Construction of edge-center two-level knowledge graph, computing power evaluation methods and systems
By constructing a two-level knowledge graph of edge and center, the problem of inaccurate computing power assessment in existing technologies is solved, and standardized assessment and overall scheduling of AI models in the power system are realized, meeting the needs of three-network coordination.
Patent Information
- Application Number
- CN202611438246.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-17
AI Technical Summary
The lack of a standardized computing power evaluation system in existing technologies makes it impossible to accurately quantify the hardware consumption of different AI models, which restricts the overall scheduling of computing power under the collaboration of three networks.
A two-level knowledge graph of edge and center is constructed. Entities and relationships are extracted from multi-source heterogeneous data, divided into edge and center fragments, and entity alignment and knowledge fusion are performed. Constraints are embedded to construct a standardized set of triples and store them in a graph database. Computing power requirements are assessed in combination with knowledge graph technology.
It provides a standardized computing power evaluation system, accurately quantifies the hardware consumption of different AI models, realizes the overall scheduling of computing power under the collaboration of three networks, and solves the semantic gap between supply and demand on the business side and the computing power side.
Smart Images

Figure CN123115433A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer software and computing power scheduling technology, specifically to the construction of edge-center two-level knowledge graphs, computing power evaluation methods and systems. Background Technology
[0002] Against the backdrop of the collaborative construction of new power systems and the three networks of the power grid (communication network, computing network, and digital network), the digitalization of power services is expanding rapidly. AI applications such as intelligent inspection of power grid equipment, new energy forecasting, dispatch voice recognition, and power risk control auditing are being implemented on a large scale. Multiple AI models with architectures such as CNN, Transformer, and MoE are being widely deployed. The demand for power computing power is characterized by its diverse types, heterogeneity, high reliability, and low latency. While power grid personnel are familiar with power regulations and business logic, they lack knowledge of AI models, GPU / NPU hardware, memory bandwidth, P95 latency, TTFT, and other computing power metrics. This makes it difficult for them to translate business constraints such as high reliability and millisecond-level response into standardized computing power parameters during the computing power resource application process, creating a semantic gap between supply and demand on the business side and the computing power side. Traditional computing power assessment relies on manual experience to fill in forms for estimation, which can easily lead to excessive idle or redundant computing resources or insufficient configuration, causing business lag and timeout failures. At the same time, there is no unified baseline library of computing power performance in the power scenario in the current technology, and there is a lack of standardized computing power assessment system. It is impossible to accurately quantify the hardware consumption of different AI models, which restricts the overall scheduling of computing power under the three-network synergy. The present invention was developed in response to this industry pain point.
[0003] Existing similar solutions fall into two categories: general commercial computing power assessment solutions and existing computing power baseline solutions in the domestic power industry. The closest to existing technology is the computing power assessment solution that relies on model-scenario mapping forms and manual baseline lookup (a combination of the domestic computing power metropolitan area network white paper implementation solution and the Azure AI Gallery assessment solution).
[0004] Existing table-based computing power assessments rely on static Excel baseline tables, requiring business personnel to manually fill in parameters and engineers to manually select hardware. This lack of a standardized computing power assessment system makes it impossible to accurately quantify the hardware consumption of different AI models, hindering the coordinated scheduling of computing power under the three-network collaboration model. Summary of the Invention
[0005] To address the lack of a standardized computing power evaluation system in existing technologies, which hinders the accurate quantification of hardware consumption for different AI models and restricts the coordinated scheduling of computing power under the three-network collaboration, this invention proposes a method for constructing an edge-center two-level knowledge graph, including: Extract entities and relationships between entities from multi-source heterogeneous data to generate an initial set of triples; The initial set of triples is divided into edge shard triples and center shard triples according to the hardware deployment level, and cross-shard connections are established. Entity alignment and knowledge fusion are performed on the edge segment triples, the center segment triples, and the cross-segment connection relationships to obtain a standardized triple set. The standardized triple set is imported into a graph database, stored in the form of nodes and edges, and constraints are embedded to construct an edge-center two-level knowledge graph.
[0006] Preferably, the step of dividing the initial set of triples into edge shard triples and center shard triples according to the hardware deployment level, and establishing cross-shard connections, includes: The triplet consisting of localized power services of substations and distribution area edge nodes, lightweight edge AI models, edge NPU computing power performance, and local low latency constraints is divided into edge sharding triplets. The triplet consisting of cross-regional power business under the overall network scheduling, cloud-based MoE large model, GPU cluster computing power indicators, unified computing power baseline of the entire network, and compliance specifications for three-network collaboration is divided into central sharding triplets. Assign a first cross-shard identifier to each entity in each edge shard triplet, and assign the same second cross-shard identifier to entities in each center shard triplet that have a business relationship with that entity; based on the consistency matching of the first cross-shard identifier and the second cross-shard identifier, establish an index association relationship between edge shards and center shards, and generate a shard association mapping table.
[0007] Preferably, the step of performing entity alignment and knowledge fusion on the edge fragment triples, the center fragment triples, and the cross-fragment connection relationships to obtain a standardized triple set includes: Local terminology unification, entity disambiguation and data deduplication are performed on edge-segmented triples to standardize the computing power parameters corresponding to fuzzy constraints in power business. The central sharding triplet integrates the multi-source computing power baseline data, model parameters, and hardware indicators of the entire network, supplements the P95 tail latency, TTFT, and TPOT long-tail performance indicators, and corrects the adaptation deviation between the general computing power baseline and the power business scenario. Based on cross-shard identifiers, shard consistency verification is performed, and the fuzzy business requirements of the power business side are automatically mapped into standardized computing power technical indicators, forming a standardized triplet set containing five layers: "business scenario - AI model - computing power hardware - performance indicators - compliance constraints".
[0008] Preferably, the step of importing the standardized triple set into a graph database, storing it in the form of nodes and edges, and embedding constraints to construct an edge-center two-level knowledge graph includes: The edge-sharded triples in the standardized triple set are stored in a lightweight graph database; The central sharded triples in the standardized triple set are stored using an enterprise-level graph database. Using entities as nodes and the relationships between entities as edges, an edge knowledge graph is constructed from edge-sharded triples. Using entities as nodes and the relationships between entities as edges, a central knowledge graph is constructed from central partitioned triples. The cross-segment identifiers between the edge and center are used as the association between the edge knowledge graph and the center knowledge graph. Combined with constraints, a two-level edge-center knowledge graph is constructed.
[0009] Preferably, it further includes: Based on the actual load test data of real load test conditions that are iteratively updated based on the computing power performance baseline library, the relationship between entities and parameter indicators stored in the edge-center secondary knowledge graph is compared in real time, and deviation items between theoretical graph rules and real load test conditions are automatically identified. Based on the relationship between the deviation term and the triple, the semantic association weights, constraint threshold ranges, and business adaptation mapping relationships are automatically and iteratively corrected, and the model-level adaptation rules and hardware performance matching boundaries are updated synchronously.
[0010] Furthermore, this invention also provides a guided computing power requirement standard evaluation method based on knowledge graph technology, including: The algorithm model entities are obtained by parsing input keywords from natural language describing business requirements and matching them with a pre-built edge-center two-level knowledge graph. An interactive slot-filling mechanism is used to guide the completion of entity information in the algorithm model. By integrating the input keywords and the completed algorithm model entity information, a structured business requirement profile that can be recognized by a machine is generated. Based on demand profiles, a preliminary set of candidate platforms for the algorithm model is obtained by retrieving the edge-center secondary knowledge graph. The suitability of algorithm models in the candidate platform set is evaluated based on the entity association relationships in the knowledge graph; Among them, the edge-center two-level knowledge graph is constructed based on the edge-center two-level knowledge graph construction method described above.
[0011] Preferably, the evaluation of the algorithm model suitability in the candidate platform set based on entity relationships in the knowledge graph includes: Based on the relationships and attributes between entities in the knowledge graph, model adaptation score, concurrency adaptation score, latency adaptation score and throughput adaptation score are respectively applied to each combination in the candidate platform set. Based on the model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score, the comprehensive matching accuracy of each combination is calculated by weighted summation. The scoring rules for model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score are determined based on the hardware physical upper limit attribute, communication link measured performance attribute, and model running baseline attribute stored in the knowledge graph. This is a quantitative mapping of hardware physical constraints and communication physical constraints.
[0012] Furthermore, this application also provides a guided computing power requirement standard evaluation system based on knowledge graph technology, including: The knowledge reasoning layer is used to parse input keywords from natural language describing business requirements and match them with a pre-built edge-center two-level knowledge graph to obtain the algorithm model entity. The interactive guidance layer is used to guide the completion of entity information in the algorithm model using an interactive slot filling mechanism based on the entity in the algorithm model. The matching calculation layer is used to integrate the input keywords and the completed algorithm model entity information to generate a machine-recognizable structured business requirement profile; and based on the requirement profile, it retrieves the edge-center secondary knowledge graph to obtain a preliminary candidate platform set for the algorithm model. The results output layer is used to verify and score the suitability of algorithm models in the candidate platform set based on entity associations in the knowledge graph, and to quantify the accuracy of computing power selection and matching. Among them, the edge-center two-level knowledge graph is constructed based on the edge-center two-level knowledge graph construction method described above.
[0013] Preferably, the knowledge reasoning layer is specifically used for; Natural language is parsed using named entity recognition and semantic role labeling techniques to obtain input keywords; Based on the input keywords, the system retrieves the corresponding business scenario entities in the pre-constructed edge-center two-level knowledge graph, and then matches them with the entities in the algorithm model.
[0014] Preferably, the input keywords include: business scenario, processing object, and business task.
[0015] Preferably, the matching calculation layer is specifically used for: The input keywords and the completed algorithm model entity information are integrated together and transformed into a structured business requirement profile that can be recognized by machines. Based on the model classification and business hardware constraints in the demand profile, the model and hardware relationship within the edge-center secondary knowledge graph is retrieved to recall a preliminary candidate platform set. The candidate platform set is obtained by removing platforms that clearly violate latency and concurrency hardware constraints from the initial candidate platform set.
[0016] Preferably, the result output layer is specifically used for: Based on the relationships and attributes between entities in the knowledge graph, model adaptation score, concurrency adaptation score, latency adaptation score and throughput adaptation score are respectively applied to each combination in the candidate platform set. Based on the model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score, the comprehensive matching accuracy of each combination is calculated by weighted summation. The scoring rules for model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score are determined based on the hardware physical upper limit attribute, communication link measured performance attribute, and model running baseline attribute stored in the knowledge graph. This is a quantitative mapping of hardware physical constraints and communication physical constraints.
[0017] In another aspect, this application also provides an electronic device, comprising: at least one processor and a memory; 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, a method for constructing an edge-center two-level knowledge graph and a guided computing power requirement standard evaluation method based on knowledge graph technology are implemented as described above.
[0018] Furthermore, this application also provides a computer-readable storage medium having an executable program stored thereon, which, when executed, implements the edge-center two-level knowledge graph construction method and the guided computing power demand standard evaluation method based on knowledge graph technology as described above.
[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method for constructing an edge-center two-level knowledge graph. This knowledge graph addresses the shortcomings of traditional graphs, such as fixed parameters, poor adaptability, and inability to meet the compliance constraints of grid edge deployments. Furthermore, this invention provides a guided computing power demand standard evaluation method based on knowledge graph technology. This method parses input keywords from natural language describing business needs and matches them with the pre-constructed edge-center two-level knowledge graph to obtain algorithm model entities. Based on these entities, an interactive slot-filling mechanism guides the completion of algorithm model entity information. The method integrates the input keywords and the completed algorithm model entity information to generate a machine-recognizable structured business demand profile. Based on this demand profile, the method retrieves the edge-center two-level knowledge graph to obtain a preliminary candidate platform set for the algorithm model. Finally, based on the entity relationships in the knowledge graph, the method verifies and scores the adaptability of the algorithm models in the candidate platform set, quantifying the accuracy of computing power selection and matching. This method bridges the semantic gap between supply and demand on the business side and the computing power side in existing technologies, provides a standardized computing power evaluation system, accurately quantifies the hardware consumption of different AI models, and achieves coordinated scheduling of computing power under the three-network collaboration model. Attached Figure Description
[0020] Figure 1 This is a flowchart of the guided computing power demand standard evaluation method based on knowledge graph technology of the present invention. Figure 2 This is a flowchart illustrating the overall process of the knowledge graph-based guided computing power demand assessment method of the present invention. Figure 3 This is a schematic diagram of an electronic device structure according to the present invention. Detailed Implementation
[0021] To better understand the present invention, the following description, in conjunction with the accompanying drawings and embodiments, will further illustrate the content of the present invention.
