An industrial chain map-based power system hierarchical classification method and system

Abstract

The application provides a power system hierarchical classification method and system based on an industry chain graph, comprising: in a pre-constructed power system industry chain graph, determining a graph node corresponding to each current power achievement data; then generating an initial power data feature vector of each current power achievement data; fusing the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain a current power fusion feature vector of each current power achievement data; and using a pre-constructed power hierarchical classification model to output an industry chain hierarchical label to which each current power achievement data belongs and classify a plurality of current power achievement data. The use of the industry chain graph makes the classification result of the current power achievement data consistent with the industry chain semantics, and provides reliable data support for the power achievement data management of the power system.

Description

Technical Field

[0001] This invention belongs to the field of power system industry chain technology, specifically relating to a power system hierarchical classification method and system based on industry chain map. Background Technology

[0002] As power systems evolve towards digitalization and intelligence, the power industry has accumulated a wealth of technological achievements and data, including research project reports, technical standards, literature, engineering acceptance documents, and academic papers. This data is widely distributed across all segments of the power industry chain, including power generation, transmission, distribution, consumption, and energy storage, and constitutes crucial knowledge assets supporting the technological evolution and industrial upgrading of the power system.

[0003] In practical applications, the scientific and rational hierarchical classification of power industry achievement data is of significant practical importance. For example, in the scenario of technology achievement management in power enterprises, achievement data from different sources and of different types need to be categorized according to their functional positioning in the industry chain to facilitate achievement retrieval, knowledge reuse, and performance evaluation. In the scenario of industry technology situation analysis, by classifying and statistically analyzing achievement data from each link in the industry chain, it is possible to reveal the level of technological innovation activity, technological evolution trends, and upstream and downstream collaborative relationships in different links, providing data support for industry planning and policy formulation. In the scenario of scientific research project application and resource allocation, accurate achievement classification helps to identify technological gaps and advantageous directions, guiding the precise allocation of scientific research investment.

[0004] However, the existing technology lacks a hierarchical classification mechanism that can simultaneously take into account the content characteristics of the power output data itself, which makes it difficult to accurately classify power outputs that cross links and levels, affecting the level of precision in power system output management. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, in a first aspect, this invention proposes a power system hierarchical classification method based on a supply chain map, comprising: Collect multiple current power output data at the power system level; In the pre-constructed power system industry chain map, determine the map node corresponding to each of the current power achievement data; Based on each current power achievement data, feature extraction is performed on the current power achievement data to generate an initial power data feature vector for each current power achievement data; the initial power data feature vector is then fused with the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain the current power fusion feature vector for each current power achievement data. For each current power fusion feature vector, use a pre-built power hierarchy classification model to output the industry chain hierarchy label to which each current power achievement data belongs; The multiple current power achievement data are classified according to the industry chain level tag to which each current power achievement data belongs.

[0006] Preferably, the pre-construction process of the power system industry chain map includes: Multiple power entities are identified from the collected historical power data at the power system level; these multiple power entities are then used as multiple graph nodes. Based on the multiple power entities, the upstream and downstream relationships between the multiple entities are determined according to their functional positioning or business process sequence in the power system; based on the upstream and downstream relationships between the multiple entities and the obtained technical reference relationships between the multiple power entities, multiple graph edges are constructed. The multiple graph nodes and multiple graph edges are integrated to form an initial power system industry chain graph; A graph embedding algorithm is used to perform representation learning on the initial power system industry chain graph to generate graph embedding vectors for multiple graph nodes; The graph embedding vector of each graph node is associated with the initial power system industry chain graph to form the final power system industry chain graph.

[0007] Preferably, the step of fusing the initial power data feature vector with the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph to obtain the current power fusion feature vector for each current power result data includes: Calculate the attention weight between the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power output data in the power system industry chain graph; Based on the attention weights, a cross-attention transformation is performed on the initial power data feature vector and the graph embedding vector to obtain the current power fusion feature vector for each current power result data.

[0008] Preferably, the calculation of the attention weight between the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power output data in the power system industry chain graph includes: Using linear transformation technology, the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph are mapped to the same space to obtain the mapped initial power data feature vector and the mapped graph embedding vector. Calculate the similarity score between the mapped initial power data feature vector and the mapped graph embedding vector to obtain the vector similarity score; The vector similarity scores are normalized to obtain attention weights.

[0009] Preferably, the current power fusion feature vector of each of the current power results data satisfies the following formula:

[0010] In the above formula, This represents the current power integration feature vector. This indicates normalization processing. Represents the initial power data feature vector. Represents a graph embedding vector. This represents the total number of multi-head attention heads in a linear transformation. Indicates the first k The learning fusion weights of individual attention heads This represents a cross-attention transformation based on attention weights.

[0011] Preferably, the pre-construction process of the hierarchical classification model includes: Multiple nodes in the power system industry chain map are used as multiple network nodes of the graph neural network; the historical power fusion feature vector corresponding to each map node is used as the initial feature of the corresponding network node. The neighbor relationships between multiple network nodes are determined based on the edge connection relationships between multiple nodes in the power system industry chain map. Based on the neighbor relationships, with the goal of fusing the topological information of its neighboring nodes with the feature representation of each network node, the initial features of multiple network nodes are iteratively updated according to the constructed node feature update model to obtain the power hierarchy classification model.

[0012] Preferably, the construction process of the node feature update model satisfies the following formula:

[0013] In the above formula, The first character of the graph neural network is represented by the second character. Layer network nodes Update features, Indicates the first Layer network nodes initial characteristics, This represents the activation function of a graph neural network. N Represents a set of network nodes. Represents network nodes The set of neighboring nodes, Represents network nodes The set of neighboring nodes, Indicates the first Layer network nodes initial characteristics, The first character of the graph neural network is represented by the second character. Layer weight matrix.

