Heterogeneous Algorithm Scheduling Method and Apparatus for Power Systems
By standardizing and dynamically scheduling heterogeneous capability objects in the power system, the problem of automatic and safe scheduling of large-scale intelligent agents with heterogeneous algorithms in the power system is solved. This achieves fully automated scheduling from task understanding to algorithm combination, thereby improving the scheduling automation and security of the power system.
Patent Information
- Application Number
- CN202611482841.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-22
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies have failed to effectively solve the problem of automatic and secure scheduling of heterogeneous algorithms by large-scale intelligent agents in power systems, especially in complex scenarios with a high proportion of renewable energy grid connection and deep interaction between power sources, grids, loads and storage. It is difficult to achieve unified encapsulation and invocation of multiple types of heterogeneous algorithms and dynamic discovery, orchestration and secure scheduling.
By profiling heterogeneous capability objects in the power system, generating structured capability description files, and encapsulating them into standardized capability units with unified standardized interfaces, these units are registered in the algorithm capability directory. Based on the structured task descriptions, candidate capability units are automatically discovered, algorithm execution plans are generated, and dynamic scheduling of the master control agent is realized.
It realizes automatic and safe scheduling of large-scale intelligent agents with heterogeneous algorithms in power systems, eliminating the burden of adapting each agent one by one and writing integration code separately, improving the level of scheduling automation and security, and ensuring reliability and flexibility in complex scenarios.
Smart Images

Figure CN123068043A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system technology, and for example to a heterogeneous algorithm scheduling method and apparatus for power systems. Background Technology
[0002] Large-scale model intelligent agent platforms are widely regarded as an important technical route to improve the automation and intelligence level of power grid operations. ReAct (Reasoning and Acting) frameworks, representing large-scale model intelligent agents, combine the language understanding and reasoning capabilities of large models with the execution capabilities of external tools through a "think-act-observe" cycle, enabling them to decompose and execute complex power dispatching tasks to a certain extent. However, new power systems are exhibiting significant characteristics such as a high proportion of renewable energy grid integration, deep interaction between power sources, grids, loads, and storage, and complex and variable operating conditions. Core business scenarios such as substation operation, distribution area management, grid dispatching, equipment inspection, customer service, and emergency response require the comprehensive utilization of various specialized algorithms, including power flow calculation, load forecasting, fault diagnosis, image recognition, time-series anomaly detection, knowledge reasoning, natural language processing, and multi-objective optimization. These algorithms are typically developed by different teams on different platforms using different languages or frameworks, and provide services externally through various means such as database interfaces, message queues, REST APIs, traditional Web Services, file exchange, and simulation program calls. When a large model intelligent agent platform needs to call these heterogeneous algorithms, it must adapt the interface protocol, data format and calling method of each algorithm one by one, write integration code separately for each algorithm, and manually maintain the availability status and version information of the algorithm.
[0003] Existing technologies have proposed solutions for encapsulating heterogeneous algorithms to achieve unified encapsulation and invocation of multiple types of heterogeneous algorithms. For example, patent application CN 202511736229.6 discloses a general model inference service framework that supports the encapsulation of multiple types of heterogeneous algorithms. Utilizing an algorithm adaptation layer and adopting a plug-in architecture design, it designs corresponding algorithm adapters for different types of algorithms, achieving unified encapsulation of multiple types of heterogeneous algorithms such as sklearn, PyTorch, and YOLO. External systems can call any model inference service through a standardized API interface.
[0004] In the process of implementing the embodiments of the present invention, at least the following problems were found in the related art: Existing solutions address the issue of unified invocation of machine learning models at the protocol level through algorithm adapters. However, they essentially only provide static encapsulation capabilities for locally deployed model inference algorithms. They do not cover heterogeneous algorithms and business systems in the power grid, nor do they provide dynamic discovery, orchestration, and secure scheduling mechanisms for power grid business tasks. This makes it difficult for large model agents to complete the automatic secure scheduling of heterogeneous algorithms in the power system. Summary of the Invention
[0005] To address the shortcomings of the existing technologies, this invention provides a method and apparatus for scheduling heterogeneous algorithms in power systems, enabling large-scale intelligent agents to automatically and securely schedule heterogeneous algorithms in power systems.
[0006] According to one aspect of the present invention, a heterogeneous algorithm scheduling method for a power system is provided, applied to a large model intelligent agent platform, comprising: profiling multiple heterogeneous capability objects in the power system and generating a structured capability description file for each heterogeneous capability object; wherein the heterogeneous capability objects include heterogeneous algorithms and heterogeneous business services for power data acquisition, power equipment operation, or grid dispatching business process processing; based on the structured capability description file, encapsulating the heterogeneous capability objects into standardized capability units with a unified standardized interface, and registering the standardized capability units to an algorithm capability directory; parsing received business task requests into structured task descriptions, and determining a set of candidate standardized capability units from the algorithm capability directory based on the structured task descriptions; generating an algorithm execution plan including sequential, parallel, or conditional branch structures according to the set of candidate standardized capability units and the orchestration constraint information in the structured task descriptions, so that the master control agent in the large model intelligent agent platform calls each standardized capability unit according to the algorithm execution plan; wherein each node in the algorithm execution plan corresponds to the call of a standardized capability unit.
[0007] Optionally, capability profiles are created for multiple heterogeneous capability objects in the power system, generating a structured capability description file for each heterogeneous capability object. This includes: establishing an initial capability description file for each heterogeneous capability object using a structured format; wherein the initial capability description file includes the functional description, input / output specifications, and data dependencies of the heterogeneous capability object; obtaining and determining semantic tags representing the functional types of the heterogeneous capability objects and applicable power grid business scenarios for the initial capability description file based on a preset set of power grid business tags, and writing the semantic tags into the initial capability description file to generate an intermediate capability description file; and writing the interface protocol type, performance indicators, permission levels, and security constraints of the heterogeneous capability objects into the intermediate capability description file to generate the final structured capability description file.
[0008] Optionally, based on the structured capability description file, heterogeneous capability objects are encapsulated into standardized capability units with unified standardized interfaces. This includes performing layer-by-layer encapsulation of heterogeneous capability objects using a three-layer encapsulation structure of bottom, middle, and top layers. At the bottom layer, according to the interface protocol type in the structured capability description file, a corresponding adapter is matched for the heterogeneous capability object to convert the raw response data output by the heterogeneous capability object using the native interface protocol into a unified first intermediate data format. At the middle layer, for the core objects in the first intermediate data format, field mapping is performed according to a unified structure expression, and professional attribute normalization is performed according to the general model of the power industry to generate a second intermediate data format with unified semantics. At the top layer, based on the second intermediate data format and the functional description, input / output specifications, and security constraints in the structured capability description file, standardized capability units for large-scale model intelligent agent platforms are generated.
[0009] Optionally, the structured task description includes: task target keywords representing the intent of power dispatching operations or power grid equipment objects; data requirement fields representing the measurement data required to perform power dispatching operations; and constraints representing the power grid operation safety rules and dispatch response timeliness requirements corresponding to the power dispatching operations. Based on the structured task description, a candidate set of standardized capability units is determined from the algorithm capability catalog, including: assembling the task target keywords and data requirement fields into functional semantic query conditions and data requirement query conditions according to a preset query semantic template; performing semantic matching between the functional semantic query conditions and the semantic tags corresponding to each standardized capability unit in the algorithm capability catalog to obtain an initial set of standardized capability units; performing availability filtering on each standardized capability unit in the initial set of standardized capability units according to the data requirement query conditions to obtain an intermediate set of standardized capability units; determining the professional suitability score of each standardized capability unit in the intermediate set of standardized capability units based on the semantic tags, historical reliability scores, and constraints; and determining the standardized capability units in the intermediate set of standardized capability units whose professional suitability scores meet a preset threshold as a set of candidate capability units.