[0022] The following section introduces some well-known basic technologies involved in the technical solution of this application.
[0023] (1) General Specifications for Evaluating Power AI Models The China Electrotechnical Society released T / CES2307, "Specification for the Construction of Scenario-based Verification and Evaluation System for Power Artificial Intelligence Models." This standard establishes model evaluation criteria based on functionality, efficiency, robustness, and security, distinguishes between assessment indicators for classification, target detection, and NLP models, and clarifies two major computing power-related indicators: time efficiency (response latency) and space efficiency (memory usage). It serves as a general industry benchmark for quantifying computing power indicators.
[0024] (2) Commonly known principles of computing power characteristics of mainstream AI models CNN, Transformer, and MoE are commonly used AI models in power scenarios, and the computing power bottlenecks of each model are different.
[0025] (3) General technologies for computing power benchmark testing The international MLPerf and NIST benchmark evaluation frameworks use batch size, input precision, and concurrency as variables to measure hardware throughput, latency, and energy efficiency. By fitting hardware performance curves through batch sampling, it is a general testing method for building performance baselines.
[0026] (4) Basic Theories of Knowledge Graphs Knowledge graphs consist of entities, relations, and triples. Relying on mature technologies such as NER entity extraction, semantic role labeling, K-hop neighborhood expansion, TransE embedding reasoning, and SHACL constraint verification, they can achieve cross-domain entity association and missing field reasoning.
[0027] (5) Computing power matching conventional scoring logic Existing resource matching generally adopts a multi-dimensional weighted scoring model, assigning segmented values according to the degree of matching between business indicators and hardware capabilities, which is used to rank candidate resources.
[0028] Example 1: A method for constructing an edge-center two-level knowledge graph, such as Figure 1 As shown, it includes: Step S1: Extract entities and relationships between entities from multi-source heterogeneous data to generate an initial set of triples; Step S2: Divide the initial set of triples into edge shard triples and center shard triples according to the hardware deployment level, and establish cross-shard connection relationships; Step S3: Perform entity alignment and knowledge fusion on the edge fragment triples, the center fragment triples, and the cross-fragment connection relationships to obtain a standardized triple set; Step S4: Import the standardized triple set into a graph database, store it in the form of nodes and edges, and embed constraints to construct an edge-center two-level knowledge graph.
[0029] The following is a further explanation of each step: Before step S1, the construction of the computing power performance baseline library is introduced, and the construction process includes: Select test subjects and choose three mainstream AI models in the power industry; Standardized stress test experiments were conducted based on the test objects and the three mainstream AI models selected in the power industry to collect raw stress test data. Standardized data is obtained by cleaning, removing outliers, and standardizing features from the raw stress test data; The correlation between performance indicators and parameters is obtained by fitting standardized data using polynomials; A performance baseline library is constructed based on the correlation between the performance metrics and parameters.
[0030] Furthermore, the relationship between performance indicators and parameters here specifically refers to the relationship between the selected AI model parameters and the computing power performance indicators of the AI model in a given scenario. These performance indicators include: P95 tail latency, TTFT, and TPOT long tail performance indicators.
[0031] Step S1: Extract entities and relationships between entities from multi-source heterogeneous data to generate an initial set of triples, including: First, the multi-source heterogeneous data sources for this data collection were identified, specifically covering four categories: core power grid business data, AI model data, computing power hardware performance data, and compliance constraint data. These categories include: 1. Power Business Data: Unstructured text and structured work order data related to AI business processes such as intelligent power grid equipment inspection, new energy forecasting, dispatch voice recognition, and power risk control auditing, including business procedures, SLA constraints, concurrency requirements, millisecond-level low latency, and high reliability operation requirements; 2. AI Model Data: Model specifications, batch parameters, inference logic, and business adaptation scenario data for multi-architecture models such as CNN, Transformer, and MoE deployed on a large scale in the power grid; 3. Computing Power Hardware Data: GPU / NPU hardware parameters, memory bandwidth, theoretical FLOPS, measured throughput, average latency, P95 tail latency, TTFT, TPOT, and other detailed computing power performance indicators; 4. Industry Baseline Data: MLPerf standardized stress test data, baseline data from the power computing power metropolitan area network white paper, and Azure AI Gallery computing power assessment reference data. The above baseline data can be obtained from the constructed computing power performance baseline library.
[0032] Secondly, an entity and relation extraction model adapted to the power computing power scenario is adopted to complete structured knowledge extraction. Entity recognition is performed using the BERT-BiLSTM-CRF model (a BERT-based bidirectional long short-term memory network-conditional random field model), accurately extracting three core entity categories: business layer entities (intelligent inspection, new energy prediction, dispatch speech recognition, power risk control auditing, etc.), model layer entities (CNN models, Transformer models, MoE hybrid expert models, etc.), and computing power layer entities (GPU / NPU hardware, video memory, bandwidth, P95 latency, TTFT, concurrent computing power, etc.). Simultaneously, a remotely supervised relation extraction algorithm is used to mine inherent semantic relationships between entities. Core relationships include: business-adapted models, model-matched computing power, computing power corresponding performance indicators, indicator constraints on business SLAs, and hardware-supported model inference.
[0033] Finally, the extracted entities and relations are uniformly encapsulated into a standard triple format of <head entity, relation, tail entity>, summarizing all power computing power business, model, hardware, and indicator data to generate an initial triple set containing tens of thousands of raw data entries. This initial triple set covers all dimensions of power computing power assessment, differing from the limitations of traditional static Excel spreadsheets that only store average latency and fixed batch performance data. For the first time, it incorporates essential sub-indicators for high-reliability power grid services, such as P95 tail latency, TTFT, and TPOT, providing a complete data foundation for subsequent accurate computing power assessment.
[0034] Step S2: Divide the initial triple set into edge shard triples and center shard triples according to the hardware deployment level, and establish cross-shard connections, including: The triplet consisting of localized power services of substations and distribution area edge nodes, lightweight edge AI models, edge NPU computing power performance, and local low latency constraints is divided into edge sharding triplets. The triplet consisting of cross-regional power business under the overall network scheduling, cloud-based MoE large model, GPU cluster computing power indicators, unified computing power baseline of the entire network, and compliance specifications for three-network collaboration is divided into central sharding triplets. Assign a first cross-shard identifier to each entity in each edge shard triplet, and assign the same second cross-shard identifier to entities in each center shard triplet that have a business relationship with that entity; based on the consistency matching of the first cross-shard identifier and the second cross-shard identifier, establish an index association relationship between edge shards and center shards, and generate a shard association mapping table; wherein, the shard association mapping table is the cross-shard connection relationship.
[0035] This step addresses the pain points of traditional graph parameters being fixed and having poor adaptability, making them unsuitable for compliant deployment at the edge of the power grid, as well as the global fixation of existing static baseline forms, which are unsuitable for distributed computing power scheduling. Combining the architectural characteristics of the power grid's three-network collaboration of "edge deployment and central coordination," this step completes knowledge sharding and layering, solving the problems of redundant computing resources and unbalanced configuration.
[0036] First, based on the physical deployment level of the power grid hardware and the computing power operation scenarios, differentiated sharding rules are formulated to accurately split the initial triplet set. Among them, the edge sharding triplet focuses on the computing power requirements of local edge services in the power grid, mainly including localized AI business knowledge of each substation and distribution area edge node, lightweight CNN / Transformer (i.e., convolutional neural network / transformer) model parameters deployed on the terminal, edge NPU hardware performance, millisecond-level latency constraints of local services, and high-concurrency operation requirements on site. This triplet is adapted to the low-latency and real-time computing power assessment business scenarios of the power grid edge, supporting the rapid selection of local edge computing power.
[0037] The central segmented triplet focuses on the overall network-wide computing power scheduling. It mainly includes global knowledge triplets such as network-wide general MoE large model parameters, cloud GPU cluster computing power indicators, network-wide unified computing power baseline standards, three-network collaborative computing power coordination rules, power industry-wide general compliance constraints, and cross-regional computing power scheduling specifications. It is deployed on the power grid intelligent computing center cloud platform and is responsible for the aggregation, standardized management and control and global scheduling of network-wide computing power resources.
[0038] Secondly, cross-shard connections are constructed to establish semantic relationships between edge local knowledge and central global knowledge. Addressing the shortcomings of existing technologies that lack semantic association and rely entirely on manual matching, a cross-shard index association mechanism is established, consisting of "edge business - central model - global computing power baseline." A first cross-shard identifier is assigned to each entity in each edge shard triplet, and the same second cross-shard identifier is assigned to entities in each central shard triplet that have a business relationship with that entity. Based on the consistency matching of the first and second cross-shard identifiers, an index association relationship is established between edge shards and central shards, generating a shard association mapping table.
[0039] The relationships include: business scenario → depends on → algorithm model, algorithm model → deployed on → computing hardware, business scenario → constraints → latency / concurrency / throughput metrics; the three major business entities include business scenario entity, algorithm model entity and computing hardware entity.
[0040] Specific relationships include: the baseline of the general Transformer model associated with the edge local intelligent inspection business, the MoE large model computing power index associated with the edge new energy prediction business, and the network-wide SLA compliance standard associated with the edge latency constraint parameters. Simultaneously, a sharding association mapping table is generated to record the correspondence between all edge and center shards, achieving linkage and matching between local business needs and global computing power standards, completely resolving the problems of large deviations and lack of semantic association in manual matching.
[0041] Step S3: Perform entity alignment and knowledge fusion on the edge fragment triples, the center fragment triples, and the cross-fragment connection relationships to obtain a standardized triple set, including: Local terminology unification, entity disambiguation and data deduplication are performed on edge-segmented triples to standardize the computing power parameters corresponding to fuzzy constraints in power business. The central sharding triplet integrates multi-source computing power baseline data, supplements P95 tail latency, TTFT, and TPOT long-tail performance indicators, and corrects the adaptation deviation between the general computing power baseline and the power business scenario. Based on cross-shard identifiers, shard consistency verification is performed, and the fuzzy business requirements of the power business side are automatically mapped and converted into standardized computing power technical indicators, resulting in a standardized set of triples.
[0042] Step S3 addresses the pain points of existing technologies, such as missing business parameters, semantic ambiguity, inconsistent baseline data, and large deviations in manual correction. It resolves the semantic gap between power business personnel and computing power technicians, eliminates fragmented knowledge conflicts, and forms a standardized and reusable power computing power knowledge system.
[0043] First, entity alignment and deduplication are performed within each shard. For edge shards, business terminology, device naming, and latency metric definitions are standardized across edge sites to eliminate ambiguity in localized expressions. For example, the computing power technical parameters corresponding to vague business descriptions such as "millisecond-level response" and "high-reliability operation" are standardized. For central shards, multi-source baseline data, model parameters, and hardware metrics across the entire network are deduplicated and integrated. Deviations between the MLPerf general baseline and the actual scenarios in the power industry are corrected, and missing P95 tail latency, TTFT, and TPOT sub-metrics in traditional static forms are supplemented to construct a complete power-specific computing power baseline knowledge system.