[0014] Preferably, the step of extracting features from each current power output data point to generate an initial power data feature vector for each current power output data point includes: Based on each of the current power output data, multiple text segments are obtained; The multiple text portions are vectorized to obtain multiple text feature vectors; The multiple text feature vectors are fused to generate an initial power data feature vector for each current power achievement data.

[0015] Secondly, a hierarchical classification system for power systems based on an industry chain map includes: The data acquisition module is used to collect multiple current power output data located at the power system level; The graph node determination module is used to determine the graph node corresponding to each current power achievement data in a pre-constructed power system industry chain graph. The feature vector fusion acquisition module is used to extract features from each current power achievement data based on each current power achievement data, and generate an initial power data feature vector for each current power achievement data; and to fuse the initial power data feature vector with the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain the current power fusion feature vector for each current power achievement data. The hierarchical label acquisition module is used to take each current power fusion feature vector and use a pre-built power hierarchical classification model to output the industry chain hierarchical label to which each current power achievement data belongs; The classification module is used to classify the multiple current power achievement data according to the industry chain level tag to which each current power achievement data belongs.

[0016] Preferably, the system further includes a power system supply chain map construction module, used for: Multiple power entities are identified from the collected historical power data at the power system level; these multiple power entities are then used as multiple graph nodes. Based on the multiple power entities, the upstream and downstream relationships between the multiple entities are determined according to their functional positioning or business process sequence in the power system; based on the upstream and downstream relationships between the multiple entities and the obtained technical reference relationships between the multiple power entities, multiple graph edges are constructed. The multiple graph nodes and multiple graph edges are integrated to form an initial power system industry chain graph; A graph embedding algorithm is used to perform representation learning on the initial power system industry chain graph to generate graph embedding vectors for multiple graph nodes; The graph embedding vector of each graph node is associated with the initial power system industry chain graph to form the final power system industry chain graph.

[0017] Preferably, the fused feature vector acquisition module includes: The attention weight calculation module is used to calculate the attention weight between the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph. The power fusion feature vector acquisition module is used to perform cross-attention transformation on the initial power data feature vector and the graph embedding vector according to the attention weight, so as to obtain the current power fusion feature vector of each current power result data.

[0018] Preferably, the attention weight calculation module is specifically used for: Using linear transformation technology, the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph are mapped to the same space to obtain the mapped initial power data feature vector and the mapped graph embedding vector. Calculate the similarity score between the mapped initial power data feature vector and the mapped graph embedding vector to obtain the vector similarity score; The vector similarity scores are normalized to obtain attention weights.

[0019] Preferably, the current power fusion feature vector of each of the current power results data satisfies the following formula:

[0020] In the above formula, This represents the current power integration feature vector. This indicates normalization processing. Represents the initial power data feature vector. Represents a graph embedding vector. This represents the total number of multi-head attention heads in a linear transformation. Indicates the firstk The learning fusion weights of individual attention heads This represents a cross-attention transformation based on attention weights.

[0021] Preferably, the system further includes a hierarchical classification model construction module, used for: Multiple nodes in the power system industry chain map are used as multiple network nodes of the graph neural network; the historical power fusion feature vector corresponding to each map node is used as the initial feature of the corresponding network node. The neighbor relationships between multiple network nodes are determined based on the edge connection relationships between multiple nodes in the power system industry chain map. Based on the neighbor relationships, with the goal of fusing the topological information of its neighboring nodes with the feature representation of each network node, the initial features of multiple network nodes are iteratively updated according to the constructed node feature update model to obtain the power hierarchy classification model.

[0022] Preferably, the construction process of the node feature update model satisfies the following formula:

[0023] In the above formula, The first character of the graph neural network is represented by the second character. Layer network nodes Update features, Indicates the first Layer network nodes initial characteristics, This represents the activation function of a graph neural network. N Represents a set of network nodes. Represents network nodes The set of neighboring nodes, Represents network nodes The set of neighboring nodes, Indicates the first Layer network nodes initial characteristics, The first character of the graph neural network is represented by the second character. Layer weight matrix.

[0024] Preferably, the fused feature vector acquisition module is specifically used for: Based on each of the current power output data, multiple text segments are obtained; The multiple text portions are vectorized to obtain multiple text feature vectors; The multiple text feature vectors are fused to generate an initial power data feature vector for each current power achievement data.

[0025] Thirdly, this application also proposes an electronic device, comprising: at least one processor and a memory; wherein the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the power system hierarchical classification method based on the industry chain map is implemented.

[0026] Fourthly, this application also proposes a readable storage medium having an executable program stored thereon, wherein when the executable program is executed, it implements the aforementioned power system hierarchical classification method based on industry chain map.

[0027] Compared with the closest prior art, the present invention application has the following beneficial effects: This invention discloses a power system hierarchical classification method and system based on a power industry chain graph, comprising: collecting multiple current power achievement data located at the power system hierarchy; determining the graph node corresponding to each current power achievement data in a pre-constructed power system industry chain graph; extracting features from each current power achievement data to generate an initial power data feature vector for each current power achievement data; fusing the initial power data feature vector with the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain a current power fused feature vector for each current power achievement data; using a pre-constructed power hierarchy classification model, outputting the industry chain hierarchy label to which each current power achievement data belongs; and classifying the multiple current power achievement data according to the industry chain hierarchy label to which each current power achievement data belongs. Collecting multiple current power achievement data at the power system level can take into account the hierarchical characteristics of the data's content. Then, using a supply chain map, the supply chain level label of the current power achievement data can be obtained, ensuring that the classification results of the current power achievement data are semantically consistent with the supply chain. This hierarchical classification mechanism can effectively avoid classification bias caused by similar power achievement data but different functional levels. It can accurately classify power achievements that cross links and levels. At the same time, it can provide reliable data support for the management of power achievement data, technology layout analysis, and collaborative development of the supply chain within the power system. Attached Figure Description

[0028] Figure 1 A flowchart of a power system hierarchical classification method based on industry chain graph provided for this invention application; Figure 2 This invention application provides a power system hierarchical structure diagram based on a power system hierarchical classification method using an industry chain graph. Figure 3 A schematic diagram of a power system supply chain map for a power system hierarchical classification method based on supply chain map provided in this invention application; Figure 4 An architecture diagram of a power system hierarchical classification system based on a supply chain graph is provided for this invention application; Figure 5 This is a schematic diagram of the operation of an electronic device provided in this invention application. Detailed Implementation

[0029] The specific embodiments of this invention will be further described in detail below with reference to the accompanying drawings.