[0010] Optionally, based on semantic tags, historical reliability scores, and constraints, a professional fit score for the standardized capability unit is determined, including: mapping semantic tags to semantic matching vectors, mapping historical reliability scores to reliability feature vectors, and encoding constraints to constraint satisfaction vectors; performing nonlinear transformations on the semantic matching vectors, reliability feature vectors, and constraint satisfaction vectors respectively to obtain semantic latent vectors, reliability latent vectors, and constraint latent vectors; wherein the semantic latent vectors and reliability latent vectors have the same vector dimension; calculating the attention weight between the semantic latent vectors and reliability latent vectors, and using the attention weight to perform weighted fusion of the semantic latent vectors and reliability latent vectors to obtain a fused latent vector; concatenating the fused latent vector with the constraint latent vector to obtain a concatenated vector, and performing a fully connected transformation and activation processing on the concatenated vector to obtain a professional fit score representing the comprehensive matching degree of the standardized capability unit to the current business task.
[0011] Optionally, determining the candidate capability unit set from the algorithm capability catalog based on the structured task description further includes: when the professional suitability scores of each standardized capability unit in the intermediate standardized capability unit set are all lower than a preset threshold, performing semantic expansion on the functional semantic query conditions to obtain extended query conditions; based on the extended query conditions, re-performing semantic matching and availability filtering on each standardized capability unit in the algorithm capability catalog to obtain a supplementary standardized capability unit set; merging the supplementary standardized capability unit set with the intermediate standardized capability unit set, and re-performing professional suitability scoring and threshold filtering to obtain a candidate standardized capability unit set.
[0012] Optionally, the orchestration constraint information includes task timing relationships, preconditions, postconditions, and power grid operation safety constraints. Based on the candidate capability unit set and the orchestration constraint information in the structured task description, an algorithm execution plan including sequential, parallel, or conditional branch structures is generated, including: mapping each standardized capability unit in the candidate capability unit set to a node in the algorithm execution plan according to the task timing relationships, and determining the initial dependencies between nodes; determining the data dependencies between nodes according to the preconditions and postconditions, marking nodes with data dependencies as sequential execution nodes, and marking nodes without data dependencies as parallel execution branches; inserting safety verification nodes before nodes with high-risk equipment operation according to power grid operation safety constraints, and configuring branch triggering conditions for nodes with conditional branches; assembling the initial dependencies, sequential and / or parallel markings, safety verification nodes, and branch triggering conditions into a structured algorithm execution plan.
[0013] Optionally, based on the candidate capability unit set and the orchestration constraint information in the structured task description, an algorithm execution plan including sequential, parallel, or conditional branch structures is generated. This further includes: for each node in the algorithm execution plan, selecting standardized capability units that meet the functional requirements from the candidate capability unit set according to the node's corresponding functional requirements, obtaining a candidate unit subset for the node; when the candidate unit subset includes multiple standardized capability units capable of performing the same function, obtaining the data quality assessment value and computational resource utilization rate of the data source corresponding to each standardized capability unit; prioritizing each standardized capability unit in the candidate unit subset based on the data quality assessment value and computational resource utilization rate, determining the standardized capability unit with the highest priority as the node's primary standardized capability unit, and writing the standardized capability units in the candidate unit subset other than the primary standardized capability unit as backup standardized capability units into the node's exception handling strategy; wherein, the exception handling strategy is used to sequentially select backup standardized capability units to initiate retry in the event of a failure in calling the primary standardized capability unit.
[0014] Optionally, the heterogeneous algorithm scheduling method for power systems further includes: before scheduling each standardized capability unit sequentially according to the algorithm execution plan, reading the caller identity information corresponding to the business task request, and reading the permission level and security constraints from the structured capability description file corresponding to the current standardized capability unit, and performing identity authentication and minimum permission verification; after the verification is passed, performing parameter legality checks on the input parameters to be passed according to the input specifications corresponding to the current standardized capability unit; after the check is passed, initiating a call to the current standardized capability unit, and monitoring the response latency and execution status of the current standardized capability unit during the call process to trigger the node anomaly handling strategy corresponding to the current standardized capability unit; after the call returns, performing integrity verification, consistency verification, and semantic verification on the returned data sequentially, and after all three verifications pass, returning a success status to the master control agent in the large model intelligent agent platform so that the master control agent can schedule the next node.
[0015] According to another aspect of the present invention, the present invention also provides a heterogeneous algorithm scheduling device for a power system, applied to a large model intelligent agent platform, comprising: a capability profiling module, configured to perform capability profiling on multiple heterogeneous capability objects in the power system and generate a structured capability description file for each heterogeneous capability object; wherein the heterogeneous capability objects include heterogeneous algorithms and heterogeneous business services for power data acquisition, power equipment operation, or grid dispatching business process processing; and a heterogeneous capability object encapsulation module, connected to the capability profiling module, configured to encapsulate the heterogeneous capability objects into standardized capability units with unified standardized interfaces based on the structured capability description files, and register the standardized capability units. The algorithm capability directory is accessed; the task parsing module, connected to the heterogeneous capability object encapsulation module, is configured to parse the received business task request into a structured task description, and determine a set of candidate standardized capability units from the algorithm capability directory based on the structured task description; the algorithm execution plan generation module, connected to the task parsing module, is configured to generate an algorithm execution plan, including sequential, parallel, or conditional branch structures, based on the set of candidate standardized capability units and the orchestration constraint information in the structured task description, so that the master agent in the large model agent platform can call each standardized capability unit according to the algorithm execution plan; wherein, each node in the algorithm execution plan corresponds to the call of a standardized capability unit.
[0016] The heterogeneous algorithm scheduling method and apparatus for power systems provided by this invention have at least the following beneficial technical effects: In the technical solution of this invention, firstly, by structurally profiling and uniformly standardizing the encapsulation of heterogeneous capability objects, heterogeneous algorithms and business services with different protocols, data formats, and security levels are transformed into standardized capability units with consistent interfaces and semantic alignment, fundamentally eliminating the burden of agents adapting one by one and writing integration code separately. Based on this, through unified registration and semantic retrieval of the algorithm capability catalog, the controlling agent can automatically discover and filter available candidate standardized capability unit sets according to business task requests, and dynamically generate algorithm execution plans containing sequential, parallel, or conditional branch structures based on the orchestration constraints in the structured task description. This allows the collaborative invocation of multiple algorithms and multiple business services to be flexibly organized according to the actual needs of the task, rather than relying on a fixed static process. Finally, the controlling agent sequentially calls each standardized capability unit according to the algorithm execution plan, achieving fully automated scheduling from task understanding to algorithm combination to secure execution. Thus, a fully automated closed loop is achieved at the unified semantic level, from task parsing, capability discovery, plan orchestration to sequential invocation, and security constraints are incorporated into every standardized capability unit invocation, enabling large-scale intelligent agents to automatically complete the scheduling of heterogeneous algorithms in the power system under safe and controllable conditions. Attached Figure Description
[0017] Figure 1 A flowchart illustrating a heterogeneous algorithm scheduling method for a power system provided in an embodiment of the present invention; Figure 2 A flowchart illustrating another heterogeneous algorithm scheduling method for power systems provided in an embodiment of the present invention; Figure 3 A flowchart illustrating another heterogeneous algorithm scheduling method for power systems provided in an embodiment of the present invention; Figure 4 A flowchart illustrating another heterogeneous algorithm scheduling method for power systems provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an algorithm execution plan provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a heterogeneous algorithm scheduling device for a power system provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of another heterogeneous algorithm scheduling device for power systems provided in an embodiment of the present invention. Detailed Implementation
[0018] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0019] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.
[0020] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.
[0021] The following describes the key terms in the embodiments of the present invention in order to better understand the technical solutions of the present invention.
[0022] A large-model intelligent agent platform refers to a software runtime environment that uses a large language model as its core inference engine and possesses capabilities for task understanding, tool invocation, planning and decision-making, and execution feedback. In this embodiment of the invention, the large-model intelligent agent platform serves as the runtime carrier for heterogeneous algorithm scheduling methods, providing inference and decision-making capabilities and runtime support environment for capability discovery, dynamic orchestration, and unified scheduling.
[0023] The master control agent refers to an agent instance deployed in a large-scale model agent platform that is responsible for the full lifecycle management of tasks. In this embodiment of the invention, the master control agent is the decision-making body for the dynamic orchestration and secure scheduling of heterogeneous algorithms. It is responsible for coordinating the calling order and exception handling strategies of various standardized capability units at the task level, and dynamically confirming or adjusting the scheduling of the next node based on the returned execution status.