[0044] Secondly, complete the consistency fusion verification of cross-shard connection relationships. Relying on cross-shard identifiers, verify the matching consistency between edge business requirements and central computing power standards, and automatically correct parameter conflicts caused by manual experience estimation. Addressing the pain point that business personnel cannot convert standardized computing power parameters, achieve automatic semantic conversion through knowledge fusion: automatically map the fuzzy business constraints of "high reliability, millisecond-level response, and high concurrency" on the power grid business side into standardized computing power technical parameters such as P95 latency threshold, TTFT inference time, GPU memory usage, and concurrent computing power specifications, completely bridging the semantic gap between business and computing power.
[0045] Finally, a unified knowledge standard is established, and a standardized set of triples is output. Industry-standard terminology for all entities, relationships, and metrics is unified, and invalid, conflicting, and redundant knowledge is eliminated. This forms a triple dataset containing five layers of standardized associations: "business scenario - AI model - computing hardware - performance metrics - compliance constraints." This addresses the problems of traditional solutions, such as the lack of a standardized evaluation system, missing parameters, and long evaluation cycles, providing standardized knowledge support for subsequent quantitative computing power evaluation and intelligent matching.
[0046] Step S4: Import the standardized triple set into a graph database, store it in the form of nodes and edges, embed constraints, and construct an edge-center two-level knowledge graph, including: The edge-sharded triples in the standardized triple set are stored in a lightweight graph database; The central sharded triples in the standardized triple set are stored using an enterprise-level graph database. Using entities as nodes and the relationships between entities as edges, an edge knowledge graph is constructed from edge-sharded triples. Using entities as nodes and the relationships between entities as edges, a central knowledge graph is constructed from central partitioned triples. The cross-segment identifiers between the edge and center are used as the association between the edge knowledge graph and the center knowledge graph. Combined with constraints, a two-level edge-center knowledge graph is constructed.
[0047] The constraints include: SHACL shape constraint verification rules and SWRL semantic logic reasoning rules, while also integrating K-hop neighborhood expansion and TransE vector reasoning capabilities.
[0048] The SHACL shape constraint verification rules include: parameter existence constraint rules, numerical range constraint rules, and business-model linkage constraint rules. The SWRL semantic logic reasoning rules include model size-resource consumption reasoning rules, high concurrency-latency linkage reasoning rules, business level-fault tolerance threshold reasoning rules, and operating condition accuracy-throughput matching reasoning rules.
[0049] When constructing the secondary knowledge graph, based on the fragmented association mapping table generated in step S2, cross-graph association nodes are created in the enterprise-level graph database of the central knowledge graph to achieve graph-level semantic association between the edge knowledge graph and the central knowledge graph. The specific steps are as follows: The first step is to read the mapping table records. Traverse the shard association mapping table. Each record contains the first cross-shard identifier, the second cross-shard identifier, the edge node ID, the center node ID, and the business association type.
[0050] The second step is to create cross-graph association nodes. For each record in the mapping table, create a cross-graph association node in the central graph library with a node type of "CrossPartitionLink". Its attributes include cross-partition identifier (take the value of the first cross-partition identifier), edge node ID, central node ID, and business association type.
[0051] The third step is to establish the connecting edges. Two directed edges are established for this cross-graph connecting node: the first edge, with the relation type "links_to_edge", points to the edge node in the edge knowledge graph identified by the edge node ID; the second edge, with the relation type "links_to_center", points to the center node in the center knowledge graph identified by the center node ID.
[0052] The fourth step is cross-graph query. When a cross-graph query is required, start from a node in the edge knowledge graph, traverse backwards along the "links_to_edge" edge to the cross-graph related node, and then traverse forwards along the "links_to_center" edge to the corresponding node in the central knowledge graph. This enables semantic association retrieval and reasoning between edge local business needs and the central global computing power baseline.
[0053] Step S4 addresses the pain points of traditional computing power assessment, such as lack of quantitative standards, subjective manual selection, high resource redundancy, and unreproducible solutions. By combining the constraints of compliant deployment at the power grid edge, it completes the implementation of a two-level graph, making computing power assessment interpretable, quantifiable, and reproducible.
[0054] First, a distributed graph database deployment architecture is adopted, with standardized triples stored in shards. At the edge, a lightweight graph database is used to store edge shard knowledge, supporting offline, low-latency computing power assessment of power grid edge nodes. At the central end, an enterprise-grade graph database is used to store global central shard knowledge, enabling unified management and coordinated resource scheduling of the entire network's computing power baseline. All standardized entities are used as graph nodes, and the standardized semantic relationships between entities are used as graph edges, constructing a structured and visualized power computing power knowledge network.
[0055] Secondly, it embeds power-specific constraints and quantitative evaluation rules to compensate for the shortcomings of traditional static forms that lack quantitative standards and ranking mechanisms. First, it embeds power grid compliance deployment constraints, binding the three-network collaborative construction specifications, power computing power SLA standards, and edge data security management rules to graph nodes and edges. Second, it embeds segmented weighted scoring rules and a Top-K candidate solution ranking mechanism, establishing differentiated computing power matching weights for different AI models (CNN, Transformer, MoE) to achieve quantitative comparison and ranking of multiple computing power configuration schemes. Third, it embeds long-tail latency constraint rules, using P95 tail latency, TTFT, and TPOT indicators as core constraints to replace the traditional single average latency evaluation standard, adapting to the power grid's high reliability and strict control of long-tail faults business requirements.
[0056] Finally, the edge-center two-level knowledge graph was completed and its functions were implemented. This graph possesses core capabilities such as automatic semantic association, intelligent requirement completion, quantitative matching of computing power, and traceable and reproducible solutions, completely solving the shortcomings of existing technologies: by replacing manual full table filling with knowledge-guided interaction, the parameter omission rate is significantly reduced; by replacing manual table lookup estimation with graph semantic association, the redundancy of computing resources is significantly reduced from 60%; and through standardized baselines and quantitative scoring mechanisms, the accuracy, interpretability, and reproducibility of multi-model computing power evaluation are achieved, fully supporting the unified scheduling of computing power and the efficient construction of intelligent computing centers under the three-network synergy of the new power system.
[0057] The present invention provides a method for constructing an edge-center two-level knowledge graph, which further includes: Based on the actual load test data of real load test conditions that are iteratively updated based on the computing power performance baseline library, the relationship between entities and parameter indicators stored in the edge-center secondary knowledge graph is compared in real time, and deviation items between theoretical graph rules and real load test conditions are automatically identified. Based on the relationship between the deviation term and the triple, the semantic association weights, constraint threshold ranges, and business adaptation mapping relationships are automatically and iteratively corrected, and the model-level adaptation rules and hardware performance matching boundaries are updated synchronously.
[0058] When the deviation between a certain edge segment triplet (intelligent inspection, requirement P95, 180ms) and the baseline library's measured P95=207ms with the same model and hardware is >15%: (a) The semantic association weight of the edge corresponding to the triple (intelligent inspection - adaptation - lightweight CNN) is changed from 0.8 to η'=η×(1 The formula 0.5×Δ / Δ0) is corrected to 0.74, where η' is the corrected semantic association weight, η is the original semantic association weight, Δ is the absolute value of the deviation between the measured value and the map rule value, and Δ0 is the threshold for deviation to trigger correction, with the same unit as Δ. In the formula, 0.5 represents the correction rate coefficient (attenuation factor), which controls the amplitude of each correction. It is a dimensionless constant with a value range of (0, 1], and its function is to prevent the map from oscillating due to excessive amplitude of a single correction. When Δ≤Δ0 (the measured deviation is within the tolerance range), Δ / Δ0≤1. However, the correction action is usually set to trigger only when Δ>Δ0. Therefore, when Δ≤Δ0, the formula is not executed and the weight remains unchanged.
[0059] When Δ > Δ0 (the measured deviation exceeds the tolerance threshold): Δ / Δ0 > 1. For example, if Δ = 30ms and Δ0 = 20ms, then Δ / Δ0 = 1.5, η' = η × 0.25, and the weight is reduced from the original η to 0.25η, which greatly weakens the confidence of the association.
[0060] The purpose of this formula is to automatically reduce the weight of the correlation when there is a large deviation between the measured stress test data and the theoretical rules stored in the graph. This reduces the reliance on the unreliable correlation in the next graph inference, until subsequent stress test data proves that the correlation has returned to stability, and then the weight is gradually restored through reverse correction.
[0061] (b) Modify the P95 threshold range in the SHACL shape constraint from (0, 200ms] to (0, 230ms], and write it into the fragment association mapping table version number v2; (c) Replace the "High Reliability" → "P95≤200ms" entry in the business adaptation mapping table with "P95≤230ms"; (d) Reload the above changes through the SWRL rule engine and synchronously update the attributes of the central shard (intelligent inspection global SLA) node.
[0062] The edge-center two-level knowledge graph constructed in this application can produce the following effects: The solution adopts a two-level split storage of initial triplets: edge sharding and central sharding. Knowledge data is physically / logically separated. The edge side only carries the shards required by its own business, while the central side stores global knowledge, which is no longer a single and fixed global graph. The sharded architecture is naturally adapted to the grid edge-center distributed deployment mode. The edge can run local graphs independently, while the center completes global fusion, improving the ability to adapt to different scenarios. Additional configurations, such as cross-shard connectivity, fusion alignment, graph database storage, and embedded constraints, can bind power grid compliance rules as constraints to the graph, meeting the compliance management requirements of edge deployments.
[0063] Example 2 Guided computing power requirement assessment methods based on knowledge graph technology include: Step 1: Parse the input keywords from the natural language describing the business requirements and match them with the pre-built edge-center two-level knowledge graph to obtain the algorithm model entities; Step 2: Based on the algorithm model entity, an interactive slot filling mechanism is used to guide the completion of the algorithm model entity information; Step 3: Integrate the input keywords and the completed algorithm model entity information to generate a machine-recognizable structured business requirement profile; Step 4: Based on the demand profile, retrieve the edge-center secondary knowledge graph to obtain a preliminary set of candidate platforms for the algorithm model; Step 5: Evaluate the suitability of algorithm models in the candidate platform set based on entity relationships in the knowledge graph; Among them, the edge-center two-level knowledge graph is constructed based on the edge-center two-level knowledge graph construction method described above.
[0064] Further, step 1: parsing input keywords from natural language describing business requirements and matching them with a pre-constructed edge-center two-level knowledge graph to obtain algorithm model entities includes: Natural language is parsed using named entity recognition and semantic role labeling techniques to obtain input keywords; Based on the input keywords, the system retrieves the corresponding business scenario entities in the pre-constructed edge-center two-level knowledge graph, and then matches them with the entities in the algorithm model.
[0065] Furthermore, the input keywords include: business scenario, processing object, and business task.
[0066] If the business scenario is: drone inspection, the object being processed is insulators, and the business task is defect detection, then the final parsed result will be ["substation", "drone inspection", "insulators", "defect detection"].
[0067] The knowledge graph information relied upon for the guided completion in step 2 is limited to the schema layer of the knowledge graph, namely metadata information such as entity type definitions, attribute field constraints, and required / optional tags, rather than the specific entity instance data in the instance layer of the knowledge graph.
[0068] Further, step 4: the step of retrieving the edge-center secondary knowledge graph based on demand profiles to obtain the candidate platform set for the algorithm model includes: Based on the model classification and business hardware constraints in the demand profile, the model and hardware relationship within the edge-center secondary knowledge graph is retrieved to recall a preliminary candidate platform set. The candidate platform set is obtained by removing platforms that clearly violate latency and concurrency hardware constraints from the initial candidate platform set.
[0069] The retrieval and matching in step 4 relies on specific entity instances and their relationships in the knowledge graph data layer.
