionality partitioning on the original architecture embedding vector to obtain the decoupled architecture feature vector includes: The original architecture is embedded into a vector input semantic decoupling head and subjected to dimensional partitioning to separate independent semantic slots corresponding to four governance perspectives: business, application, data, and technology. Based on the independent semantic slots, a decoupling architecture feature vector containing the saliency scores of each dimension's role is output.

5. The method for quantifying and evaluating operational deviations of power grid enterprise architecture according to claim 4, characterized in that, The original architecture is embedded into the vector input semantic decoupling head and then divided into four independent semantic slots corresponding to the four governance perspectives of business, application, data and technology. Specifically, after the original architecture is embedded into the vector input semantic decoupling head, it is divided into four independent semantic slots for the four governance perspectives of business, application, data and technology by linear projection and multi-dimensional masking segmentation operators.

6. The method for quantifying and evaluating operational deviations of power grid enterprise architecture according to claim 2, characterized in that, Step S1 specifically includes: Standardized parsing and processing are performed on the original model files in the power grid enterprise architecture blueprint to identify and separate four types of basic architecture elements: business component nodes, application service nodes, data entity nodes, and technical component nodes. An initial architecture node set is generated based on the basic architecture elements. Based on the initial set of architecture nodes, a hierarchical affiliation mapping operation is performed to associate each node with the corresponding governance level dimension, generating a hierarchical architecture node sequence. For each node in the hierarchical architecture node sequence, a responsibility definition extraction process is performed. Functional verbs and object nouns in the node description text are parsed using natural language processing technology to generate a responsibility-enhanced node list. Based on the aforementioned list of enhanced responsibility nodes and the architecture governance rule base, a matching operation is performed to conduct a logical comparison and generate a set of constraint-marked nodes. The constraint marker node set is structurally encapsulated to construct a semantic anchor object with complete governance metadata description capabilities. Based on the semantic anchor object, a semantic anchor set carrying governance metadata is finally generated.

7. The method for quantifying and evaluating operational deviations of power grid enterprise architecture according to claim 6, characterized in that, The architecture governance rule base is a pre-built, dynamically updated structured knowledge base used to store and manage various compliance restrictions that must be followed during the design, development, and operation of power grid enterprise architecture.

8. A device for quantifying and evaluating operational deviations of a power grid enterprise architecture, characterized in that, The method for quantifying deviations from the operational structure of a power grid enterprise, as described in any one of claims 1 to 7, includes: The semantic anchor generation module is used to structurally parse core semantic units from the power grid enterprise architecture blueprint and generate a set of semantic anchors carrying governance metadata. The local topological subgraph sequence generation module is used to perform anchor-aware heterogeneous neighborhood sampling operations based on the semantic anchor set to generate a local topological subgraph sequence with semantic traction characteristics. The constraint parameter set construction module is used to calculate a weighted semantic similarity index based on the local topological subgraph sequence and construct an anchor point projection consistency constraint parameter set for constraining the distribution of the constraint vector space. The original architecture embedding vector generation module is used to use the local topological subgraph sequence as input to the graph structure fidelity channel, and to construct a semantic regularization channel in combination with the anchor projection consistency constraint parameter set, and generate the original architecture embedding vector based on the semantic regularization channel. The decoupled architecture feature extraction module is used to perform dimensionality partitioning on the original architecture embedding vector to obtain the decoupled architecture feature vector. The deviation type identification module is used to identify abnormal patterns such as score missing, score redundancy and dimension mismatch based on the comparison results between the feature vector of the decoupled architecture and the preset requirements of the power grid enterprise architecture blueprint, and generate architecture deviation type labels classified as missing, redundant and mismatched. The interpretable backtrace chain generation module is used to reverse-search for the nearest neighbor semantic anchors of the same type for abnormal nodes marked as architectural deviation type labels, generate an interpretable backtrace chain of architectural deviation, and complete the quantitative assessment of the deviation degree of the power grid enterprise's architecture operation based on the interpretable backtrace chain of architectural deviation.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the quantification assessment method for deviation of power grid enterprise architecture operation as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the power grid enterprise architecture operation deviation quantification assessment method as described in any one of claims 1 to 7.

Images