[0030] Example 1: like Figure 1 As shown, this invention application proposes a hierarchical classification method for power systems based on industry chain diagrams, which may include the following steps: Step 1: Collect multiple current power performance data at the power system level; Step 2: In the pre-constructed power system industry chain map, determine the map node corresponding to each of the current power achievement data; Step 3: Based on each current power achievement data, perform feature extraction on the current power achievement data to generate an initial power data feature vector for each current power achievement data; fuse the initial power data feature vector with the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain the current power fusion feature vector for each current power achievement data; Step 4: Using a pre-built power hierarchy classification model, output the industry chain hierarchy label to which each current power achievement data belongs for each current power fusion feature vector; Step 5: Classify the multiple current power achievement data according to the industry chain level tag to which each current power achievement data belongs.

[0031] In step 1 above, the power system hierarchy refers to the functional hierarchy based on the power system industry chain structure, used to characterize the specific position of power output data in the stages of power production, transmission, use, and storage. In a specific implementation, such as Figure 2As shown, the power system hierarchy includes at least four main levels: generation side, grid side, load side, and energy storage side. The types of power output data include, but are not limited to, technical reports, acceptance documents, academic papers, and technical standard documents for power projects. These data originate from various business segments of the power system and reflect the technical achievements or entity information of different nodes in the industry chain. Data collection methods may include database retrieval, file import, or web crawling that meets regulatory requirements.

[0032] In step 2 above, a graph node corresponding to each current power achievement data is determined within a pre-constructed power system industry chain graph. Specifically, each current power achievement data is identified as an entity, and based on the entity identification result (i.e., the corresponding power entity or power technology achievement), it is mapped to a graph node in the power system industry chain graph. This graph node represents the power entity (such as an enterprise or research institution) or power technology achievement (such as a standard or technical solution) corresponding to the power achievement data. The mapping method can employ a text similarity-based matching method or utilize information such as the name and attributes of graph nodes in the power system industry chain graph for association.

[0033] In practical applications, a single power achievement data set often contains attribute information across multiple dimensions. For example, a technical report may cover a company's R&D background (corresponding to the company entity node), specific engineering implementation plans (corresponding to the technical achievement node), and relevant industry standards (corresponding to the standard document node). Directly mapping the entire document to a single graph node can easily lead to mapping bias due to the mixed attributes, thus affecting the accuracy of subsequent classification. To address this issue, this solution designs a graph node mapping strategy based on multi-attribute priority in step 2. Specifically, it includes the following steps: Step 2.1: Perform fine-grained entity extraction on the current power results data and divide it into several attribute units according to preset attribute categories. The attribute categories include, but are not limited to: subject attributes (enterprises, research institutions, etc. involved), technical attributes (technical solutions, process parameters, etc. involved), standard attributes (technical standards, specification documents, etc. involved), and project attributes (project names, engineering codes, etc. involved).

[0034] Step 2.2: Based on the characteristics of the power system hierarchical classification task, pre-set the mapping priority of different attribute categories. The mapping priority takes into account that the hierarchical classification of the industrial chain depends more on the technical and standard attributes to which the result data belongs, while the dependence on the subject attributes is relatively low. Therefore, the technical and standard attributes are set as high priority, and the subject attributes and project attributes are set as low priority.

[0035] Step 2.3: When mapping graph nodes, priority is given to matching graph nodes based on high-priority attribute units. If the matching results of high-priority attribute units are missing or the confidence level is lower than a preset threshold, then low-priority attribute units are used to supplement the matching step by step. Through the above method, it is ensured that the mapped graph nodes can reflect the essential belonging of power results data at the industry chain level to the greatest extent, rather than being affected by non-core attributes in the document.

[0036] This multi-attribute priority mapping strategy differs fundamentally from conventional entity linking methods: conventional methods typically aim to link the entire document or the main body of the document to a "most relevant" node in the graph, ignoring the primary and secondary relationships when multiple attributes coexist within the document; while this scheme, by explicitly distinguishing attribute priorities, integrates the goal orientation of the classification task into the mapping process, realizing "graph association guided by classification goals," which significantly improves the ability of subsequent fusion features to represent the hierarchy of the power industry chain.

[0037] By mapping power output data to nodes in the power system industry chain map and using the hierarchical relationships of these nodes as strong prior knowledge, semantic ambiguity in the textual data of power output data can be effectively suppressed, thereby improving the accuracy and consistency of subsequent classification results.

[0038] Step 2 above, the pre-construction process of the power system industry chain map, includes: Step a: Identify multiple power entities from the collected historical power data at the power system level; and use these multiple power entities as multiple graph nodes. Step b: Based on the multiple power entities, determine the upstream and downstream relationships between the multiple entities according to their functional positioning or business process sequence in the power system; based on the upstream and downstream relationships between the multiple entities and the obtained technical reference relationships between the multiple power entities, construct multiple graph edges; Step c: Integrate the multiple graph nodes and multiple graph edges to form an initial power system industry chain graph; Step d: Using a graph embedding algorithm, representation learning is performed on the initial power system supply chain graph to generate graph embedding vectors for multiple graph nodes; Step e: Associate the graph embedding vector of each graph node with the initial power system industry chain graph to form the final power system industry chain graph.