[0070] Further, step 5: evaluating the adaptability of algorithm models in the candidate platform set based on entity relationships in the knowledge graph includes: Based on the relationships and attributes between entities in the knowledge graph, model adaptation score, concurrency adaptation score, latency adaptation score and throughput adaptation score are respectively applied to each combination in the candidate platform set. Based on the model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score, the comprehensive matching accuracy of each combination is calculated by weighted summation. The scoring rules for model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score are determined based on the hardware physical upper limit attribute, communication link measured performance attribute, and model running baseline attribute stored in the knowledge graph. This is a quantitative mapping of hardware physical constraints and communication physical constraints.
[0071] Example 3 This invention presents a guided computing power demand standard evaluation method based on knowledge graph technology. This method is designed for the "three-network coordination" construction scenario of new power systems, targeting AI services such as intelligent power grid inspection, NLP analysis, and multimodal recognition. It integrates four core technologies: computing power performance baseline modeling, cross-domain knowledge graph construction, interactive slot filling, and segmented weighted interpretable scoring. This achieves fully automated evaluation from business demand input, automatic constraint completion, intelligent computing power matching to configuration list output. Figure 2 As shown.
[0072] The overall architecture is divided into five layers: data foundation layer, knowledge reasoning layer, guided interaction layer, matching calculation layer, and result output layer. The execution flow is as follows: construction of computing power baseline library → construction of three-layer knowledge graph → business requirement-guided completion → multi-dimensional quantitative scoring and ranking → generation and export of evaluation results. This method does not require business personnel to have professional computing power knowledge. It relies on knowledge graphs to automatically convert business language into computing power indicators, and combines actual baseline data to complete accurate computing power selection. At the same time, it ensures that the evaluation process is interpretable, the results are reproducible, and the data is auditable, solving the problems of low efficiency, large bias, and incomplete indicators in the traditional manual evaluation mode.
[0073] 1. Construct a multi-dimensional computing power performance baseline library (preparatory step) This step provides measured data support for the entire evaluation method, unlike traditional static forms that only record average performance. First, the test objects are selected, choosing three mainstream AI models in the power industry: CNN (image recognition, defect detection), Transformer (text analysis, intelligent customer service), and MoE hybrid expert models (large model inference). Simultaneously, the entire range of existing computing hardware platforms in the power grid is included, encompassing different card counts and inference framework configurations. Second, standardized stress testing experiments are conducted. Under a unified testing environment, different batch sizes, computational accuracies (FP32 / FP16 / INT8), and concurrent request counts are tested, continuously collecting core metrics: maximum concurrency, actual throughput, TTFT (first-word latency P95), TPOT (total latency P95), end-to-end total latency P95, memory usage, and hardware utilization. These core metrics are referred to as raw stress test data. Third, data fitting and storage: The original stress test data is cleaned and outliers are removed. A performance baseline function is constructed using polynomial fitting to describe the nonlinear relationship between the model, hardware, input parameters and various performance indicators. All indicators, baseline functions and hardware parameters are stored in a unified computing power performance baseline library, which supports real-time query and retrieval.
[0074] This step relies on a standardized, full-scale, discrete dataset of stress tests and employs a multivariate polynomial fitting method to construct a performance baseline function that can be solved in real time. Detailed explanations are provided regarding the fitting data source, construction process, function definition, and nonlinear relationship modeling methods, addressing key concerns from experts. The aforementioned computing power performance baseline library was reused, and standardized stress testing and data cleaning were performed. A performance baseline function was constructed through polynomial fitting. A multivariate high-order polynomial was used as the basic model for fitting to obtain the baseline function corresponding to each performance index.
[0075] The input samples for polynomial fitting in this invention are all derived from the raw data of standardized stress tests conducted under the unified environment of this scheme, without the use of external general benchmark datasets. The stress test covers three types of power AI models: CNN, Transformer, and MoE, and a full range of computing hardware, covering multiple batch sizes, FP32 / FP16 / INT8 calculation precision, and gradient concurrent request counts. Discrete measured points such as maximum concurrency, service throughput, TTFT first-word P95 latency, TPOT word-by-word P95 latency, end-to-end total P95 latency, memory usage, and hardware utilization are collected for each set of operating conditions. After cleaning the raw data, removing outliers, and standardizing features, standardized data is obtained. All sampled data is used as the training set for polynomial fitting, which differs from existing technologies that only use a small number of static tables for fixed operating conditions, realizing full-condition measured data-driven modeling.
[0076] 2. Construct a three-layer cross-domain knowledge graph consisting of "business scenarios - algorithm models - computing hardware" (preparatory steps) This invention abandons the traditional, publicly available knowledge graph architecture of one-time static construction and centralized, fixed deployment. It specifically addresses the industry shortcomings of traditional graphs, such as fixed parameters, poor adaptability, and inability to meet the compliance constraints of power grid edge deployment. It innovatively designs a dynamic closed-loop correction mechanism for stress test data and a two-level edge-center graph segmentation and collaborative deployment mechanism to achieve dynamic iteration, precise adaptation, and compliant deployment of the knowledge graph. The overall graph model is a directed multigraph G=(E,R,T), where E is the set of entities, R is the set of relations, and T is the set of triples. The specific construction and iterative deployment process is as follows: (1) Multi-dimensional entity extraction: Three types of core entities are extracted from power grid business specifications, AI model documents, and computing hardware manuals: business scenario entities (intelligent inspection, intelligent customer service, new energy prediction, etc., bound to real-time and reliability business attributes), algorithm model entities (CNN, Transformer, MoE and sub-models, bound to parameter quantity, FLOPs, KV cache and other computing power attributes), and computing hardware entities (various models of inference equipment, bound to peak computing power, video memory, bandwidth, concurrency limit and other hardware attributes). Among them, the algorithm model entities adopt a two-level hierarchical structure. The top level is the three major model categories, and the lower level is subdivided into ResNet, YOLO, BERT, GPT, power sparse large model and other sub-entities, fully covering the mainstream power AI models.
[0077] (2) Standardized relationship definition: Configure multiple sets of association relationships, including core semantic relationships such as "business scenario → depends on → algorithm model", "algorithm model → deployed on → computing hardware" and "business scenario → constraint → latency / concurrency / throughput indicators", to connect the association links of the three-layer entities.
[0078] (3) Rule Embedding and Capability Integration: Self-developed customized SHACL shape constraint verification rules and SWRL semantic logic reasoning rules are embedded in the graph. Simultaneously, K-hop neighborhood expansion and TransE vector reasoning capabilities are integrated to achieve data compliance verification, implicit knowledge reasoning, and semantic expansion functions. It should be clearly stated that SHACL and SWRL are general semantic rule frameworks and are well-known basic technologies in this field. However, the custom constraint statements, business reasoning logic, and threshold linkage rules customized for the power computing power assessment scenario in this invention are not common knowledge and belong to the differentiated innovation content of this invention. Specific customized rules are detailed below: ① This invention provides customized SHACL constraint verification rules (specifically for power computing scenarios) Based on the W3C SHACL standard framework, three types of custom rigid constraints are defined for the power AI computing power demand assessment scenario. These constraints are used to automatically verify the completeness of business requirement parameters, the legality of numerical values, and the model-hardware matching correlation, preventing issues such as missing parameters, out-of-bounds errors, and business adaptation conflicts. The specific rules are as follows: 1) Parameter Existence Constraint Rules: Forced verification that the slots for power computing power assessment must not be empty. Rule logic: Business scenario instances must be bound to seven attributes: model classification, model size, maximum concurrency C, P95-TTFT upper limit, P95-TPOT upper limit, end-to-end P95 latency upper limit, and minimum throughput Tmin. If any attribute is empty, the requirement is determined to be incomplete, triggering the guided completion mechanism.
[0079] 2) Numerical range constraint rules: Adapt to the power grid SLA operation specifications, customize the legal threshold range of power computing parameters, and the rule logic is: the number of concurrent users C≥1, the delay parameters are all >0 and do not exceed the maximum tolerance threshold of the power grid business, and the throughput Tmin>0. If the parameters exceed the limit, they will be directly intercepted and prompted for correction to avoid the problem of computing power matching distortion caused by invalid parameter input.
[0080] 3) Business-Model Linkage Constraint Rules: Establish a binding and verification relationship between power business and AI model. The rule logic is as follows: image inspection business is only allowed to match the CNN model system, text interaction business is only allowed to match the Transformer model system, and multimodal integrated business is only allowed to match the MoE model system, so as to prevent incorrect adaptation scenarios where the business and model architecture do not match.
[0081] ② This invention customizes SWRL semantic logic reasoning rules (dedicated computing power matching reasoning). Based on the SWRL rule-based reasoning framework, a custom logic for inferring implicit relationships across domains in power computing power is developed. This logic can automatically derive implicit computing power constraints based on explicit business parameters, overcoming the limitation of traditional graphs that can only store explicit data and cannot intelligently deduce and adapt rules. The specific reasoning rules are as follows: 1) Model Scale - Resource Consumption Inference Rules: If the business is bound to a model with a large number of parameters (≥10B), then the automatic inference will derive constraints: the required video memory bandwidth and the single card video memory usage threshold will be adjusted simultaneously to automatically match high bandwidth and large video memory computing power hardware, and avoid the problems of large model video memory overflow and inference lag.
[0082] 2) High concurrency-latency linkage reasoning rules: If the maximum concurrency of the service C ≥ 32 channels (power high concurrency scenario), then the automatic reasoning will derive constraints: forcibly tighten the TTFT, TPOT, and P95 latency fault tolerance thresholds, select high-stability computing power platforms, and adapt to the high reliability operation requirements of the power grid peak.
[0083] 3) Business Level - Fault Tolerance Threshold Reasoning Rules: If the business is a core production business such as power dispatching or risk control auditing, the following constraints will be automatically derived: 1.2 times the lenient latency threshold will fail, and 1.0 times the strict standard will be implemented to prevent the risk of latency exceeding the standard in core businesses; non-core operation and maintenance businesses will retain 1.2 times the engineering fault tolerance threshold to balance stability and adaptability.
[0084] 4) Operating Accuracy-Throughput Matching Inference Rules: If the service uses FP16 / INT8 low-precision inference, the hardware throughput adaptation range will be automatically inferred, and low-precision high-throughput computing power equipment will be matched first to achieve accurate matching between computing power performance and service operating conditions.
[0085] This invention, through the coordinated linkage of the aforementioned SHACL and SWRL customized rules, differs from the industry's general basic framework, and achieves automated verification of power computing power requirements, intelligent derivation of implicit constraints, and adaptive threshold for specific scenarios, thus completely solving the problems of traditional graph verification and inference being disconnected, data being invalid, and conclusions being distorted.
[0086] (4) Dynamic Closed-Loop Correction of Load Testing Data: Unlike the traditional fixed mode where knowledge graphs remain static after construction, this invention establishes a two-way dynamic closed-loop iteration mechanism between baseline load testing data and the knowledge graph. Relying on the continuously updated full-condition load testing data from the computing power performance baseline library, the model-hardware adaptation relationship, latency / concurrency / throughput constraint thresholds, and resource association attributes stored in the graph are compared in real time to automatically identify deviations between theoretical graph rules and actual load testing conditions. For deviation entities and triple relationships, the semantic association weights, constraint threshold ranges, and business adaptation mapping relationships are automatically iteratively corrected, and the model-level adaptation rules and hardware performance matching boundaries are updated synchronously. Through continuous feedback of the test data to the graph iteration, the problems of rigid rules, lagging adaptation, and low matching accuracy of traditional static graphs are completely solved, and the accuracy and scenario adaptability of knowledge graph computing power matching reasoning are greatly improved. The semantic association weights are the weight values on the edges of the knowledge graph triple relationships, which are the quantified weights of the relationships between entities; the constraint threshold ranges are numerical parameters embedded within the SHACL / SWRL rules (e.g., SHACL numerical range constraints: concurrency C≥1, P95 latency ≤X; here 1 and the latency upper limit X are the constraint threshold ranges); the business adaptation mapping relationships are association relationships (e.g., entity-entity mapping: business scenario to adaptation algorithm model, algorithm model to adaptation hardware; entity-parameter index mapping: business scenario entities are mapped to the corresponding latency, concurrency, and throughput SLA index sets); updating model-level adaptation rules is a synchronous update of model-level adaptation rules, for example, it cannot be simply and completely equivalent to SHACL / SWRL, but part of it will be implemented as SHACL / SWRL, and the other part is the hierarchical constraint metadata of the graph ontology layer.