[0024] Understandably, in addition to the master control agent, the large-scale model agent platform also includes a unified scheduling runtime engine. The unified scheduling runtime engine is responsible for initiating calls to each standardized capability unit sequentially according to the algorithm execution plan under the decision instructions of the master control agent; performing identity authentication, least privilege verification, and parameter validity checks before the call; monitoring response latency and execution status during the call; and performing integrity verification, consistency verification, and semantic verification after the call. The unified scheduling runtime engine returns a success status or exception information to the master control agent, which then determines the scheduling of the next node based on the returned information.
[0025] Combination Figure 1 As shown, this embodiment of the invention provides a heterogeneous algorithm scheduling method for power systems, applied to a large model intelligent agent platform, including the following steps: S10: Create capability profiles for multiple heterogeneous capability objects in the power system and generate a structured capability description file for each heterogeneous capability object.
[0026] Among them, heterogeneous capability objects include heterogeneous algorithms and heterogeneous business services used for power data acquisition, power equipment operation or grid dispatch business process processing.
[0027] Heterogeneous algorithms are used to complete data calculation, pattern recognition, or optimization tasks in specific power business scenarios, including N-1 security verification algorithm, regulation resource assessment algorithm, load forecasting algorithm, fault diagnosis algorithm, image recognition algorithm, time series anomaly detection algorithm, knowledge reasoning algorithm, natural language processing algorithm, and multi-objective optimization algorithm.
[0028] Heterogeneous business services refer to business systems and their external interfaces used in power systems for power data acquisition, power equipment operation, or grid dispatch business process processing. Their core function is to complete specific business actions in grid operation, including real-time measurement query, power flow calculation program, dispatch operation ticket generation system, simulation and pre-drilling system, and operation approval process in the dispatch automation system (SCADA / EMS).
[0029] S20, based on the structured capability description file, encapsulates heterogeneous capability objects into standardized capability units with a unified standardized interface, and registers the standardized capability units to the algorithm capability directory.
[0030] Here, a standardized capability unit refers to a capability invocation unit with a unified standardized interface, formed by encapsulating heterogeneous capability objects through a unified interface. Standardized capability units provide the upper-layer master control agent with a consistent interface, semantic alignment, and clearly defined security constraints as an invocation entry point, enabling the master control agent to initiate invocations of heterogeneous algorithms or heterogeneous business services at a unified semantic level without needing to be aware of underlying protocol differences.
[0031] The Algorithm Capability Directory is a directory service used for the unified registration and management of all standardized capability units. It stores a structured capability description file and its semantic tags for each standardized capability unit, and supports searching and matching by dimensions such as functional semantics, applicable scenarios, input / output types, performance metrics, and permission levels.
[0032] S30: Parse the received business task request into a structured task description, and determine a set of candidate standardized capability units from the algorithm capability catalog based on the structured task description.
[0033] S40 generates an algorithm execution plan, including sequential, parallel, or conditional branch structures, based on the candidate standardized capability unit set and the orchestration constraint information in the structured task description, so that the master agent in the large model agent platform can call each standardized capability unit according to the algorithm execution plan.
[0034] Each node in the algorithm execution plan corresponds to the invocation of a standardized capability unit.
[0035] By employing the heterogeneous algorithm scheduling method for power systems provided in this disclosure, the large-scale intelligent agent can automatically discover and combine heterogeneous algorithms and business services with different protocols and semantics in the power system at a unified semantic level, without having to write integration code separately for each algorithm or rely on manually preset fixed processes. At the same time, since security constraints are bound to the entire process of encapsulation and invocation of standardized capability units, the master control agent is always within the controllable boundary in the execution of power equipment operation and high-risk scheduling business processes, thereby improving the level of scheduling automation while ensuring the safety and reliability of the large-scale intelligent agent in complex power grid scenarios.
[0036] Optionally, in step S10, capability profiles are created for multiple heterogeneous capability objects in the power system, and a structured capability description file is generated for each heterogeneous capability object. This includes: establishing an initial capability description file for each heterogeneous capability object using a structured format; wherein the initial capability description file includes the functional description, input / output specifications, and data dependencies of the heterogeneous capability object; obtaining and determining semantic tags representing the functional type and applicable power grid business scenarios of the heterogeneous capability object for the initial capability description file based on a preset set of power grid business tags, and writing the semantic tags into the initial capability description file to generate an intermediate capability description file; and writing the interface protocol type, performance indicators, permission level, and security constraints of the heterogeneous capability object into the intermediate capability description file to generate the final structured capability description file.
[0037] Here, the functional description of a heterogeneous capability object refers to a brief textual description of the power business functions or calculation logic implemented by the heterogeneous capability object. For example, the functional description of the N-1 safety verification algorithm is "for a given grid section, simulate the power flow distribution after a single component is taken out of operation, and determine whether there are other components exceeding limits or voltage anomalies." The functional description of the SCADA measurement query service is "to obtain real-time or historical measurement data such as active power, reactive power, voltage amplitude, and switch position of a specified substation or equipment within a specified time range." The input and output specifications refer to the field names of the input data items and output data items required by the heterogeneous capability object when performing its functions. The structured definition of data types, value ranges, and mandatory requirements; for example, the input specifications for the N-1 safety verification algorithm include the power grid topology model (CIM / E format), real-time or planned sections, a list of components to be verified, and safety limits; the output specifications include the N-1 over-limit equipment list, the degree of over-limit, critical sections, and verification conclusions. Data dependency refers to the data source, data format, and data freshness requirements that heterogeneous capability objects must obtain from external sources before execution. For example, the data dependency of the power flow calculation program is "dependent on real-time measurement data provided by the SCADA system and CIM / E topology files provided by the power grid model system, with a data freshness requirement of no more than 5 minutes."
[0038] Semantic tags for heterogeneous capability objects refer to standardized labels that characterize the functional type and applicable power grid business scenarios of heterogeneous capability objects based on a pre-defined set of power grid business tags. For example, the semantic tag for the N-1 safety verification algorithm is "{safety verification, power flow calculation, N-1, scheduling operation}".
[0039] The interface protocol type of a heterogeneous capability object refers to the technical protocol or calling method used when the heterogeneous capability object provides services to the outside world. For example, the interface protocol type of the N-1 security verification algorithm is "REST API, returning JSON format". The performance index refers to the time characteristics, throughput characteristics and resource consumption characteristics of the heterogeneous capability object during normal operation. For example, the performance index of the N-1 security verification algorithm is "average response latency of 15 seconds and maximum concurrency of 5". The permission level refers to the minimum identity permission level required to call the heterogeneous capability object. For example, the permission level of the N-1 security verification algorithm is "advanced scheduling permission". The security constraint refers to the security restrictions that the heterogeneous capability object must comply with during the call. For example, the security constraint of the N-1 security verification algorithm is "only reading power grid model and measurement data is allowed, and direct issuance of control commands is not allowed".
[0040] In some possible implementations, based on a pre-defined set of power grid business tags, semantic tags representing the functional types of heterogeneous capability objects and applicable power grid business scenarios are determined for the initial capability description file. This includes: obtaining the input / output specifications and data dependency fields from the initial capability description file; matching the input / output specifications and data dependency fields with the pre-defined set of power grid business tags to determine the functional type tags corresponding to the heterogeneous capability objects; wherein the power grid business tag set adopts a hierarchical structure, including computational tags, data acquisition tags, equipment operation tags, and process processing tags; based on the functional type tags, determining the applicable power grid business scenario tags for the heterogeneous capability objects from the hierarchical structure, and using the functional type tags and power grid business scenario tags together as semantic tags. By matching the input / output specifications and data dependency fields with the hierarchical set of power grid business tags, first determining the functional type tags, and then determining the applicable scenario tags accordingly, the standardization and consistency of semantic tags are ensured. This enables the large model intelligent agent platform to achieve accurate retrieval and scenario matching of heterogeneous capability objects based on a unified semantic tag system in the subsequent capability discovery stage.