[0039] In one specific implementation of the example, the pre-construction process of the aforementioned power system industry chain map is described in detail. Taking the wind power generation field as an example, multiple power entities are identified as map nodes from the collected historical power achievement data (including technical reports, project documents, industry standards, etc.) at the power system level: On the generation side, entities such as wind turbine manufacturers, wind farm developers, wind turbine blade manufacturers, wind turbine gearbox manufacturers, and wind turbine converter manufacturers are extracted; on the grid side, entities such as transmission operators, step-up substations, transmission lines, distribution network operators, and smart distribution systems are extracted; on the load side, entities such as industrial users (e.g., large manufacturing enterprises), commercial users (e.g., commercial complexes), residential users, electric vehicle charging operators, and demand response service providers are extracted; on the energy storage side, entities such as pumped storage power stations, lithium-ion battery manufacturers, flywheel energy storage system integrators, and hydrogen energy storage technology research and development institutions are extracted. After treating the aforementioned power entities as multiple graph nodes, the upstream and downstream relationships between these entities are determined based on their functional positioning and business process sequence within the power system, as shown in Table 1. For example, there is an upstream supply relationship between wind turbine blade manufacturers and wind turbine manufacturers; an equipment supply relationship between wind turbine manufacturers and wind farm developers; a power grid connection relationship between wind farm developers and transmission operators; a power transmission relationship between transmission operators and distribution network operators; and a power supply relationship between distribution network operators and industrial users. Simultaneously, combining the technology reference relationships between multiple power entities obtained from technical report citation data and technical standard association data—for example, the technology of wind turbine converter manufacturers is referenced by the technical solutions of wind turbine manufacturers, and the technical standards of energy storage systems are referenced by the grid connection solutions of wind farm developers—multiple graph edges are constructed. Integrating these multiple graph nodes and edges forms the initial power system industry chain graph. Then, a graph embedding algorithm can be used to learn the representation of the initial power system supply chain graph. For example, the Node2Vec (Node to Vector) algorithm can be used to learn the low-dimensional vector representation of the graph nodes by sampling the graph node sequence through a biased random walk, generating a graph embedding vector for each graph node. Finally, the graph embedding vector of each graph node is associated and stored with the corresponding graph node in the initial power system supply chain graph to form the final power system supply chain graph. This final power system supply chain graph (e.g., ...) Figure 3 Each graph node in the diagram contains node attribute information and its corresponding graph embedding vector, which is used for feature fusion of the current power results data.

[0040] Table 1 Relationships between power entities

[0041]

[0042] The construction of the power system supply chain map in this scheme does not simply use the map nodes as retrieval results. Instead, it deeply fuses the graph embedding vectors of the map nodes with the initial feature vectors extracted from the text during the feature fusion stage. This design allows the structured topological information contained in the power system supply chain map (such as upstream and downstream relationships and hierarchical dependencies between map nodes) to participate in classification decisions in a vectorized manner, compensating for the lack of global structure perception capability in pure text features. Compared to the conventional approach of using the map only for entity links, this achieves the synergistic expression of "structural knowledge" and "content knowledge," significantly enhancing the power system supply chain's ability to discriminate hierarchical boundaries.

[0043] In step 3 above, the step of extracting features from each current power output data point to generate an initial power data feature vector for each current power output data point may include the following steps: Step A3.1: Based on each of the current power output data, obtain multiple text segments; Step A3.2: Vectorize the multiple text portions to obtain multiple text feature vectors; Step A3.3: Fuse the multiple text feature vectors to generate an initial power data feature vector for each current power result data.

[0044] Steps A3.1-A3.3 above aim to extract an initial power data feature vector from each current power output data. A specific example of this process can be as follows: In step 3.1, based on the determined type of current power achievement data, the corresponding parsing rules are determined. Then, each current power achievement data is parsed, and according to the parsing rules corresponding to its type, it is divided into multiple text parts. For technical report types, it is divided into project name, executive summary, introduction, technical solution, experimental data, results analysis, and conclusion. For academic paper types, it is divided into title, abstract, introduction, methods, experiments, and conclusion. For standard document types, it is divided into standard name, scope, normative references, terms and definitions, technical requirements, and test methods. Furthermore, for power achievement data where the text part contains structured data such as charts, formulas, or tables, corresponding image recognition, formula parsing, or structured data encoding techniques can be used to extract corresponding multimodal features for subsequent fusion with multiple text feature vectors. Through the above division method, each current power achievement data is parsed into multiple text parts. Given the length limit of the BERT (Bidirectional Encoder Representations from Transformers) model for input text (typically 512 tokens), directly inputting complete power data into the model would result in truncation of excessively long portions, leading to information loss. Therefore, by splitting each current power data point into multiple text parts according to its semantic structure, each text part can meet the input length requirement of the BERT model, thus achieving complete encoding of the full-text information and avoiding feature loss due to truncation.

[0045] In step 3.2, the BERT model can be used to vectorize each text segment, converting each text segment into a fixed-dimensional vector representation. Specifically, for each text segment, it is input into the pre-trained BERT model, and the vector corresponding to the label output by the model is taken as the text feature vector of that text segment, resulting in the text feature vector of each text segment, denoted as t1, t2...ti...tn, where each ti can be a 768-dimensional or 1024-dimensional real-valued vector. In terms of BERT model application, the BERT feature vectors of multiple text segments are fused (e.g., by weighted averaging, concatenation, etc.) to form a complete document-level initial power data feature vector while maintaining the independent representation capability of each text segment.

[0046] In step 3.3, multiple text feature vectors (and optional multimodal features) are fused to generate an initial power data feature vector for each current power result data.