[0087] (5) Edge-Center Two-Level Collaborative Deployment: Addressing the stringent compliance constraints of power grid deployment, this invention solves the problem of adapting traditional single centralized graph architecture to substation edge nodes. Publicly available research solutions generally adopt a centralized full-scale graph deployment mode, which has two major pain points in power grid adaptation: First, the computing power and storage resources of substation edge nodes are limited, prohibiting the deployment of massive full-scale model graph data; second, the external network connectivity of power grid edge stations is limited, and remote calls to centralized graphs pose network latency and compliance access risks. This invention designs a two-level collaborative deployment architecture of central full-domain graph + edge sharded graph: The central node deploys a complete three-layer cross-domain knowledge graph, responsible for global rule updates, data closed-loop correction, and overall matching of computing power resources; each substation edge node deploys a scenario-specific sharded graph as needed, storing only simplified entities and association rules of local business, adaptation models, and edge computing hardware, without deploying full redundant data. Simultaneously, a data synchronization and rule linkage mechanism between the edge and center is established. The map rules updated iteratively by the center are automatically pushed down and synchronized to the edge shards. The measured data from the edge is transmitted back to the center to complete the closed-loop correction. This not only meets the deployment requirements of lightweight, compliant, and offline usable power grid edge nodes, but also ensures the consistency of map data and the accuracy of inference across the entire domain.
[0088] (6) Graph entry and implementation: Complete the binding and entry of entities, relationships, attributes, dynamic rules, and two-level sharding deployment strategies into the database, forming a standardized three-layer cross-domain dynamic knowledge graph database that can be iterated, corrected, and deployed in layers.
[0089] 3. Receive initial business requirements input Business personnel submit initial requirements via a web interface or form, using natural language keywords / short phrases such as "substation drone inspection" or "intelligent customer service dialogue system," without needing to fill in professional computing power parameters. The system receives the input text and standardizes the data format.
[0090] 4. Intent Recognition and Graph Entity Mapping Business personnel input natural language business descriptions into the web interface, and the backend calls Named Entity Recognition (NER) and Semantic Role Labeling (SRP) technologies to parse the input keywords; it then searches the three-layer knowledge graph to match the natural language descriptions to the corresponding business scenario entities; based on the graph relationship paths, it initially associates the default matching algorithm model categories for the business. The first step is entity extraction, which extracts three types of entities from power grid business specifications, model documents, and hardware manuals: business scenario entities (intelligent inspection, intelligent customer service, etc., bound to real-time and reliability attributes), algorithm model entities, and computing power hardware entities (various models of inference devices, bound to peak computing power, video memory, bandwidth, and other attributes).
[0091] Input keywords include: business scenario (e.g., drone inspection), processing object (e.g., insulator), business task (e.g., defect detection). The final parsed result is ["substation", "drone inspection", "insulator", "defect detection"]. Then, based on the keywords "drone inspection" and "defect detection", the knowledge graph business scenario entity "[intelligent inspection of power grid equipment]" is matched. Based on the business scenario entity relationships, the top-level algorithm model category is automatically associated: CNN model category (YOLO / ResNet).
[0092] The algorithm model entities are structured in a two-level hierarchy. The top level comprises major model categories, including CNN, Transformer, and MoE. Each major category has subcategories; for example, the CNN category includes image detection subcategories like ResNet and YOLO, the Transformer category includes power dialogue subcategories like BERT and GPT, and the MoE category includes sparse power multimodal models. Regardless of the major or subcategories, all entities are uniformly bound to inherent computational power attributes such as parameter count, theoretical FLOPs, KV cache usage, computation-to-memory ratio, and inference latency baseline. The system first matches the top-level model category corresponding to the business using graph relationship paths, and then confirms the subcategories and model size parameters through an interactive guided completion process, thus completing intent localization.
[0093] 5. Missing slot detection and interactive guided completion An interactive slot filling mechanism is adopted, with predefined mandatory slots for model size, maximum concurrency, TTFT upper limit, TPOT upper limit, end-to-end latency upper limit, and minimum throughput requirements. Missing slots are detected based on SHACL rules from the knowledge graph, and simple guidance questions are automatically generated and pushed to the interface. User responses are received and their validity is verified; if parameters are abnormal, corrections are prompted until all slots are filled. The predefined mandatory slots for model size, maximum concurrency, TTFT upper limit, TPOT upper limit, end-to-end latency upper limit, and minimum throughput requirements are all declared uniformly in the three-layer cross-domain knowledge graph ontology layer. Semantic constraint rules are written in the SHACL language and embedded in the knowledge graph storage. The rules include three core verification logics: mandatory attribute existence constraints, parameter value range constraints, and business-model linkage dependency constraints. Missing slot detection relies on the built-in SHACL inference engine of the knowledge graph to achieve complete semantic verification. The execution flow is as follows: The business natural language is identified and mapped by NER to a temporary instance of the graph business scenario, which only carries basic business entity information and does not carry complete constraint slot parameters; The knowledge reasoning layer calls the SHACL reasoning engine to match and verify the temporary requirement instances with the SHACL shape rules of the preloaded graph, and checks all the required slots defined in the ontology one by one through the existence constraint sh:minCount=1. If a temporary instance is missing a slot attribute, the attribute is empty, or the parameter value is outside the valid range, then the slot is determined to be missing / the parameter is abnormal, and a list of missing slots is output. The system generates simplified interactive questions based on the missing list and pushes them to the front-end guidance interaction layer. After receiving the user's supplementary parameters, it updates the temporary requirement instance of the graph and performs SHACL validation repeatedly until all required slots are complete, parameters are compliant, and the requirement completeness reaches 100%.
[0094] 6. Generate standardized business requirement profiles By integrating the initial keywords with the completed constraint parameters, a machine-recognizable structured business requirement vector is generated. (r=(mcat,msize,C,TTFT 95max,TPOT95max,LAT95max,Tmin)) where: mcat is the model classification, msize is the model size, C is the number of concurrent connections, and the last four terms are the P95 latency constraint and the minimum throughput requirement, which are used as unified inputs for subsequent matching calculations. The structured business requirement vector is generated by the knowledge reasoning layer. The guidance and interaction layer only completes the initial collection of business keywords and interactive slot parameters. After all slots are verified as complete and valid through SHACL rules, the original business data is forwarded to the knowledge reasoning layer. The initial value of the model classification mcat comes from the business default matching algorithm model category obtained in step 4 based on the graph relationship path retrieval, namely the three top-level models: CNN, Transformer, and MoE. If there are customized models or multi-model deployment scenarios in the business, business personnel can manually modify the model category in the interactive guidance completion stage, and mcat will be updated based on the user's confirmation result. The knowledge reasoning layer extracts the model classification, model size, concurrency, TTFT upper limit, and TPOT from the graph matching results and the completed completion parameters. The raw information of the upper limit, end-to-end latency limit, and minimum throughput is used to perform digital encoding mapping on the model classification mcat. At the same time, the model size msize, concurrency C, various P95 latency constraints, and minimum throughput Tmin are uniformly converted into numerical values under the corresponding units. Natural language text is removed, and then the data is assembled into a fixed-length one-dimensional structured demand vector according to a fixed field order. This vector is then output to the matching calculation layer as a unified input for computing power screening and weighted scoring.
[0095] 7. Knowledge Graph Selection of Candidate Computing Power Platforms Based on model classification and business hard constraints in the demand profile, the "model-hardware" relationship is retrieved within the knowledge graph to recall a preliminary set of candidate platforms. Platforms that clearly violate latency and concurrency hard constraints are eliminated to narrow the computational scope and reduce computational overhead. Here, business hard constraints are derived from standardized business demand vectors and are mandatory requirements of the business SLA. They include model classification constraints, concurrency hard constraints, TTFT95max upper limit, TPOT95max upper limit, end-to-end P95 latency upper limit, and minimum throughput requirements. If a candidate platform fails to meet any of these hard constraints, it is directly filtered out and does not participate in subsequent scoring calculations. Among them, the concurrency hard constraint refers to the maximum concurrency C in the business demand profile, that is, the upper limit of the number of simultaneous requests allowed by the business. The maximum concurrency that a candidate platform can support, calculated by the performance baseline function, must be greater than or equal to the maximum business concurrency C. For example, in the intelligent customer service business, the maximum concurrent requests are 50. If a certain computing platform can only support 30 concurrent requests under the corresponding model conditions, it will be directly eliminated because it does not meet the hard concurrency constraints. The complete set of hard constraints for this business is as follows: support for Transformer14B model, maximum concurrency of no less than 50, P95 latency of TTFT no greater than 800ms, P95 latency of TPOT no greater than 100ms, total end-to-end P95 latency no greater than 1500ms, and minimum throughput of no less than 120 requests per second.
[0096] 8. Multi-dimensional weighted matching score calculation It adopts a weighted additive scoring function and a segmented penalty mechanism, with a total score range of [0, 100], divided into four dimensions.
[0097] (1) Model fit score (30 points in total, or 10 points in total if there is no model size): This invention is based on the two-level hierarchical system of the knowledge graph (model category - sub-model / size), establishes standardized matching judgment rules, clarifies the definition conditions of three matching scenarios, and the entire process is automatically completed by the system for graph reasoning matching without human subjective judgment. The specific judgment criteria are as follows: ① Only the classification is consistent (10 points): The business needs and the computing power platform only match the top-level model category (both are CNN / Transformer / MoE), but the sub-model series, model parameter quantity, and model scale are completely mismatched. It only meets the basic architecture adaptation and cannot meet the business's refined inference conditions. ② Partial match (15 points): The business requirements are consistent with the major category and sub-series of the computing power platform model, but the number of model parameters and scale are not completely equivalent. It belongs to the approximate adaptation of the same architecture and series, which can be compatible with business inference but is not the optimal working condition. ③ Perfect Scale Match (30 points): Business requirements are fully aligned with the main categories, sub-series, and parameter levels of the computing platform model, and are perfectly matched with the standard working conditions of the baseline library stress test, making it the optimal model specification for business needs.
[0098] The overall model adaptation score is driven entirely by the system: First, it automatically identifies missing business parameters through the knowledge graph and guides the completion of these parameters to obtain accurate model type and scale parameters. Then, it automatically completes the model adaptation verification and scoring based on the entity association relationships in the knowledge graph, quantifying the accuracy of the computing power selection and matching. A complete scale match scores 30 points, a partial match scores 15 points; only the classification is consistent and the scale does not match scores 10 points; no adapted model specifications score 0 points.
[0099] (2) Concurrency Adaptation Score (25 points): This invention establishes standardized concurrency matching judgment rules based on measured data from the computing power performance baseline library. The entire process is automatically calculated and judged by the system without any subjective human intervention, accurately defining and improving the standard and insufficient scenarios. The specific judgment criteria are as follows: ① Platform concurrency meets requirements (25 points): The system calls the corresponding baseline function to solve for the maximum number of concurrent requests that the computing power platform can stably handle based on the model specifications, calculation accuracy, batch size and other working parameters of the business requirements. When the maximum number of concurrent requests that the platform can support is greater than or equal to the maximum concurrency constraint C set by the business requirements, it is considered that the concurrency capability meets the standard and can stably handle the peak concurrent requests of the business without the risk of concurrent blocking or queuing timeout, and gets a full score of 25 points.