[0041] In some possible implementations, the interface protocol type, performance metrics, permission levels, and security constraints of heterogeneous capability objects are written into the intermediate capability description file to generate the final structured capability description file. This includes: obtaining the interface protocol type of the heterogeneous capability object and writing it into the intermediate capability description file to trigger the matching of the corresponding protocol adapter during the encapsulation phase; obtaining the average response latency, maximum concurrency, and data freshness requirements of the heterogeneous capability object in historical calls and writing these as performance metrics into the intermediate capability description file to trigger availability filtering during the capability discovery phase; and obtaining the permission level and security constraints of the heterogeneous capability object and writing them into the intermediate capability description file to trigger security verification strategies before scheduling calls. By writing the interface protocol type, performance metrics, permission levels, and security constraints into the intermediate capability description file, these parameters are transformed into structured control parameters that trigger protocol adapter matching during the encapsulation phase, availability filtering during the capability discovery phase, and security verification strategies before scheduling calls, providing a unified data-driven foundation for the automatic access and secure scheduling of heterogeneous capability objects.
[0042] For example, Table 1 below shows the structured capability description file table for the N-1 security verification algorithm.
[0043] Table 1. Structured capability description file of the N-1 security verification algorithm
[0044] Similarly, capability description files can be created for "power flow calculation program", "regulation resource assessment algorithm", "multi-objective optimization algorithm", "operation ticket generation system", "simulation pre-play system", etc., and registered in the algorithm capability directory.
[0045] The structured format uniformly records the functional descriptions, input / output specifications, and data dependencies of heterogeneous capability objects. It also assigns semantic tags representing functional types and applicable scenarios to these objects through a pre-defined set of power grid business tags. This enables heterogeneous algorithms and business services, which were originally semantically diverse and difficult to understand, to be identified and retrieved by the large model intelligent agent platform at a unified semantic level. At the same time, by writing interface protocol types, performance indicators, permission levels, and security constraints into the final structured capability description file, subsequent protocol adaptation, capability discovery, and pre-call security verification can all be automatically triggered directly based on this capability description file. This provides a machine-readable, semantically clear, and security-informed data foundation for the unified encapsulation and secure scheduling of heterogeneous capability objects.
[0046] Optionally, in step S20, based on the structured capability description file, heterogeneous capability objects are encapsulated into standardized capability units with unified standardized interfaces. This includes performing layer-by-layer encapsulation of heterogeneous capability objects using a three-layer encapsulation structure of bottom, middle, and top layers. At the bottom layer, according to the interface protocol type in the structured capability description file, a corresponding adapter is matched for the heterogeneous capability object to convert the raw response data output by the heterogeneous capability object using the native interface protocol into a unified first intermediate data format. At the middle layer, for the core objects in the first intermediate data format, field mapping is performed according to a unified structure expression, and professional attribute normalization is performed according to the general model of the power industry to generate a second intermediate data format with unified semantics. At the top layer, based on the second intermediate data format and the functional descriptions, input / output specifications, and security constraints in the structured capability description file, standardized capability units for large-scale intelligent agent platforms are generated.
[0047] The following uses the N-1 safety verification algorithm as an example to illustrate the three-layer encapsulation process of heterogeneous capability objects. The original form of the N-1 safety verification algorithm is a REST API service deployed by a certain vendor. It requires a power grid model and cross-section file in a specific XML format as input, and outputs a complex JSON containing many vendor-private fields. Low-level encapsulation: The structured capability description file records the interface protocol type of the N-1 safety verification algorithm as "REST API, returning JSON format." Based on this, the low-level layer matches the corresponding REST API protocol adapter for the N-1 safety verification algorithm. When the N-1 safety verification algorithm is called, its private JSON response, output in the native protocol, is converted by the adapter into a unified first intermediate data format, thus shielding the protocol differences in field naming and data structure between different vendors. Middle-level encapsulation: Field mapping is performed on the core objects (equipment, measurement, topology, etc.) in the first intermediate data format, and professional attributes are standardized according to the common model of the power industry; for example, the vendor's private p... flow Mapped to the standard field activePowerMW, the per-unit value v rate =0.98, converted to engineering unit voltage = 215.6kV based on the reference voltage, and then the equipment identification line is set. A The data is standardized into a 220kV XX line conforming to the unified naming convention of the power grid, thereby generating a second intermediate data format with unified semantics. Upper-level encapsulation: Based on the second intermediate data format, combined with the functional description ("simulating power flow distribution after N-1 faults for a given section"), input and output specifications (input is the topology model and section data, output is a list of out-of-limit devices) and safety constraints ("only reading model and measurement data is allowed, and direct issuance of control commands is not allowed") in the structured capability description file, a standardized tool description for the large model intelligent agent platform is generated, ultimately forming a standardized capability unit for the N-1 safety verification algorithm.
[0048] Through a three-layer encapsulation approach—lower-layer protocol adaptation, middle-layer semantic normalization, and upper-layer interface exposure—the raw data output by heterogeneous capability objects using different native protocols is uniformly converted into an intermediate format with unified semantics. This ultimately generates standardized capability units for large-scale intelligent agent platforms. As a result, when large-scale intelligent agents call various heterogeneous algorithms and business services, they can complete cross-system and cross-professional capability calls at a semantically consistent level through a unified standardized interface. This significantly reduces the integration cost between intelligent agents and heterogeneous power systems and improves data consistency and call reliability during the scheduling process.
[0049] Optionally, in step S30, the structured task description includes: task target keywords characterizing the intent of power dispatching operations or power grid equipment objects; data requirement fields characterizing the measurement data required to perform power dispatching operations; and constraints characterizing the power grid operation safety rules and dispatching response time requirements corresponding to the power dispatching operations. For example, Table 2 below shows a structured task description table corresponding to a business task request "The power flow of a certain 220kV line A phase exceeds the limit; please provide handling suggestions."
[0050] Table 2. Structured Task Description Table
[0051] Combination Figure 2 As shown, based on the structured task description, the set of candidate standardized capability units is determined from the algorithm capability catalog, including the following steps: S301, assemble the task target keywords and data requirement fields into functional semantic query conditions and data requirement query conditions according to the preset query semantic template.
[0052] For example, the large-scale intelligent agent platform parses business task requests into structured task descriptions. The task objective keywords are "power flow overrun handling" and "220kV line"; the data requirement fields are "real-time measurement data (active power, reactive power, voltage amplitude), time range within the last 5 minutes of the current moment, spatial range within the line and associated substations"; and the constraints are "dispatch response delay not exceeding 30 seconds," "equipment operation requires advanced dispatching privileges," and "must not cause new equipment overruns." The master control agent assembles the above information into two sets of query conditions according to a preset query semantic template: the functional semantic query condition is "power flow overrun handling + line equipment," and the data requirement query condition is "measurement type = active / reactive / voltage, data time range = last 5 minutes of the current moment, topology range = line and associated substations." S302, perform semantic matching between the functional semantic query conditions and the semantic tags corresponding to each standardized capability unit in the algorithm capability catalog to obtain the initial set of standardized capability units.
[0053] For example, the controlling agent matches the functional semantic query conditions with the semantic tags of each standardized capability unit in the algorithm capability catalog. For instance, the semantic tag for "Power Flow Calculation Program" is "{Power Flow Calculation, Scheduling Operation}", the semantic tag for "N-1 Safety Verification Algorithm" is "{Safety Verification, Power Flow Calculation, N-1, Scheduling Operation}", the semantic tag for "Regulation Resource Assessment Service" is "{Resource Assessment, Scheduling Operation}", and the semantic tag for "Image Recognition Algorithm" is "{Defect Identification, Equipment Inspection}". In the matching results, the semantic tags of the first three are semantically related to "Power Flow Exceeding Limits Handling" and are included in the initial standardized capability unit set; the image recognition algorithm is excluded because its semantic tag is irrelevant to the task objective.
[0054] S303, based on the data requirement query conditions, perform availability filtering on each standardized capability unit in the initial standardized capability unit set to obtain an intermediate standardized capability unit set.