[0047] In one specific implementation of the example, average pooling can be used for fusion, that is, the arithmetic mean of all text feature vectors is calculated as the initial power data feature vector. The average pooling method is introduced here mainly because it is simple to calculate, does not require additional training parameters, and can efficiently fuse the feature vectors of multiple text parts into an initial power data feature vector of fixed-dimensional document-level representation. This can meet the requirements of input vector dimension consistency for subsequent tasks such as classification, retrieval or similarity calculation.

[0048] The initial power data feature vectors described above can satisfy the following formula:

[0049] In the above formula, Let represent the initial power data feature vector, n represent the total number of text parts, i represent the i-th text part, and ti represent the i-th text feature vector.

[0050] In step 3 above, fusing the initial power data feature vector with the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph to obtain the current power fusion feature vector for each current power result data may include the following steps: Step B3.1: Calculate the attention weight between the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph; Step B3.2: Based on the attention weights, perform cross-attention transformation on the initial power data feature vector and the graph embedding vector to obtain the current power fusion feature vector for each current power result data.

[0051] Steps B3.1-B3.2 above aim to fuse the initial power data feature vector with the graph embedding vector of the corresponding graph node to generate the fused current power fusion feature vector. This current power fusion feature vector retains the textual semantic information of the power data itself while incorporating its topological structure information in the power system industry chain graph, providing a more comprehensive feature representation for subsequent hierarchical classification. When obtaining the graph embedding vector, it is obtained by learning the representation of a pre-constructed power system industry chain graph using a graph embedding algorithm, and is used to characterize the positional information of the graph node in the power system hierarchy. Examples of graph embedding algorithms include Node2Vec, GraphSAGE (GraphSample and AggregatE, inductive graph representation learning), and other graph embedding algorithms suitable for isomorphic graphs. In step B3.1 above, calculating the attention weight between the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power output data in the power system industry chain graph may include the following steps: Step 3.1.1: Using linear transformation technology, the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph are mapped to the same space to obtain the mapped initial power data feature vector and the mapped graph embedding vector; Step 3.1.2: Calculate the similarity score for the mapped initial power data feature vector and the mapped graph embedding vector to obtain the vector similarity score; Step 3.1.3: Normalize the vector similarity scores to obtain attention weights.

[0052] In steps 3.1.1-3.1.3 above, firstly, a linear transformation technique is used to map the initial power data feature vector and the graph embedding vector to the same space, respectively, to obtain the mapped initial power data feature vector and the mapped graph embedding vector located in the same space; then, the similarity score between the mapped initial power data feature vector and the mapped graph embedding vector is calculated to obtain the vector similarity score; finally, the softmax function can be used to normalize the vector similarity score to obtain the attention weight.

[0053] The above-mentioned linear transformation technique maps the initial power data feature vector and the graph embedding vector of the corresponding graph node in the power system industry chain graph to the same space. This is mainly because the initial power data feature vector and the graph embedding vector may be in different vector space distributions in the original space. The former comes from the semantic space after text feature fusion, while the latter comes from the topological space obtained by graph structure learning. The two differ in dimension, magnitude, and semantic distribution, and direct similarity calculation would produce a large deviation. By introducing the linear transformation technique, a learnable linear projection matrix is ​​applied to the two types of feature vectors respectively, mapping them to the same shared space. This achieves the alignment and unified representation of heterogeneous features, laying a foundation for the same distribution for subsequent similarity calculation.

[0054] In step B3.2 above, the step of performing cross-attention transformation on the initial power data feature vector and the graph embedding vector according to the attention weight to obtain the current power fusion feature vector for each current power result data includes: First, cross-attention transformation is used. Achieve initial power data feature vector With graph embedding vectors Cross-modal interaction, in which As a query As keys and values, they are computed through an attention mechanism. and The correlation between various nodes in the graph is analyzed, and the graph embedding vectors are weighted and aggregated according to the correlation weights, thereby integrating structured knowledge of the industry chain graph into the graph. In the semantic representation. Based on this, a multi-head attention mechanism is introduced, and a setting is made. Multiple parallel attention heads, each independently learning a different linear transformation matrix, form multiple subspaces; initial power data feature vectors are captured from these multiple subspaces. With graph embedding vectors The interaction relationships between them; at the same time, learning fusion weights are introduced for each attention head. ,right The outputs of each attention head are weighted and summed. This learning fusion weight is adaptively adjusted during the interaction process, enabling the fusion mechanism to dynamically allocate the contribution of each attention head to the final current power fusion feature vector according to the characteristics of different types of power output data.

[0055] Secondly, a residual connection structure is adopted to combine the weighted summation of the multi-attention head output with the initial power data feature vector. Add them together to form This design borrows the residual concept from the Transformer (a neural network architecture based on self-attention), allowing the original input information to bypass the attention transformation layer and be directly transmitted to the output. This alleviates the gradient vanishing problem in deep networks and ensures that the fused feature vectors are preserved. The original text semantic information is preserved to avoid the loss or distortion of original features due to the attention mechanism's over-focus on graph information.

[0056] Finally, the results after residual connection are normalized using LayerNorm to obtain the final current power fusion feature vector. Layer normalization can stabilize the distribution of features within a range of mean 0 and variance 1, effectively accelerating model convergence and enhancing the numerical stability of features. This makes the fused feature vector comparable across different samples, facilitating processing by subsequent hierarchical classification models. Through the above design, It achieves deep integration of textual semantic features and industry chain graph structure features, which not only retains the core information of the initial power results data, but also adaptively incorporates relevant knowledge associations in the graph, providing an information-rich and stable fusion feature foundation for subsequent hierarchical classification tasks.