[0100] ② Insufficient Platform Concurrency (Proportional Conversion + Penalty Coefficient): If the maximum concurrent connection that the computing power platform's baseline solution can support is less than the maximum concurrent constraint C of the business requirements, it is considered to have insufficient concurrency capacity and cannot fully handle the peak access pressure of the business, which is prone to problems such as request queuing, business response timeouts, and task backlog. In this case, the base score is converted according to the ratio of "platform's actual maximum concurrency / business requirement concurrency". The final score = converted base score × 0.5 penalty coefficient. The penalty mechanism weakens the recommendation priority of platforms that do not meet the concurrency standards, ensuring that the computing power selection meets the peak operation requirements of the business.
[0101] (3) Latency Adaptation Score (25 points in total): This invention addresses three core power business latency indicators: TTFT first-word P95 latency, TPOT word-by-word P95 latency, and end-to-end total P95 latency. It introduces a 1.2 times lenient threshold mechanism to adapt to engineering implementation scenarios and establishes standardized three-level latency judgment rules. The entire process is automatically judged by the system based on baseline measured data without human subjective intervention, accurately distinguishing between three operating conditions: meeting the standard, slightly exceeding the standard, and seriously exceeding the standard. The specific judgment criteria are as follows: ① Latency meets the standard (full marks 25 points): The actual measured latency of each item on the computing power platform is ≤ the corresponding latency upper limit threshold set by the business requirements, fully meets the business SLA latency constraints, and has no risk of timeout. Full marks can be obtained for this dimension if all three latency indicators meet the standard.
[0102] ② Slight Exceedance (Half a point): Any latency measured by the computing power platform is greater than the business standard threshold, but less than or equal to the business standard threshold × 1.2 times the lenient threshold. This is a working condition that slightly exceeds the theoretical standard but is still compatible with stable business operation. It will not cause obvious business lag or timeout failure. After proportional conversion, it can receive a maximum of 50% of the score for this dimension.
[0103] ③ Severely Exceeds Standards (0 points): Any latency measured by the computing power platform is greater than the business standard threshold × 1.2 times the lenient threshold, which exceeds the engineering's compatibility and fault tolerance range. This will lead to operational risks such as business response lag, interaction timeout, and task failure. The latency adaptation is directly judged as not meeting the standards, and this dimension will receive 0 points.
[0104] (4) Throughput Adaptation Score (20 points in total): Based on the measured throughput data of the computing power performance baseline library, this invention establishes a standardized throughput matching, proportional conversion, and redundancy prevention scoring mechanism. The entire process is automatically calculated by the system without human intervention, accurately distinguishing between scenarios where throughput meets the standard and scenarios where throughput is insufficient, while suppressing the waste of excessive computing power configuration. The specific judgment and scoring rules are as follows: ① Throughput Meets Requirement (Proportional Scoring, Suppressing Over-Configuration): Based on business models, concurrency, accuracy, and other operating conditions, the system calls a baseline function to calculate the measured steady-state maximum throughput of the computing platform. When the platform's measured throughput is greater than or equal to the minimum throughput requirement Tmin for the business, the throughput capability is considered to meet the standard. To avoid excessive redundancy and waste of computing resources, this invention does not adopt a "meeting the standard equals full marks" mechanism. Instead, it suppresses over-configuration through proportional scoring based on demand. The scoring formula is: Score = 20 × (Business Required Throughput ÷ Platform Measured Throughput). The closer the platform throughput is to the business's essential needs and the smaller the redundancy, the closer the score is to the full 20 points, completely solving the problem of excessive over-configuration and idle resources in traditional computing power selection.
[0105] ② Insufficient Throughput (Proportional Conversion + Penalty Factor): If the platform's actual measured throughput is less than the minimum throughput requirement Tmin for the business, it is considered to have insufficient throughput capacity and cannot handle the steady-state inference load of the business, which is prone to problems such as task backlog, insufficient processing throughput, and degraded business throughput. First, the score is calculated based on the basic proportional conversion: Basic Score = 20 × (Platform Actual Measured Throughput ÷ Business Required Throughput), and then a penalty mechanism is added to weaken the priority of the substandard platform. The final score = Basic Score × 0.5.
[0106] The overall four-dimensional scores are all automatically calculated based on the platform's measured baseline data, without any subjective human correction. Finally, the scores of model adaptation, concurrency adaptation, latency adaptation, and throughput adaptation are accumulated to obtain the final comprehensive matching score of each candidate computing power platform, realizing a fully quantitative, reproducible, and auditable computing power matching evaluation.
[0107] This invention designs a scenario-adaptive weighted scoring function with a segmented penalty mechanism, breaking through the commonly used manual, globally fixed static weight mode in existing technologies. Addressing the differentiated priority requirements of various power business scenarios, it innovatively introduces a knowledge graph-based business entity-specific weight vector adaptive optimization mechanism, enabling intelligent dynamic adaptation of scoring weights to business scenarios, with the total score range fixed at [0, 100]. The overall scoring dimensions include four core dimensions: model adaptation, concurrency adaptation, latency adaptation, and throughput adaptation. The baseline full score weights for each dimension are 30, 25, 25, and 20 points, respectively. Specifically, a dedicated weight vector is bound to the knowledge graph entity attributes, replacing the globally fixed weights, accurately aligning with the core performance requirements of different power businesses. The specific mechanism and scoring rules for each dimension are as follows: The core mechanism of adaptive weight optimization: This invention pre-binds scenario-specific weight vectors for various power businesses into the business scenario entity layer of the three-layer cross-domain knowledge graph. It automatically optimizes the scoring weights across four dimensions based on the project priority and SLA constraint focus of the business scenario, eliminating the use of uniform fixed weights. After the system completes the mapping of business scenario entities, it automatically reads the pre-set weight configuration of that business entity and dynamically allocates the scoring proportions for model adaptation, concurrency, latency, and throughput, achieving scenario-based adaptive adaptation of the scoring engine. For example, for core production businesses such as power dispatching and risk control auditing, the weights of latency and concurrency are automatically increased to prioritize business operational reliability; for offline inference businesses such as drone inspections and batch image analysis, the weights of model adaptation and throughput are automatically increased to prioritize inference efficiency and model adaptation accuracy; for interactive businesses such as intelligent customer service, the fine-grained latency weights of TTFT and TPOT are emphasized to meet the essential needs of interactive experience. The weight vectors are fixed in the business entity nodes of the knowledge graph and continuously optimized with the dynamic iteration of the graph, possessing the characteristics of being interpretable, traceable, and adaptive.
[0108] (1) Model Adaptation Dimensions (Benchmark maximum score of 30 points, or 10 points for models without scale): This invention establishes standardized matching judgment rules based on a two-level hierarchical system (model category - sub-model / scale) pre-defined by the knowledge graph. The entire process of graph reasoning and matching is automatically completed by the system without any subjective human judgment. The scores for each dimension are the upper limit of the scenario-adaptive benchmark. The actual score ratio is dynamically optimized by the exclusive weight vector of the knowledge graph business entity to adapt to the different priorities of different power businesses. The specific matching judgment criteria are as follows: ① Only the classification is consistent (score of 10 points under the corresponding scenario weight): The business needs and the computing power platform only match the top-level model category (both are CNN / Transformer / MoE), but the sub-model series, model parameter quantity, and model scale are completely mismatched. It only meets the basic architecture adaptation and cannot fit the business's refined inference conditions. ② Partial matching (gets a score of 15 points under the corresponding scenario weight): The business requirements are consistent with the main category and sub-series of the computing power platform model, but the number of model parameters and scale are not completely equivalent. It belongs to the approximate adaptation of the same architecture and series, which can be compatible with business inference but is not the optimal working condition. ③ Scale is perfectly matched (score of 30 points under the corresponding scenario weight): The business requirements are fully aligned with the major categories, sub-series of models, and model parameter quantity / scale levels of the computing power platform, and are fully matched with the standard working conditions of the baseline library stress test, which is the optimal model specification for the business.
[0109] The overall model adaptation score is driven by the system in a fully automated manner: first, the system automatically identifies missing business parameters through the knowledge graph and guides the completion of these parameters to obtain accurate model type and scale parameters; then, based on the entity association relationship in the knowledge graph, the system automatically completes the model adaptation verification and scoring of the computing power platform, quantifying the accuracy of computing power selection and matching.
[0110] (2) Concurrency Adaptation Dimension (Benchmark Maximum Score: 25 points): This invention establishes standardized concurrency matching judgment rules based on actual test data from the computing power performance baseline library. The entire process is automatically calculated and judged by the system without any subjective human intervention. The score for this dimension is the upper limit of the scenario adaptive benchmark. The actual score weight is dynamically adapted by the knowledge graph business entity vector, accurately defining and developing the target and insufficient scenarios. The specific judgment criteria are as follows: ① Platform concurrency meets requirements (full score of 25 points under the corresponding scenario weight): The system calls the corresponding baseline function to solve for the maximum number of concurrent requests that the computing power platform can stably bear based on the model specifications, calculation accuracy, batch size and other working parameters of the business requirements. When the maximum number of concurrent requests that the platform can support is ≥ the maximum concurrency constraint C set by the business requirements, it is considered that the concurrency capability meets the standard and can stably bear the peak concurrent requests of the business without the risk of concurrent blocking or queuing timeout.
[0111] ② Insufficient Platform Concurrency (Proportional Conversion + Penalty Coefficient): If the maximum concurrent connection that the computing power platform's baseline solution can support is less than the maximum concurrent constraint C of the business requirements, it is considered to have insufficient concurrency capacity and cannot fully handle the peak access pressure of the business, which is prone to problems such as request queuing, business response timeouts, and task backlog. In this case, the base score is converted according to the ratio of "platform's actual maximum concurrency / business requirement concurrency". The final score = converted base score × 0.5 penalty coefficient. The penalty mechanism weakens the recommendation priority of platforms that do not meet the concurrency standards, ensuring that the computing power selection meets the peak operation requirements of the business.
[0112] (3) Latency Adaptation Dimension (Benchmark Maximum Score 25): This invention addresses three core power business latency indicators: TTFT first-letter P95 latency, TPOT word-by-word P95 latency, and end-to-end total P95 latency. It introduces a 1.2x lenient threshold mechanism to adapt to engineering implementation scenarios and establishes standardized three-level latency judgment rules. The score for this dimension is the upper limit of the scenario-adaptive benchmark. The actual weight allocation is dynamically optimized based on knowledge graph business entities. The entire process is automatically judged based on baseline measured data without subjective human intervention, accurately distinguishing between three operating conditions: compliant, slightly exceeding the standard, and severely exceeding the standard. The specific judgment criteria are as follows: ① Latency meets the standard (get the full score of 25 points under the corresponding scenario weight): The actual latency of each item measured by the computing power platform is ≤ the corresponding latency upper limit threshold set by the business requirements, fully meets the business SLA latency constraints, and there is no risk of timeout. If all three latency indicators meet the standard, the full score for this dimension can be obtained.
[0113] ② Slight Exceedance (half a point): Any latency measured by the computing power platform is greater than the business standard threshold, but less than or equal to the business standard threshold × 1.2 times the lenient threshold. This is a case of slightly exceeding the theoretical standard but still compatible with stable business operation. It will not cause obvious business lag or timeout failure. After proportional conversion, it can receive a maximum of 50% of the score for this dimension.
[0114] ③ Severely Exceeds Standards (0 points): Any latency measured by the computing power platform is greater than the business standard threshold × 1.2 times the lenient threshold, which exceeds the engineering's compatibility and fault tolerance range. This will lead to operational risks such as business response lag, interaction timeout, and task failure. The latency adaptation is directly judged as not meeting the standards, and this dimension will receive 0 points.