[0055] In some possible implementations, based on the data requirement query conditions, availability filtering is performed on each standardized capability unit in the initial standardized capability unit set. This includes: comparing the data freshness requirements in the structured capability description file corresponding to each standardized capability unit with the data time range constraints in the data requirement query conditions, and eliminating standardized capability units whose data freshness requirements do not meet the data time range constraints; comparing the performance indicators in the structured capability description file corresponding to each standardized capability unit with the scheduling response latency constraints in the structured task description, and eliminating standardized capability units whose average response latency exceeds the scheduling response latency constraints; comparing the permission levels in the structured capability description file corresponding to each standardized capability unit with the caller permission information in the structured task description, and eliminating standardized capability units whose permission levels are higher than the caller permission information; and determining the remaining standardized capability units after comparing data freshness requirements, performance indicators, and permission levels as the intermediate standardized capability unit set.
[0056] For example, the controlling agent performs availability filtering on the initial standardized set based on data demand query conditions. For instance, the data freshness requirement for power flow calculation program A is "measurement data no more than 5 minutes," which meets the current task requirements; the data freshness requirement for power flow calculation program B is "measurement data no more than 15 minutes," which is acceptable but has poor timeliness; the performance metric of the N-1 safety verification algorithm is "average response latency of 15 seconds," which meets the 30-second scheduling response latency constraint; the performance metric of the adjustment resource assessment service is "average response latency of 40 seconds," exceeding the 30-second constraint, and is therefore removed from the initial set. After filtering, power flow calculation program A, power flow calculation program B, and the N-1 safety verification algorithm are retained and enter the intermediate standardized capability unit set.
[0057] S304. Based on the semantic tags, historical reliability scores, and constraints corresponding to each standardized capability unit in the intermediate standardized capability unit set, determine the professional suitability score of the standardized capability unit.
[0058] In some possible implementations, the professional suitability score of the standardized capability unit is determined based on semantic tags, historical reliability scores, and constraints. This includes: mapping semantic tags to semantic matching vectors, mapping historical reliability scores to reliability feature vectors, and encoding constraints into constraint satisfaction vectors; performing nonlinear transformations on the semantic matching vectors, reliability feature vectors, and constraint satisfaction vectors to obtain semantic latent vectors, reliability latent vectors, and constraint latent vectors; wherein the semantic latent vectors and reliability latent vectors have the same vector dimension; calculating the attention weight between the semantic latent vectors and reliability latent vectors, and using the attention weights to perform weighted fusion of the semantic latent vectors and reliability latent vectors to obtain a fused latent vector; concatenating the fused latent vector with the constraint latent vector to obtain a concatenated vector, and performing a fully connected transformation and activation processing on the concatenated vector to obtain a professional suitability score representing the comprehensive matching degree of the standardized capability unit to the current business task.
[0059] Furthermore, semantic tags are mapped to semantic matching vectors, historical reliability scores are mapped to reliability feature vectors, and constraints are encoded into constraint satisfaction vectors. This includes: calculating the semantic relevance between semantic tags and task target keywords in the structured task description to obtain the matching score between each tag item in the semantic tags and the task target keywords; arranging the matching scores according to a preset tag order to generate a semantic matching vector; obtaining the historical reliability score of the standardized capability unit, which includes at least two of the following: historical call success rate, average response latency, maximum concurrency, and number of historical failures; normalizing each historical reliability score to a preset numerical range and concatenating the normalized historical reliability scores in a preset order to generate a reliability feature vector; obtaining the satisfaction status of each constraint in the constraints for the standardized capability unit; assigning a first encoding value to satisfied constraints, a second encoding value to unsatisfied constraints, and a third encoding value between the first and second encoding values to partially satisfied constraints; and combining the encoding values corresponding to each constraint according to the order of the constraints to generate a constraint satisfaction vector.
[0060] Furthermore, nonlinear transformations are performed on the semantic matching vector, reliability feature vector, and constraint satisfaction vector respectively to obtain semantic latent vector, reliability latent vector, and constraint latent vector. This includes: inputting the semantic matching vector into a first encoding sub-network, where the first encoding sub-network performs at least one fully connected transformation and nonlinear activation processing on the semantic matching vector, outputting the semantic latent vector; inputting the reliability feature vector into a second encoding sub-network, where the second encoding sub-network performs at least one fully connected transformation and nonlinear activation processing on the reliability feature vector, outputting the reliability latent vector; wherein the second encoding sub-network has the same network structure as the first encoding sub-network, so that the semantic latent vector and the reliability latent vector have the same vector dimension; and inputting the constraint satisfaction vector into a third encoding sub-network, where the third encoding sub-network performs at least one fully connected transformation and nonlinear activation processing on the constraint satisfaction vector, outputting the constraint latent vector; wherein the third encoding sub-network has a different network structure than the first encoding sub-network, to preserve the independent semantics between the constraint terms in the constraint satisfaction vector.
[0061] The semantic latent vector serves to provide a feature carrier for the soft indicator "functional matching," allowing it to interact and be calculated with other indicators in the same numerical space. The reliability latent vector provides a comparable and interactive feature representation for the soft indicator "historical reliability," which is in the same numerical space as the functional matching indicator. The constraint latent vector acts as an independent carrier of rigid constraint information. After being concatenated with the fusion latent vector, it is processed uniformly by the prediction sub-network, ensuring that the constraint satisfaction state participates in the final fitness score calculation without being diluted. Through independent nonlinear transformations of the three encoding sub-networks, the three heterogeneous information types—semantics, reliability, and constraints—are mapped to their respective best-fit latent spaces. Semantics and reliability are aligned to the same space for attention interaction, while constraints retain their independent space for rigid participation in prediction after concatenation. This eliminates the differences in the original dimensions and lays the numerical and structural foundation for the subsequent hierarchical calculation architecture of soft indicator fusion and independent rigid constraints.
[0062] Further, the attention weights between the semantic latent vector and the reliability latent vector are calculated, and the attention weights are used to weighted fuse the semantic latent vector and the reliability latent vector to obtain the fused latent vector. This includes: mapping the semantic latent vector to a first query vector through a first query projection matrix, mapping the reliability latent vector to a first key vector through a first key projection matrix, and mapping the reliability latent vector to a first value vector through a first value projection matrix; calculating the first dot product similarity between the first query vector and the first key vector, and performing normalization on the first dot product similarity to obtain the first attention weight; and mapping the reliability latent vector to a first value vector through a second query projection matrix. The query projection matrix is mapped to the second query vector. The semantic latent vector is mapped to the second key vector through the second key projection matrix. The semantic latent vector is also mapped to the second value vector through the second value projection matrix. The second dot product similarity between the second query vector and the second key vector is calculated. The second dot product similarity is normalized to obtain the second attention weight. The first value vector is weighted using the first attention weight to obtain the first weighted result. The second value vector is weighted using the second attention weight to obtain the second weighted result. The first weighted result and the second weighted result are added element-wise to obtain the fused latent vector.
[0063] The first query projection matrix is a weight matrix used to linearly map semantic latent vectors from the semantic latent space to the first query space. This matrix is a learnable parameter matrix in the fitness evaluation model, with the number of rows equal to the vector dimension of the first query space and the number of columns equal to the vector dimension of the semantic latent vectors, capturing the semantic focus pattern on reliability. The second query projection matrix is a weight matrix used to linearly map reliability latent vectors from the reliability latent space to the second query space. This matrix is also a learnable parameter matrix in the fitness evaluation model, with the number of rows equal to the vector dimension of the second query space and the number of columns equal to the vector dimension of the reliability latent vectors, capturing the reliability focus pattern on semantics.
[0064] By introducing independent query projection matrices, key projection matrices, and value projection matrices for bidirectional attention, and by querying the reliability latent vector with the semantic latent vector and querying the semantic latent vector with the reliability latent vector, the differentiated attention weights of the two towards each other are obtained, and the weights are applied to the corresponding value vectors. This allows the fused latent vector to truly reflect the asymmetric relationship between the degree of mutual attention between functional matching and historical reliability.