[0057] Therefore, the current power fusion feature vector of each of the aforementioned current power results data can be expressed by the following formula:

[0058] In the above formula, This represents the current power integration feature vector. This indicates normalization processing. Represents the initial power data feature vector. Represents a graph embedding vector. This represents the total number of multi-head attention heads in a linear transformation. Indicates the first k The learning fusion weights of individual attention heads This represents a cross-attention transformation based on attention weights.

[0059] In step 4 above, the pre-construction process of the hierarchical classification model includes: Step 4.1: Use multiple nodes in the power system industry chain map as multiple network nodes of the graph neural network; use the historical power fusion feature vector corresponding to each map node as the initial feature of the corresponding network node; Step 4.2: Determine the neighbor relationships between multiple network nodes based on the edge connection relationships between multiple nodes in the power system industry chain graph; Step 4.3: Based on the neighbor relationships, with the goal of fusing the topological information of the neighbor nodes with the feature representation of each network node, the initial features of multiple network nodes are iteratively updated according to the constructed node feature update model to obtain the power hierarchy classification model.

[0060] In a specific implementation of this example, the pre-construction process of the hierarchical classification model described above is explained in detail. This example continues the aforementioned power system industry chain map in the field of wind power generation, constructing a classification model for predicting the industry chain level to which a map node belongs. First, based on the pre-constructed power system industry chain map in the field of wind power generation, this map contains a total of 120 map nodes, including the generation side, grid side, load side, and energy storage side, as well as 380 upstream and downstream relationship edges and technology reference relationship edges between map nodes. In the pre-construction process of the hierarchical classification model, the 120 map nodes in the power system industry chain map are first used as 120 network nodes of the graph neural network; at the same time, the historical power fusion feature vector corresponding to each map node (this vector is obtained by fusing the initial features of the historical power results data of the map node with the graph embedding vector through the aforementioned feature fusion method) is used as the initial feature of the corresponding network node, and each initial feature vector is a 256-dimensional real number vector.

[0061] Then, the neighbor relationships between multiple network nodes are determined based on the edge connections between multiple graph nodes in the power system supply chain graph. For example, for a graph node of a technology company, it has upstream supply relationship edges with wind turbine blade manufacturers, wind turbine gearbox manufacturers, and wind turbine converter manufacturers in the supply chain graph. Therefore, its neighbor nodes include the network nodes corresponding to the aforementioned entities.

[0062] The above-mentioned process for constructing the node feature update model includes: Step f1: Based on the propagation mechanism of Graph Convolutional Networks (GCNs), the core idea is to update the representation of the current network node by aggregating the features of neighboring nodes, while preserving the feature information of the network node itself. The formula consists of two parts: the first part is the feature transformation of the node itself. Through the graph neural network's first Layer weight matrix The first part linearly maps the current features of network nodes, then introduces non-linearity through an activation function, enabling the network model to learn the semantic information of the network nodes themselves. The second part aggregates the features of neighboring nodes. Using symmetric normalization coefficients A weighted summation of neighbor node features is performed. This coefficient, derived from the normalization of the Laplacian matrix in GCN, effectively alleviates the problem of excessively large eigenvalues ​​for high-degree nodes during aggregation, while maintaining the symmetry of the graph structure. The sum of the two parts achieves the fusion of node characteristics and neighbor topological information, resulting in a more accurate and comprehensive graph. Layer network nodes Update features It retains the original semantics of the nodes while incorporating the upstream and downstream relationships in the industry chain map.

[0063] Step f2: In choosing the activation function, ReLU (Rectified Linear Unit) can be selected. It has the advantages of simple calculation, mitigating gradient vanishing, and accelerating convergence, and is the most widely used in practical applications.

[0064] Step f3: The construction of the node feature update model is based on the following derivation logic: First, the graph nodes and their edge relationships in the industry chain graph are abstracted into graph structure data; second, drawing on the theoretical framework of GCN, the update method of graph node features is defined as a weighted sum of self-transformation and normalized neighbor aggregation; finally, through multi-level iterative stacking, each node can aggregate information from multi-hop neighbors, thereby capturing the deep structural features of indirect upstream and downstream connections in the industry chain. Through the above design, the node feature update model can effectively encode the structured knowledge of the industry chain graph into high-quality feature representations of nodes, providing information-rich input for subsequent hierarchical classification tasks.

[0065] Therefore, the construction process of the node feature update model satisfies the following formula:

[0066] In the above formula, The first character of the graph neural network is represented by the second character. Layer network nodes Update features, Indicates the first Layer network nodes initial characteristics, This represents the activation function of a graph neural network. N Represents a set of network nodes. Represents network nodes The set of neighboring nodes, Represents network nodes The set of neighboring nodes, Indicates the first Layer network nodes initial characteristics, The first character of the graph neural network is represented by the second character. Layer weight matrix.

[0067] In addition, the following criteria were designed to judge the classification accuracy of the power hierarchical classification model: the classification accuracy is judged by a combination of hierarchical classification accuracy, cross-entropy loss function, macro average F1 score and validation set performance monitoring. The hierarchical classification accuracy is calculated by weighting the classification results of each level, the cross-entropy loss is used as the main optimization objective for weighted summation, the macro average F1 score is used to balance the sample imbalance problem, and the validation set performance is used to dynamically monitor the model's generalization ability.

[0068] In addition, the termination conditions for the power hierarchy classification model iteration can include: termination when the preset maximum number of training rounds (e.g., 200 rounds) is reached; early termination when the validation set loss no longer decreases or the decrease is below a threshold for 10 consecutive rounds; early stopping and restoration of optimal parameters when the validation set accuracy does not exceed the historical best value for 15 consecutive rounds; and early termination when the gradient norm is below a preset minimum value. Through the above design, it is ensured that the model terminates iteration in a timely manner when it reaches optimal classification performance, balancing classification accuracy and training efficiency.