[0115] (4) Throughput Adaptation Dimension (Benchmark Maximum Score: 20 points): This invention establishes a standardized throughput matching, proportional conversion, and anti-redundancy scoring mechanism based on measured throughput data from the computing power performance baseline library. The score for this dimension is the upper limit of the scenario adaptive benchmark. The actual weight is dynamically allocated by the entities in the knowledge graph business scenario, realizing differentiated priority adaptation for different businesses. The entire process is automatically calculated by the system without manual intervention, accurately distinguishing between scenarios with sufficient throughput and those with insufficient throughput, while suppressing the waste of excessive computing power configuration. The specific judgment and scoring rules are as follows: ① Throughput Meets Requirement (Proportional Scoring, Suppressing Over-Configuration): Based on business models, concurrency, accuracy, and other operating conditions, the system calls a baseline function to calculate the measured steady-state maximum throughput of the computing platform. When the platform's measured throughput is ≥ the minimum throughput requirement Tmin for the business, it is considered to meet the throughput requirement. To avoid excessive redundancy and waste of computing resources, this invention does not adopt a "meeting the requirement equals full marks" mechanism. Instead, it suppresses over-configuration through proportional scoring based on demand. The scoring formula is: Score = Full marks for the corresponding dimension × (Business demand throughput ÷ Platform measured throughput). The closer the platform throughput is to the business's essential needs and the smaller the redundancy, the closer the score is to the upper limit of the dimension's full marks, completely solving the problems of excessive over-configuration and idle resources in traditional computing power selection.
[0116] ② Insufficient Throughput (Proportional Conversion + Penalty Factor): If the platform's actual measured throughput is less than the minimum throughput requirement Tmin for the business, it is considered to have insufficient throughput capacity and cannot handle the steady-state inference load of the business, which is prone to problems such as task backlog, insufficient processing throughput, and degraded business throughput. First, the score is converted according to the basic proportion: Basic score = full score of the corresponding dimension × (platform's actual measured throughput ÷ business required throughput), and then a penalty mechanism is added to weaken the priority of the substandard platform. The final score = basic score × 0.5.
[0117] The overall four-dimensional scoring is automatically calculated based on the platform's measured baseline data. The proportion of each dimension is dynamically adjusted by combining the adaptive weight vector of the knowledge graph. There is no manual subjective setting of fixed weights. Finally, the adaptive scores of each dimension are accumulated to obtain the final comprehensive matching score of the candidate computing power platform, realizing a fully quantitative, reproducible, and auditable computing power matching evaluation.
[0118] Finally, based on the scene-specific weight vectors that are adaptively optimized from the knowledge graph, the scores of the four dimensions are accumulated through a weighted summation formula to obtain the comprehensive matching score of each candidate computing power platform. The entire process combines dynamic scoring based on scene priority with a segmented penalty mechanism, which completely solves the technical shortcomings of traditional fixed weights that cannot adapt to the differentiated needs of multiple power scenarios.
[0119] 9. Top-K ranking of candidate platforms All candidate platforms are sorted from highest to lowest according to their matching scores, and the Top-K (default Top-5) are selected as the recommendation results; the detailed scores of each dimension and the matching status of the indicators are recorded simultaneously to provide a basis for interpretable output.
[0120] This invention does not select a single highest-scoring platform, but instead outputs a fixed Top-5 candidate set. The core technical reasons are threefold: First, power computing power deployment faces diverse engineering constraints. A single highest-scoring platform only represents the theoretical optimal match, and there may be implicit implementation limitations such as insufficient on-site hardware inventory, limited data center space, cluster interconnection compatibility, and project procurement cost control. Outputting only the optimal value can easily lead to poor solution implementation tolerance and a lack of backup options. Second, the Top-5 multiple candidate results provide a tiered and comparable configuration scheme spectrum, supporting technical personnel to make secondary selections based on business priorities, cost budgets, and on-site conditions, balancing theoretical optimal matching with engineering feasibility. Third, it meets the standardized management requirements of power project audit comparison, scheme selection, and backup options. Multiple candidate results can fully demonstrate the adaptation differences, scoring gradients, and advantages / disadvantages of different computing power platforms, offering greater interpretability and selectivity compared to a single optimal value, and adapting to business scenarios of intensive deployment of power grid computing power and multi-solution comparison and review.
[0121] 10. Generate an interpretable evaluation report and a standardized computing power configuration list. The output includes business information, constraint parameters, candidate platform rankings, score breakdown, indicator comparison, risk warnings, and a list of optimal hardware configurations; it supports web visualization and export in multiple formats such as Excel and JSON, meeting the requirements for project application, auditing, and archiving.
[0122] 11. Data Archiving and Experiment Reproduction The original requirements, constraints, scoring data, and results of this assessment will be archived uniformly. With the baseline data, graph rules, and algorithm logic unchanged, the assessment results can be reliably reproduced using the same inputs, meeting the requirements for power project auditing and review.
[0123] 12. Signal and data flow between modules The user input data collected by the interaction layer is transmitted to the knowledge reasoning layer in the form of text signals. The knowledge reasoning layer retrieves the knowledge graph and baseline library data from the data base layer, completes semantic parsing and slot completion, and then transmits the standardized requirement vector to the matching calculation layer. The matching calculation layer reads the platform performance indicators, performs scoring and sorting, and transmits the results and score details to the interface and output layer. The interface and output layer displays the data externally, exports files, and writes all the data back to the data base layer for archiving.
[0124] A new inference framework entity subset and compatibility association mapping mechanism has been added: Based on the existing three-layer knowledge graph entity set of "business scenario - algorithm model - computing hardware," a new independent inference framework entity subset has been added, fully covering the mainstream inference engine types deployed in the power grid, including commonly used power inference frameworks such as MindIE, TensorRT, ONNX Runtime, and OpenVINO. For the same "algorithm model - computing hardware" combination, the traditional fixed framework binding mode has been abandoned. Independent "model-hardware-inference framework" performance mapping edges have been constructed for each combination, recording the measured performance parameters, adaptation characteristics, and compatibility constraints of different inference frameworks under corresponding hardware and model operating conditions, forming a multi-dimensional framework adaptation graph system. Based on this mechanism, the system does not require users to input or fix inference framework parameters in advance. During the scoring and evaluation stage, it can automatically traverse each framework adaptation branch, intelligently selecting the optimal inference framework combination based on business latency, accuracy, and concurrency constraints, completely solving the problem of computing power configuration recommendation deviation caused by performance differences between different inference frameworks.
[0125] A new scenario label for tolerance of accuracy loss in inference frameworks has been added: A dedicated accuracy loss tolerance label for power business has been added to the entity attributes of inference frameworks, with differentiated labeling based on power grid business operation specifications and scenario characteristics. Core production businesses such as power grid dispatching and risk control auditing are labeled "zero accuracy tolerance," strictly selecting high-fidelity, zero-accuracy-loss native inference frameworks; visual businesses such as intelligent equipment inspection and defect recognition are labeled "low accuracy tolerance," compatible with INT8 / FP16 quantized inference frameworks, balancing inference speed and recognition accuracy; interactive businesses such as intelligent customer service and text analysis are labeled "mild accuracy tolerance," compatible with lightweight optimized inference frameworks. During the computing power matching and evaluation process, the system can automatically match the corresponding accuracy constraint threshold according to the business scenario type, selecting suitable inference frameworks to achieve precise three-in-one matching of computing power hardware, AI models, and inference frameworks.
[0126] Enhanced end-to-end data flow logic: After completing business requirement completion and generating structured profiles, the knowledge reasoning layer synchronously calls the compatibility relationship and precision tag attributes of the entity subset of the reasoning framework. Combined with the business SLA precision and latency constraints, it automatically filters incompatible and substandard framework branches. The matching calculation layer adds implicit constraints on the adaptability of the reasoning framework to the original four-dimensional scoring to complete multi-dimensional comprehensive scoring and ranking. Finally, the optimal computing power configuration list is output and synchronously matched with the optimal reasoning engine solution to achieve full-process, no manual intervention, and intelligent adaptation of all parameters.
[0127] Taking an intelligent customer service dialogue system (NLP-based interactive business) as an example: The baseline library and knowledge graph deployment have been completed in the pre-processing stage; business personnel input the keyword "intelligent customer service dialogue system"; the system automatically detects missing parameters and guides users to supplement them with a maximum concurrency of 50 paths, a total P95 latency of ≤15000ms, and a TPOT of ≤100ms; after generating a standardized requirement profile, the knowledge graph filters all NLP-based inference platforms; after weighted scoring calculation, MindIE1.0-14B (2 cards) scored the highest; finally, a Top 5 recommendation list and configuration list are output. This evaluation took approximately 1 minute, with a requirement completeness rate of 100% and computing power redundancy controlled within 30%, meeting the business operation requirements.
[0128] This invention systematically optimizes several inherent defects in existing computing power assessment technologies. Verified through engineering experiments, it achieves several breakthrough technical effects while ensuring operational reliability and assessment accuracy, thus meeting the standardized, intensive, and intelligent computing power scheduling needs of the power industry. 1. Improve the computing power evaluation index system and significantly enhance the accuracy of configuration matching: Abandon the traditional single evaluation mode that only uses average latency, and add core power business indicators such as P95 tail latency, TTFT first word latency, and TPOT word-by-word latency to build a full-dimensional computing power performance baseline. This accurately adapts to the high reliability of the power grid and the business characteristics of strictly controlling long-tail latency, effectively avoids the risk of peak load operation, and greatly improves the fit between computing power configuration and actual business SLA requirements.
[0129] 2. Completely solve the problem of missing parameters and achieve a significant improvement in evaluation efficiency: Relying on the interactive guided completion mechanism of the knowledge graph, business personnel do not need to have professional computing power knowledge. It automatically identifies and guides the completion of missing parameters, achieving a 100% completeness rate in describing business requirements. The traditional manual evaluation cycle of several days is compressed to minutes, greatly shortening the entire process cycle of power computing power application, review, and configuration, and improving the overall scheduling efficiency of computing power under the three-network collaboration.
[0130] 3. Significantly reduce computing power redundancy and save construction and operation costs: By using a three-layer knowledge graph semantic association and quantitative matching of measured baseline data, the drawbacks of subjective estimation based on human experience are completely eliminated. According to the "Comprehensive Evaluation Report on Intelligent Computing Centers (2024)" by the China Academy of Information and Communications Technology and the measured statistics of computing power scheduling of the State Grid Intelligent Computing Center, under the traditional manual experience selection and static form matching mode, the phenomenon of conservative over-configuration of computing power for power AI business is common, with the average redundancy of computing power generally reaching more than 60%. This invention, through an anti-over-configuration scoring mechanism and accurate fitting of baselines under all operating conditions, can accurately control the average redundancy of power computing power to within 30% under the premise of meeting the intensive construction requirements of the "Computing Power Standard System Construction Guide (2025 Edition)" and reserving the necessary fluctuation tolerance margin for power projects. This effectively reduces hardware idle waste, lowers the costs of hardware procurement, data center deployment, and daily operation and maintenance of intelligent computing centers, and helps to achieve intensive and refined utilization of power computing power resources.
[0131] 4. Achieve interpretable, reproducible, and auditable assessments: Construct a unified quantitative scoring and ranking standard, decompose the scores for each computing power selection result, and trace the basis for the results, solving the problems of traditional solutions lacking unified standards and being highly subjective and arbitrary. The entire assessment process, data, and algorithm rules are solidified and traceable, and the same inputs can stably reproduce the results, fully meeting the stringent management requirements of power project technical review, compliance audit, and post-event review.
[0132] Example 4 Based on the same inventive concept, this invention also provides a guided computing power requirement standard evaluation system based on knowledge graph technology, including: The knowledge reasoning layer is used to parse input keywords from natural language describing business requirements and match them with a pre-built edge-center two-level knowledge graph to obtain the algorithm model entity. The interactive guidance layer is used to guide the completion of entity information in the algorithm model using an interactive slot filling mechanism based on the entity in the algorithm model. The matching calculation layer is used to integrate the input keywords and the completed algorithm model entity information to generate a machine-recognizable structured business requirement profile; and based on the requirement profile, it retrieves the edge-center secondary knowledge graph to obtain a preliminary candidate platform set for the algorithm model. The results output layer is used to verify and score the suitability of algorithm models in the candidate platform set based on entity associations in the knowledge graph, and to quantify the accuracy of computing power selection and matching. The edge-center two-level knowledge graph is constructed based on business entities, the relationship between business entities and algorithm models, and the set constraints.