[0065] Furthermore, a fully connected transformation and activation process are performed on the concatenated vector to obtain a professional fit score representing the comprehensive matching degree of the standardized capability unit to the current business task. This includes: performing at least one fully connected transformation on the concatenated vector to map it into an intermediate prediction vector; wherein, the fully connected transformation performs a weighted interaction on the fusion latent vector part and the constraint latent vector part in the concatenated vector through a weight matrix, so that the constraint information has a numerical correction effect on the fusion information; using a nonlinear activation function with one-sided inhibition characteristics to perform activation processing on the intermediate prediction vector, so that the negative features representing constraint violations in the intermediate prediction vector are truncated to zero or smoothly decayed to obtain an activated vector; performing an output mapping on the activated vector to map the activated vector into a scalar form of fit prediction value, and using the fit prediction value as the professional fit score.
[0066] By using a fully connected transformation, the fused latent vector and the constraint latent vector interact at the numerical level rather than being simply parallel. Then, through nonlinear suppression of activation processing and scalarization of the output mapping, the final professional fit score can simultaneously reflect the degree of functional matching, the level of historical reliability, and the state of constraint satisfaction in a single numerical value. Moreover, when any hard constraint is violated, the activation processing can significantly suppress the score output, thereby ensuring that the score result is highly sensitive to safety constraints.
[0067] In this implementation, the semantic matching vector, reliability feature vector, and constraint satisfaction vector are encoded in parallel. Then, through a hierarchical calculation path of "attention fusion of semantics and reliability, and concatenation prediction of fusion result and constraint vector", the professional suitability score no longer relies on simple weighting of a single dimension. Instead, it can simultaneously consider the interaction between functional matching, historical reliability, and constraint satisfaction within a unified calculation framework, thereby outputting a comprehensive matching score that is more in line with the actual needs of power grid dispatching tasks.
[0068] In other possible implementations, the professional suitability score of the standardized capability unit is determined based on semantic tags, historical reliability scores, and constraints. This includes: calculating the semantic relevance between the semantic tags of the standardized capability unit and the keywords of the task target to obtain the first score component; using the historical reliability score of the standardized capability unit as the second score component; mapping the degree of satisfaction of the standardized capability unit with the constraints to the third score component; and performing a fusion calculation on the first, second, and third score components to obtain the professional suitability score of the standardized capability unit.
[0069] Taking the intermediate standardized capability unit "Power Flow Calculation Program A" in the "Regional Power Flow Exceedance Handling" task as an example: The master control agent calculates the semantic relevance between the semantic label "{Power Flow Calculation, Scheduling Operation}" of Power Flow Calculation Program A and the task objective keyword "Power Flow Exceedance Handling", obtaining the first score component as 0.90; Power Flow Calculation Program A has a 95% success rate in historical scheduling tasks, and its historical reliability score is 0.95, which is used as the second score component; The current task constraint requires "scheduling response delay not exceeding 30 seconds", and the average response delay of Power Flow Calculation Program A is 12 seconds, which fully meets the constraint, and the constraint satisfaction degree is mapped to the third score component 1.0; The above three score components are weighted and fused (with weights of 0.4, 0.3, and 0.3 respectively), and the professional adaptability score of Power Flow Calculation Program A is obtained as 0.90×0.4 + 0.95×0.3 + 1.0×0.3 = 0.945.
[0070] In this implementation, the professional suitability score is decomposed into three independent and interpretable components: semantic relevance, historical reliability, and constraint satisfaction. Then, a fusion calculation is performed. This ensures that the score retains the independent contribution of each dimension to the suitability, making it easy to trace and audit. Furthermore, the fusion calculation unifies the evaluation information of different dimensions and semantics, thus enabling a stable output of a comprehensive score that matches the power grid task requirements while maintaining low computational complexity.
[0071] S305, the standardized capability units in the intermediate standardized capability unit set whose professional suitability scores meet the preset threshold are determined as the candidate capability unit set.
[0072] For example, the preset threshold is set to 0.80. Power flow calculation program A (0.94) and N-1 safety verification algorithm (0.90) exceed the preset threshold and are identified as candidate standardized capability units; power flow calculation program B (0.74) is excluded from the candidate set because it is below the preset threshold. The final candidate standardized capability unit set is: power flow calculation program A (primary choice) and N-1 safety verification algorithm.
[0073] In some possible implementations, determining the candidate standardized capability unit set from the algorithm capability catalog based on the structured task description also includes the following steps: when the professional suitability score of each standardized capability unit in the intermediate standardized capability unit set is lower than a preset threshold, semantic expansion is performed on the functional semantic query conditions to obtain extended query conditions; based on the extended query conditions, semantic matching and availability filtering are re-performed on each standardized capability unit in the algorithm capability catalog to obtain a supplementary standardized capability unit set; after merging the supplementary standardized capability unit set with the intermediate standardized capability unit set, professional suitability scoring and threshold filtering are re-performed to obtain a candidate standardized capability unit set.
[0074] Furthermore, semantic expansion is performed on the functional semantic query conditions to obtain extended query conditions, including: extracting power grid synonyms or superordinate terms that are semantically related to the task target keywords from the structured task description; supplementing the power grid synonyms or superordinate terms into the functional semantic query conditions to obtain extended query conditions; wherein, the power grid synonyms or superordinate terms are determined based on the term mapping relationship pre-established in the unified power semantic ontology; the term mapping relationship stores power dispatch operation intent words, power grid equipment object words and corresponding synonym expressions and superordinate concept words. By supplementing the functional semantic query conditions with synonyms or superordinate terms of the task target keywords based on the pre-set term mapping relationship in the unified power semantic ontology, extended queries are performed, so that even when the wording of the original keywords and the semantic tags in the capability catalog is inconsistent, standardized capability units that match the function can still be retrieved through semantic expansion. This effectively improves the tolerance and recall rate of the inconsistency of terminology in heterogeneous power grid systems during the capability discovery stage, and avoids the candidate set being empty or the available standardized capability units being missed due to the failure of matching a single keyword.
[0075] When the initial semantic matching result fails, the system expands the semantic query conditions and initiates a second matching and filtering process instead of directly returning failure or transferring to manual intervention. This allows the system to automatically expand the search scope and rediscover available capability units when there are inconsistencies in terminology or insufficient semantic coverage in the initial query. This significantly improves the recall capability and robustness of the capability discovery phase and ensures that subsequent orchestration and scheduling will not be interrupted due to an empty candidate set.
[0076] It is understandable that, when the professional suitability scores of all standardized capability units in the intermediate standardized capability unit set are lower than the preset threshold, in addition to performing semantic expansion and initiating secondary matching, other processes can also be executed according to the needs of the actual application scenario. For example, the preset threshold can be lowered to relax the screening conditions, the standardized capability unit with the highest historical reliability score can be selected from the algorithm capability catalog as a fallback candidate, the master control agent can be notified to send a manual takeover request to the collaborative cockpit, or the process of determining the current candidate standardized capability unit set can be directly terminated and the capability discovery failure status can be returned. This invention does not impose any limitations on this.
[0077] By parsing business task requests into structured task descriptions that include power dispatching operation intentions, data requirements, and power grid operation safety constraints, and through a multi-level screening process of "semantic matching—availability filtering—professional suitability scoring—threshold screening," a set of candidate standardized capability units is automatically determined from the algorithm capability catalog. This allows large model agents to handle complex power grid tasks without manually specifying the calling objects. Based on the task objectives and constraints, units that do not meet the standards for data freshness and performance are further eliminated based on semantic matching. The available units are then quantitatively ranked by comprehensively considering semantic relevance, historical reliability, and constraint satisfaction. This ensures that the candidate capability units are functionally matched while also taking into account their historical execution reliability and compliance with the current task's safety constraints. This provides a high-quality and reliable set of candidate standardized capability units for subsequent dynamic orchestration, significantly improving the accuracy and robustness of task execution.
[0078] Optionally, in step S40, the arrangement constraint information includes task timing relationships, preconditions, postconditions, and power grid operation safety constraints. Specifically: task timing relationships refer to the chronological order that must be followed between the corresponding business actions of each standardized capability unit, as specified in the structured task description. For example, in the "Regional Power Flow Exceedance Handling" task, the task timing relationship stipulates that "real-time measurement data must be acquired" first, then "power flow calculation" is performed, and finally "operation ticket is generated." Preconditions refer to the prerequisite states or data conditions that must be met before a standardized capability unit is invoked. For example, the preconditions for the "Multi-Objective Optimization Scheme Generation" node are "power flow calculation results have been acquired" and "power flow calculation results have been obtained." "Obtain N-1 safety verification results" and "Obtain adjustment resource assessment results" are all indispensable; otherwise, the node cannot be started. Postconditions refer to the output data items or status changes that should be generated after the standardized capability unit is successfully called and returns results. For example, the postconditions of the "N-1 safety verification algorithm" node are "output N-1 over-limit device list", "output key section information", and "output verification conclusion". Grid operation safety constraints refer to restrictive rules that are directly related to the stable operation of the power system and the safety of equipment operation. For example, in the "regional power flow over-limit handling" task, grid operation safety constraints include "all control commands must be verified by N-1 and simulated before they can be issued".