[0069] Step 5 above provides a detailed explanation of the process of classifying power achievement data based on industry chain hierarchical labels. Assume that 100 current power achievement data points have been collected and processed through the aforementioned steps, including different types such as technical reports, project documents, and industry standards. Each power achievement data point has had its industry chain hierarchical label predicted using the power hierarchical classification model, resulting in the following classification results: multiple generation-side hierarchical achievement data points, multiple grid-side hierarchical achievement data points, multiple load-side hierarchical achievement data points, and multiple energy storage-side hierarchical achievement data points. After obtaining the above classification results, based on the industry chain hierarchical label of each current power achievement data point, the multiple current power achievement data points are classified and organized. According to the four levels of generation-side, grid-side, load-side, and energy storage-side, the multiple power achievement data points are respectively categorized into corresponding category directories, forming a structured achievement database. For achievement data within the same level, further secondary classification can be performed according to technical sub-fields, time dimensions, achievement types, etc. Through the above classification process, the automated and hierarchical organization and management of a large amount of data on power system achievements have been realized. The classification results can be used to support technology retrieval and knowledge discovery in the power industry, serve the analysis of industrial technology trends to reveal the innovative development trends of different links, provide data support for policy making and resource allocation, and guide precise decision-making on scientific research investment and industrial layout.

[0070] In summary, the power system hierarchical classification method based on the industry chain graph provided in this embodiment collects current power achievement data, maps it to corresponding nodes in the power system industry chain graph, extracts the textual semantic features of the current power achievement data and the topological structural features of the graph nodes, and uses an attention mechanism to fuse them into a fused feature vector. This vector is then input into a hierarchical classification model built on a graph neural network, outputting the industry chain hierarchical label to which the achievement data belongs, thus achieving automated and refined classification of power achievement data. This method fully utilizes the structural information of the industry chain graph and the content information of the achievement data, effectively overcoming the classification bias problem caused by similar achievement content but different functional levels in traditional methods. It improves the consistency and accuracy of the industry chain semantics of the classification results, providing reliable technical support for power system achievement management, technology situation analysis, and industrial collaborative development.

[0071] Example 2: like Figure 4 As shown, the present invention also provides a power system hierarchical classification system based on the industry chain map, including: The data acquisition module is used to collect multiple current power output data located at the power system level; The graph node determination module is used to determine the graph node corresponding to each current power achievement data in a pre-constructed power system industry chain graph. The feature vector fusion acquisition module is used to extract features from each current power achievement data based on each current power achievement data, and generate an initial power data feature vector for each current power achievement data; and to fuse the initial power data feature vector with the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain the current power fusion feature vector for each current power achievement data. The hierarchical label acquisition module is used to take each current power fusion feature vector and use a pre-built power hierarchical classification model to output the industry chain hierarchical label to which each current power achievement data belongs; The classification module is used to classify the multiple current power achievement data according to the industry chain level tag to which each current power achievement data belongs.

[0072] Furthermore, the system also includes a power system supply chain mapping module, used for: Multiple power entities are identified from the collected historical power data at the power system level; these multiple power entities are then used as multiple graph nodes. Based on the multiple power entities, the upstream and downstream relationships between the multiple entities are determined according to their functional positioning or business process sequence in the power system; based on the upstream and downstream relationships between the multiple entities and the obtained technical reference relationships between the multiple power entities, multiple graph edges are constructed. The multiple graph nodes and multiple graph edges are integrated to form an initial power system industry chain graph; A graph embedding algorithm is used to perform representation learning on the initial power system industry chain graph to generate graph embedding vectors for multiple graph nodes; The graph embedding vector of each graph node is associated with the initial power system industry chain graph to form the final power system industry chain graph.

[0073] Furthermore, the fused feature vector acquisition module includes: The attention weight calculation module is used to calculate the attention weight between the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph. The power fusion feature vector acquisition module is used to perform cross-attention transformation on the initial power data feature vector and the graph embedding vector according to the attention weight, so as to obtain the current power fusion feature vector of each current power result data.

[0074] Furthermore, the attention weight calculation module is specifically used for: Using linear transformation technology, the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph are mapped to the same space to obtain the mapped initial power data feature vector and the mapped graph embedding vector. Calculate the similarity score between the mapped initial power data feature vector and the mapped graph embedding vector to obtain the vector similarity score; The vector similarity scores are normalized to obtain attention weights.

[0075] Furthermore, the current power fusion feature vector of each of the current power results data satisfies the following formula:

[0076] In the above formula, This represents the current power integration feature vector. This indicates normalization processing. Represents the initial power data feature vector. Represents a graph embedding vector. This represents the total number of multi-head attention heads in a linear transformation. Indicates the first k The learning fusion weights of individual attention heads This represents a cross-attention transformation based on attention weights.

[0077] Furthermore, the system also includes a hierarchical classification model construction module, used for: Multiple nodes in the power system industry chain map are used as multiple network nodes of the graph neural network; the historical power fusion feature vector corresponding to each map node is used as the initial feature of the corresponding network node. The neighbor relationships between multiple network nodes are determined based on the edge connection relationships between multiple nodes in the power system industry chain map. Based on the neighbor relationships, with the goal of fusing the topological information of its neighboring nodes with the feature representation of each network node, the initial features of multiple network nodes are iteratively updated according to the constructed node feature update model to obtain the power hierarchy classification model.

[0078] Furthermore, the construction process of the node feature update model satisfies the following formula:

[0079] In the above formula, The first character of the graph neural network is represented by the second character. Layer network nodes Update features, Indicates the first Layer network nodes initial characteristics, This represents the activation function of a graph neural network. N Represents a set of network nodes. Represents network nodes The set of neighboring nodes, Represents network nodes The set of neighboring nodes, Indicates the first Layer network nodes initial characteristics, The first character of the graph neural network is represented by the second character. Layer weight matrix.