[0133] Furthermore, the knowledge reasoning layer is specifically used for; Natural language is parsed using named entity recognition and semantic role labeling techniques to obtain input keywords; Based on the input keywords, the system retrieves the corresponding business scenario entities in the pre-constructed edge-center two-level knowledge graph, and then matches them with the entities in the algorithm model.
[0134] Furthermore, the input keywords include: business scenario, processing object, and business task.
[0135] Furthermore, the matching computation layer is specifically used for: The input keywords and the completed algorithm model entity information are integrated together and transformed into a structured business requirement profile that can be recognized by machines. Based on the model classification and business hardware constraints in the demand profile, the model and hardware relationship within the edge-center secondary knowledge graph is retrieved to recall a preliminary candidate platform set. The candidate platform set is obtained by removing platforms that clearly violate latency and concurrency hardware constraints from the initial candidate platform set.
[0136] Furthermore, the output layer is specifically used for: Based on the relationships and attributes between entities in the knowledge graph, model adaptation score, concurrency adaptation score, latency adaptation score and throughput adaptation score are respectively applied to each combination in the candidate platform set. Based on the model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score, the comprehensive matching accuracy of each combination is calculated by weighted summation. The scoring rules for model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score are determined based on the hardware physical upper limit attribute, communication link measured performance attribute, and model running baseline attribute stored in the knowledge graph. This is a quantitative mapping of hardware physical constraints and communication physical constraints.
[0137] The guided computing power demand standard evaluation system based on knowledge graph technology of the present invention also includes: a data base layer, used to collect multi-source heterogeneous data and build a computing power baseline library based on the multi-source heterogeneous data.
[0138] The overall architecture of this invention consists of five layers: a data foundation layer, a knowledge reasoning layer, a guided interaction layer, a matching calculation layer, and a result output layer. This invention integrates four key technologies: baseline testing, knowledge reasoning, guided interaction, and interpretable scoring, addressing industry pain points such as the semantic gap between business operations and computing power, low efficiency of manual assessment, and high resource redundancy in the power sector.
[0139] The baseline library supports real-time querying and dynamic invocation, providing real-world data support for subsequent computing power matching. It protects the business scenario, algorithm model, and computing hardware through a three-layer knowledge graph architecture, defining entities, relationships, triples, and attribute systems to achieve semantic mapping from business language to computing power metrics.
[0140] The rule embedding method of the knowledge graph integrates SHACL constraint verification, SWRL logical reasoning, K-hop neighborhood expansion, and TransE vector reasoning rules, and has the capabilities of requirement integrity verification, semantic expansion, and logical reasoning.
[0141] The interactive slot filling guidance mechanism based on knowledge graphs predefines mandatory constraint slots for evaluation, automatically detects missing parameters through the graph, generates easy-to-understand guidance questions, and guides business personnel step by step to fill in key information such as concurrency, latency, and model specifications, achieving 100% completeness of requirement descriptions.
[0142] The slot data validity verification logic performs compliance checks on the parameters entered by the user, and automatically prompts for correction of abnormal parameters.
[0143] The computational power matching scoring algorithm uses a weighted scoring system with segmented penalties. It divides the scoring into four dimensions: model adaptation, concurrency adaptation, latency adaptation, and throughput adaptation, and sets differentiated full scores, scoring rules, and penalty coefficients.
[0144] The time delay dimension has a graded scoring logic, which introduces a 1.2 times lenient threshold to distinguish between three operating conditions: strict compliance, slight exceedance, and serious exceedance, and scores them separately, adapting to the actual application scenarios of power engineering.
[0145] The candidate platform's Top-K quantitative ranking mechanism sorts candidates from highest to lowest based on their comprehensive scores and simultaneously retains detailed scores for each dimension, ensuring that the results are interpretable and traceable.
[0146] Protect the standardized business requirement vector construction method, integrate the completed business constraints into a machine-recognizable structured requirement profile, and use it as a unified input standard for computing power matching.
[0147] Candidate platform screening mechanism: First, candidate platforms are recalled through semantic association, and then unqualified platforms are eliminated based on hard business constraints, thus narrowing the calculation scope and improving evaluation efficiency.
[0148] The entire process of data archiving and experiment reproduction mechanism ensures that the original evaluation data, calculation rules, and results are archived in a unified manner, so as to ensure that the evaluation results can be reproduced repeatedly under the same input, thus meeting the requirements of power project auditing and review.
[0149] The system offers a multi-format output mechanism, supporting the export of evaluation reports and computing power configuration lists in various formats, including web visualization, Excel, and JSON.
[0150] The system has a layered architecture, consisting of five layers: the guidance and interaction layer, the knowledge reasoning layer, the matching and calculation layer, the interface and output layer, and the data foundation layer, as well as the data and interface interaction logic between each layer.
[0151] Example 5 like Figure 3 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.
[0152] 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 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 edge-center two-level knowledge graph construction method or / and the guided computing power demand standard evaluation method based on knowledge graph technology in the above embodiments.
[0153] Example 6 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 edge-center two-level knowledge graph construction method and / or the guided computing power demand standard evaluation method based on knowledge graph technology described in the above embodiments.
[0154] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present 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.
[0155] This invention 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0156] 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.
[0157] 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.
[0158] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.
Claims
1. A method for constructing an edge-center two-level knowledge graph, characterized in that, include: Extract entities and relationships between entities from multi-source heterogeneous data to generate an initial set of triples; The initial set of triples is divided into edge shard triples and center shard triples according to the hardware deployment level, and cross-shard connection relationships are established. Entity alignment and knowledge fusion are performed on the edge segment triples, the center segment triples, and the cross-segment connection relationships to obtain a standardized triple set. The standardized triple set is imported into a graph database, stored in the form of nodes and edges, and constraints are embedded to construct an edge-center two-level knowledge graph.
2. The method as described in claim 1, characterized in that, The step of dividing the initial set of triples into edge shard triples and center shard triples according to the hardware deployment level, and establishing cross-shard connections, includes: The triplet consisting of localized power services of substations and distribution area edge nodes, lightweight edge AI models, edge NPU computing power performance, and local low latency constraints is divided into edge sharding triplets. The triplet consisting of cross-regional power business under the overall network scheduling, cloud-based MoE large model, GPU cluster computing power indicators, unified computing power baseline of the entire network, and compliance specifications for three-network collaboration is divided into central sharding triplets. Assign a first cross-shard identifier to each entity in each edge shard triplet, and assign the same second cross-shard identifier to entities in each center shard triplet that have a business relationship with that entity; based on the consistency matching of the first cross-shard identifier and the second cross-shard identifier, establish an index association relationship between edge shards and center shards, and generate a shard association mapping table.
3. The method as described in claim 1, characterized in that, The entity alignment and knowledge fusion of the edge fragment triples, the center fragment triples, and the cross-fragment connections yield a standardized triple set, including: Local terminology unification, entity disambiguation and data deduplication are performed on edge-segmented triples to standardize the computing power parameters corresponding to fuzzy constraints in power business. The central sharding triplet integrates the multi-source computing power baseline data, model parameters, and hardware indicators of the entire network, supplements the P95 tail latency, TTFT, and TPOT long-tail performance indicators, and corrects the adaptation deviation between the general computing power baseline and the power business scenario. Based on cross-shard identifiers, shard consistency verification is performed, and the fuzzy business requirements of the power business side are automatically mapped into standardized computing power technical indicators, forming a standardized triplet set containing five layers: "business scenario - AI model - computing power hardware - performance indicators - compliance constraints".
4. The method as described in claim 1, characterized in that, The process of importing the standardized triple set into a graph database, storing it as nodes and edges, and embedding constraints to construct an edge-center two-level knowledge graph includes: The edge-sharded triples in the standardized triple set are stored in a lightweight graph database; The central sharded triples in the standardized triple set are stored using an enterprise-level graph database. Using entities as nodes and the relationships between entities as edges, an edge knowledge graph is constructed from edge-sharded triples. Using entities as nodes and the relationships between entities as edges, a central knowledge graph is constructed from central partitioned triples. The cross-segment identifiers between the edge and center are used as the association between the edge knowledge graph and the center knowledge graph. Combined with constraints, a two-level edge-center knowledge graph is constructed.
5. The method as described in claim 1, characterized in that, Also includes: Based on the actual load test data of real load test conditions that are iteratively updated based on the computing power performance baseline library, the relationship between entities and parameter indicators stored in the edge-center secondary knowledge graph is compared in real time, and deviation items between theoretical graph rules and real load test conditions are automatically identified. Based on the relationship between the deviation term and the triple, the semantic association weights, constraint threshold ranges, and business adaptation mapping relationships are automatically and iteratively corrected, and the model-level adaptation rules and hardware performance matching boundaries are updated synchronously.
6. A guided computing power requirement standard evaluation method based on knowledge graph technology, characterized in that, include: The algorithm model entities are obtained by parsing input keywords from natural language describing business requirements and matching them with a pre-built edge-center two-level knowledge graph. An interactive slot-filling mechanism is used to guide the completion of entity information in the algorithm model. By integrating the input keywords and the completed algorithm model entity information, a structured business requirement profile that can be recognized by a machine is generated. Based on demand profiles, a preliminary set of candidate platforms for the algorithm model is obtained by retrieving the edge-center secondary knowledge graph. The suitability of algorithm models in the candidate platform set is evaluated based on the entity association relationships in the knowledge graph; The edge-center two-level knowledge graph is constructed based on a construction method for an edge-center two-level knowledge graph as described in any one of claims 1-5.
7. The method as described in claim 6, characterized in that, The evaluation of the algorithm model suitability in the candidate platform set based on entity relationships in the knowledge graph includes: Based on the relationships and attributes between entities in the knowledge graph, model adaptation score, concurrency adaptation score, latency adaptation score and throughput adaptation score are respectively applied to each combination in the candidate platform set. Based on the model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score, the comprehensive matching accuracy of each combination is calculated by weighted summation. The scoring rules for model adaptation score, concurrency adaptation score, latency adaptation score, and throughput adaptation score are determined based on the hardware physical upper limit attribute, communication link measured performance attribute, and model running baseline attribute stored in the knowledge graph. This is a quantitative mapping of hardware physical constraints and communication physical constraints.
8. A guided computing power demand standard evaluation system based on knowledge graph technology, characterized in that: include: The knowledge reasoning layer is used to parse input keywords from natural language describing business requirements and match them with a pre-built edge-center two-level knowledge graph to obtain the algorithm model entity. The interactive guidance layer is used to guide the completion of entity information in the algorithm model using an interactive slot filling mechanism based on the entity in the algorithm model. The matching calculation layer is used to integrate the input keywords and the completed algorithm model entity information to generate a machine-recognizable structured business requirement profile. Based on the demand profile, the edge-center secondary knowledge graph is retrieved to obtain a preliminary set of candidate platforms for the algorithm model; The results output layer is used to verify and score the suitability of algorithm models in the candidate platform set based on entity associations in the knowledge graph, and to quantify the accuracy of computing power selection and matching. The edge-center two-level knowledge graph is constructed based on a construction method for an edge-center two-level knowledge graph as described in any one of claims 1-5.
9. 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, the method for constructing an edge-center two-level knowledge graph as described in any one of claims 1 to 5, and the guided computing power requirement standard evaluation method based on knowledge graph technology as described in any one of claims 6 to 7 are implemented.
10. A readable storage medium, characterized in that, It contains an execution program, which, when executed, implements a method for constructing an edge-center two-level knowledge graph as described in any one of claims 1 to 5, and a guided computing power requirement standard evaluation method based on knowledge graph technology as described in any one of claims 6 to 7.