[0079] Combination Figure 3 As shown, based on the candidate capability unit set and the orchestration constraint information in the structured task description, an algorithm execution plan including sequential, parallel, or conditional branch structures is generated, including the following steps: S401, based on the task sequence relationship, maps each standardized capability unit in the candidate capability unit set to a node in the algorithm execution plan, and determines the initial dependency relationship between the nodes.
[0080] S402, based on the preconditions and postconditions, determine the data dependencies between nodes, mark nodes that have data dependencies as sequential execution, and mark nodes that do not have data dependencies as branches for parallel execution.
[0081] S403, based on power grid operation safety constraints, inserts a safety verification node before nodes where high-risk equipment operations exist, and configures branch triggering conditions for nodes with conditional branches.
[0082] S404 assembles initial dependencies, sequence and / or parallel tags, security check nodes, and branch triggering conditions into a structured algorithm execution plan.
[0083] Here, the algorithm execution plan is a structured execution scheme dynamically generated by the master control agent based on the candidate standardized capability unit set and the orchestration constraint information in the structured task description, which is used to guide the collaborative scheduling of heterogeneous algorithms and business services.
[0084] In some possible implementations, an algorithm execution plan, including sequential, parallel, or conditional branch structures, is generated based on the candidate capability unit set and the orchestration constraint information in the structured task description. This also includes: for each node in the algorithm execution plan, selecting standardized capability units that meet the functional requirements from the candidate capability unit set, obtaining a candidate unit subset for the node; when the candidate unit subset includes multiple standardized capability units capable of performing the same function, obtaining the data quality assessment value and computational resource utilization rate of the data source corresponding to each standardized capability unit; prioritizing each standardized capability unit in the candidate unit subset based on the data quality assessment value and computational resource utilization rate, determining the highest-priority standardized capability unit as the node's primary standardized capability unit, and writing the remaining standardized capability units in the candidate unit subset (excluding the highest-priority standardized capability unit) as backup standardized capability units into the node's exception handling strategy; wherein, the exception handling strategy is used to sequentially select backup standardized capability units to initiate retry in the event of a failure in calling the primary standardized capability unit.
[0085] Among them, the data quality assessment value refers to the quantitative indicator obtained after assessing the availability and accuracy of the data source on which the standardized capability unit depends at the current moment; the computing resource utilization rate refers to the proportion of computing resources used by the data source or operating environment on which the standardized capability unit depends at the current moment.
[0086] Furthermore, based on data quality assessment values and computing resource utilization rates, the standardized capability units in the candidate unit subset are prioritized and ranked. This includes: performing a first ranking of the standardized capability units in the candidate unit subset according to their data quality assessment values from high to low; when two or more standardized capability units have the same data quality assessment value, or when the difference between their data quality assessment values is less than a preset difference threshold, performing a second ranking of the two or more standardized capability units according to their computing resource utilization rates from low to high; and determining the order obtained after the first and second rankings as the priority ranking result of the standardized capability units in the candidate unit subset. By performing a two-level ranking first according to data quality assessment values from high to low, and then according to computing resource utilization rates from low to high, the primary standardized capability unit not only possesses the optimal data availability at the current moment, but also takes into account the light load of the operating environment, thereby reducing the response latency and call failure risk caused by resource contention while ensuring the reliability of the call results.
[0087] By prioritizing each node in the algorithm execution plan from the candidate unit subset according to data quality assessment value and computing resource utilization, and setting the ranking result as the primary and backup standardized capability units, the execution plan pre-embeds the call priority and failure retry path based on the real-time data source status for each node during the generation stage. This allows for a quick switch to the backup standardized capability unit without rediscovery and orchestration when the primary standardized capability unit fails to be called, significantly improving the robustness of scheduling execution and the efficiency of anomaly recovery.
[0088] Combination Figure 5 As shown, the algorithm execution plan consists of multiple nodes (nodes 1-9), each corresponding to the invocation of a standardized capability unit. Nodes are associated through initial dependencies, data dependencies, sequence markers, parallel markers, or conditional branch triggering conditions. Specifically, the algorithm execution plan includes at least one of the following: sequential execution structure, parallel execution structure, and conditional branch structure. Sequential execution structures (e.g., node 1→2, node 5→6) represent nodes that must execute sequentially due to data dependencies or business timing requirements; parallel execution structures (e.g., nodes 2, 3, 4) represent nodes that do not have data dependencies and can initiate calls simultaneously to shorten the total task execution time; conditional branch structures (e.g., node 6→6a, node 6→6b) represent multiple mutually exclusive execution paths between nodes, with the specific path selection determined by the execution result status of the previous node or key variables in the task context. Furthermore, a security verification node (node 7) is inserted before the node (node 8) where high-risk equipment operation occurs, and a primary and backup standardized capability unit are assigned to each node, enabling the algorithm execution plan to have anomaly degradation and rapid recovery capabilities during scheduling.
[0089] Continue to combine Figure 5As shown, taking the "Regional Power Flow Exceeding Limit Handling" task as an example, the algorithm execution plan is explained. Node 1 is the real-time measurement and cross-sectional data acquisition node. It is a data acquisition node (business service call), which obtains real-time active power, reactive power, voltage amplitude, switch position, and current cross-sectional data of the exceeding-limit lines and associated substations from the dispatch automation system (SCADA / EMS). This provides a data foundation for subsequent power flow calculation and safety verification. The primary standardized capability unit is SCADA real-time measurement query service, and the backup standardized capability unit is EMS cross-sectional data query service. Node 2 is the power flow calculation node, which is a calculation node (heterogeneous algorithm call). Based on the real-time cross-sectional data acquired by Node 1, it executes power flow calculation... The calculation confirms the power flow distribution, degree of exceeding limits, and voltage levels of associated substations for lines exceeding limits. The primary standardized capability unit is power flow calculation program A, and the backup standardized capability unit is power flow calculation program B. Node 3 is the N-1 safety verification node, which is a calculation-type node (heterogeneous algorithm call). Based on the current section, it simulates the power flow distribution after a single component (line, main transformer, busbar) is taken out of operation, identifies whether there are other components exceeding limits or voltage anomalies, and outputs a list of critical vulnerable components. The primary standardized capability unit is N-1 safety verification algorithm A, and the backup standardized capability unit is N-1 simplified verification algorithm B. Node 4 is for regulating resource assessment. Node 1 is a computational node (heterogeneous algorithm call) that evaluates the adjustable generator output adjustment, transferable load, reactive power compensation device adjustment capability, and response delay of various resources under the current cross-section, providing adjustable resource boundaries for subsequent scheme generation. The primary standardized capability unit is the adjustable resource assessment service A, and the backup standardized capability unit is the rule assessment fallback service. Node 5 is a multi-objective optimization scheme generation node, which is a computational node (heterogeneous algorithm call). It integrates power flow calculation results, N-1 verification results, and adjustable resource list to generate multiple candidate disposals with the objectives of eliminating over-limits, minimizing operating costs, and minimizing load losses. The primary standardized capability unit is multi-objective optimization algorithm A, and the backup standardized capability unit is rule expert scheme generation algorithm B. Node 6 is a power flow calculation node, which is a conditional branch node (business service call). It performs safety and effect simulations on the candidate disposal schemes generated by node 5. When the number of candidate schemes is greater than 1, node 6 expands into two parallel sub-branches, node 6a and node 6b, and calls the simulation simulation system for different schemes respectively. When the number of candidate schemes is 1, node 6 directly performs the simulation on a single scheme. The primary standardized capability unit is simulation simulation system A, and the backup standardized capability unit is offline simulation deduction tool B.Node 7 is the power flow calculation node, which is also a safety verification node (insertion node). Based on power grid operation safety constraints, it performs rule verification and regulatory agent verification on candidate schemes that have passed the pre-simulation, checking for safety risks such as opening or closing disconnect switches under load, violating protection coordination principles, or causing large-scale power outages for users. The primary standardized capability unit is the safety rule verification service, and the backup standardized capability unit is the regulatory agent review process. Node 8 is the operation ticket generation node, which is a process processing node (business service call). For the handling schemes that have passed the safety verification, it calls the dispatch operation ticket generation system to automatically generate operation tickets containing the operation steps. The scheduling operation ticket, including steps, operation objects, safety measures, and precautions, is primarily standardized by the scheduling operation ticket generation system A, with backup standardized by the scheduling operation ticket generation system B. Node 9 is the manual approval node (for high-risk write operations). It submits the operation ticket generated by node 8 to the authorized dispatcher for manual approval. Upon approval, the unified scheduling runtime engine executes and issues the operation ticket. If approval fails, it reverts to node 5 to regenerate the plan or transfers it to manual takeover. The primary standardized by the system is the operation approval process service, and the backup standardized by the system is the collaborative cockpit manual takeover process.