[0080] Furthermore, the fused feature vector acquisition module is specifically used for: Based on each of the current power output data, multiple text segments are obtained; The multiple text portions are vectorized to obtain multiple text feature vectors; The multiple text feature vectors are fused to generate an initial power data feature vector for each current power achievement data.

[0081] Example 3: like Figure 5 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0082] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the power system hierarchical classification method based on the industry chain map in the above embodiments.

[0083] Example 4: Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the power system hierarchical classification method based on a supply chain map in the above embodiments.

[0084] Those skilled in the art will understand that embodiments of this invention can be provided as methods, systems, or computer program products. Therefore, this invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0085] This invention application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims pending approval.

Claims

1. A hierarchical classification method for power systems based on industry chain graphs, characterized in that, include: Collect multiple current power output data at the power system level; In the pre-constructed power system industry chain map, determine the map node corresponding to each of the current power achievement data; Based on each current power achievement data, feature extraction is performed on the current power achievement data to generate an initial power data feature vector for each current power achievement data; the initial power data feature vector is then fused with the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain the current power fusion feature vector for each current power achievement data. For each current power fusion feature vector, use a pre-built power hierarchy classification model to output the industry chain hierarchy label to which each current power achievement data belongs; The multiple current power achievement data are classified according to the industry chain level tag to which each current power achievement data belongs.

2. The method according to claim 1, characterized in that, The pre-construction process of the power system industry chain map includes: Multiple power entities are identified from the collected historical power data at the power system level; these multiple power entities are then used as multiple graph nodes. Based on the multiple power entities, the upstream and downstream relationships between the multiple entities are determined according to their functional positioning or business process sequence in the power system; based on the upstream and downstream relationships between the multiple entities and the obtained technical reference relationships between the multiple power entities, multiple graph edges are constructed. The multiple graph nodes and multiple graph edges are integrated to form an initial power system industry chain graph; A graph embedding algorithm is used to perform representation learning on the initial power system industry chain graph to generate graph embedding vectors for multiple graph nodes; The graph embedding vector of each graph node is associated with the initial power system industry chain graph to form the final power system industry chain graph.

3. The method according to claim 1, characterized in that, The process of fusing the initial power data feature vector with the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain the current power fusion feature vector for each current power achievement data includes: Calculate the attention weight between the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power output data in the power system industry chain graph; Based on the attention weights, a cross-attention transformation is performed on the initial power data feature vector and the graph embedding vector to obtain the current power fusion feature vector for each current power result data.

4. The method according to claim 3, characterized in that, The calculation of the attention weight between the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power output data in the power system industry chain graph includes: Using linear transformation technology, the initial power data feature vector and the graph embedding vector of the graph node corresponding to the current power result data in the power system industry chain graph are mapped to the same space to obtain the mapped initial power data feature vector and the mapped graph embedding vector. Calculate the similarity score between the mapped initial power data feature vector and the mapped graph embedding vector to obtain the vector similarity score; The vector similarity scores are normalized to obtain attention weights.

5. The method according to claim 4, characterized in that, The current power fusion feature vector of each of the current power results data satisfies the following formula: In the above formula, This represents the current power integration feature vector. This indicates normalization processing. Represents the initial power data feature vector. Represents a graph embedding vector. This represents the total number of multi-head attention heads in a linear transformation. Indicates the first k The learning fusion weights of individual attention heads This represents a cross-attention transformation based on attention weights.

6. The method according to claim 1, characterized in that, The pre-construction process of the hierarchical classification model includes: Multiple nodes in the power system industry chain map are used as multiple network nodes of the graph neural network; the historical power fusion feature vector corresponding to each map node is used as the initial feature of the corresponding network node. The neighbor relationships between multiple network nodes are determined based on the edge connection relationships between multiple nodes in the power system industry chain map. Based on the neighbor relationships, with the goal of fusing the topological information of its neighboring nodes with the feature representation of each network node, the initial features of multiple network nodes are iteratively updated according to the constructed node feature update model to obtain the power hierarchy classification model.

7. The method according to claim 6, characterized in that, The construction process of the node feature update model satisfies the following formula: In the above formula, The first character of the graph neural network is represented by the second character. Layer network nodes Update features, Indicates the first Layer network nodes initial characteristics, This represents the activation function of a graph neural network. N Represents a set of network nodes. Represents network nodes The set of neighboring nodes, Represents network nodes The set of neighboring nodes, Indicates the first Layer network nodes initial characteristics, The first character of the graph neural network is represented by the second character. Layer weight matrix.

8. The method according to claim 1, characterized in that, The step of extracting features from each current power output data point to generate an initial power data feature vector for each current power output data point includes: Based on each of the current power output data, multiple text segments are obtained; The multiple text portions are vectorized to obtain multiple text feature vectors; The multiple text feature vectors are fused to generate an initial power data feature vector for each current power achievement data.

9. A hierarchical classification system for power systems based on industry chain graphs, characterized in that, include: The data acquisition module is used to collect multiple current power output data located at the power system level; The graph node determination module is used to determine the graph node corresponding to each current power achievement data in a pre-constructed power system industry chain graph. The feature vector fusion acquisition module is used to extract features from each current power achievement data based on each current power achievement data, and generate an initial power data feature vector for each current power achievement data; and to fuse the initial power data feature vector with the graph embedding vector of the graph node corresponding to the current power achievement data in the power system industry chain graph to obtain the current power fusion feature vector for each current power achievement data. The hierarchical label acquisition module is used to take each current power fusion feature vector and use a pre-built power hierarchical classification model to output the industry chain hierarchical label to which each current power achievement data belongs; The classification module is used to classify the multiple current power achievement data according to the industry chain level tag to which each current power achievement data belongs.

10. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a power system hierarchical classification method based on the industry chain map as described in any one of claims 1 to 8 is implemented.

Images