[0090] Based on the task sequence relationships, preconditions, postconditions, and power grid operation safety constraints in the structured task description, candidate standardized capability units are automatically mapped into algorithm execution plans containing sequential, parallel, or conditional branch structures. This allows the master control agent to flexibly organize the collaborative invocation of multiple algorithms and multiple service functions according to the actual needs of the task and the data dependencies between nodes. At the same time, by forcibly inserting safety verification nodes before high-risk equipment operation nodes and configuring explicit trigger conditions for conditional branch nodes, power grid operation safety constraints are physically transformed into unavoidable execution links at the planning level. This not only improves the parallelism and execution efficiency of multi-capability unit collaborative scheduling but also ensures the safety, determinism, and controllability of algorithm execution plans when involving high-risk operations and complex logical branches.
[0091] Optionally, combined Figure 4 As shown, the heterogeneous algorithm scheduling method for power systems includes the following steps: S10: Create capability profiles for multiple heterogeneous capability objects in the power system and generate a structured capability description file for each heterogeneous capability object.
[0092] S20, based on the structured capability description file, encapsulates heterogeneous capability objects into standardized capability units with a unified standardized interface, and registers the standardized capability units to the algorithm capability directory.
[0093] S30: Parse the received business task request into a structured task description, and determine a set of candidate standardized capability units from the algorithm capability catalog based on the structured task description.
[0094] S40 generates an algorithm execution plan, including sequential, parallel, or conditional branch structures, based on the candidate standardized capability unit set and the orchestration constraint information in the structured task description.
[0095] S50 reads the caller identity information corresponding to the business task request, and reads the permission level and security constraints from the structured capability description file corresponding to the current standardized capability unit, and performs identity authentication and least privilege verification.
[0096] In some possible implementations, performing identity authentication and least privilege verification includes: reading the caller identity information corresponding to the business task request from the task context, where the caller identity information includes the caller identifier and the caller role; reading the permission level and security constraints from the structured capability description file corresponding to the current standardized capability unit; comparing the caller identifier with a preset registry of legitimate callers to confirm that the caller corresponding to the caller identifier is in a valid authorized state; comparing the permission level corresponding to the caller role with the permission level of the current standardized capability unit; if the caller identifier is not confirmed to be in a valid authorized state, or if the permission level corresponding to the caller role is lower than the permission level of the current standardized capability unit, the identity authentication and least privilege verification are deemed to have failed; if the caller identifier is confirmed to be in a valid authorized state, and the permission level corresponding to the caller role is lower than the permission level of the current standardized capability unit, the identity authentication and least privilege verification are deemed to have passed.
[0097] By combining the valid authorization status confirmation of the caller's identifier with the hierarchical comparison of the caller's role and permission levels, the two basic security prerequisites of "whether the caller is legitimate" and "whether the caller has exceeded its authority" can be verified before each standardized capability unit call. This prevents unauthorized or unauthorized access at the call entry point and ensures the security boundary of the grid heterogeneous capability call.
[0098] S60 triggers parameter reassembly based on the error type if the verification fails.
[0099] In some possible implementations, if the verification still fails after parameter reassembly, an alternative solution is read from the node exception handling strategy corresponding to the current standardized capability unit, and an error status is returned to the master control agent.
[0100] S70, after the verification is passed, perform a parameter validity check on the input parameter to be passed according to the input specification corresponding to the current standardized capability unit.
[0101] In some possible implementations, parameter validity checks are performed on the parameters to be input based on the input specifications corresponding to the current standardized capability unit. This includes: reading the input specifications from the structured capability description file corresponding to the current standardized capability unit, which includes the input parameter name, data type, value range, and mandatory requirement; obtaining the set of parameters to be input to the current standardized capability unit from the task context; checking each parameter in the parameter set to see if it meets the mandatory requirement, data type requirement, and value range requirement of the corresponding parameter in the input specifications; if there are missing mandatory parameters, mismatched parameter data types, or parameter values exceeding the preset range, the parameter validity check is deemed to have failed; otherwise, the parameter validity check is deemed to have passed.
[0102] By automatically comparing the parameters to be passed with the predefined input specifications in the structured capability description file, problems such as missing required parameters, type mismatch, and out-of-bounds values can be intercepted before the call is initiated, thus preventing invalid calls or erroneous data from entering heterogeneous algorithms and business services.
[0103] S80, if the check fails, blocks the invocation of the current standardized capability unit.
[0104] In some possible implementations, if the check fails, after blocking the call to the current standardized capability unit, an alternative standardized capability unit call is triggered according to the node exception handling strategy corresponding to the current standardized capability unit.
[0105] S90, after the check is passed, initiate a call to the current standardized capability unit, and monitor the response latency and execution status of the current standardized capability unit during the call process to trigger the node exception handling strategy corresponding to the current standardized capability unit.
[0106] In some possible implementations, the response latency and execution status of the current standardized capability unit are monitored during the invocation process to trigger the node exception handling strategy corresponding to the current standardized capability unit. This includes: recording the invocation timestamp when initiating a call to the current standardized capability unit; obtaining the real-time execution status of the current standardized capability unit according to a preset monitoring period while waiting for a return, including waiting, executing, returning successfully, or returning failed; calculating the real-time response latency of the current standardized capability unit based on the time difference between the invocation timestamp and the current time; comparing the real-time response latency with the latency threshold in the performance metrics corresponding to the current standardized capability unit; and triggering the node exception handling strategy corresponding to the current standardized capability unit if the real-time response latency exceeds the latency threshold or the execution status is returned as failed.
[0107] By continuously monitoring response latency and execution status at preset intervals during the invocation process and comparing them in real time with the latency threshold in the performance indicators of the standardized capability unit, invocation timeouts or execution failures can be detected immediately and trigger the corresponding node exception handling strategy, thereby preventing a single node from blocking the execution plan for a long time and ensuring the timeliness and controllability of the overall scheduling process.
[0108] S100, after the call returns, performs integrity verification, consistency verification, and semantic verification on the returned data in sequence.
[0109] In some possible implementations, the returned data undergoes sequential integrity verification, consistency verification, and semantic verification, including: reading the output specification from the structured capability description file corresponding to the current standardized capability unit, the output specification including the output field names, data types, and mandatory requirements; checking whether the returned data contains all mandatory output fields according to the output specification, and whether the data types of each output field conform to the output specification, thus completing the integrity verification; after the integrity verification passes, cross-referencing the parameters of the same device or the same measurement in the returned data with the returned results of other standardized capability units stored in the task context, checking whether the parameters of the same device or the same measurement meet the preset deviation threshold in value, thus completing the consistency verification; after the consistency verification passes, checking whether the values of each field in the returned data are within the reasonable semantic range corresponding to that field according to the semantic rules of the power industry, thus completing the semantic verification.
[0110] By performing three progressive checks in sequence, it can be ensured that only fully verified results can be updated to the task context and drive the scheduling of subsequent nodes, effectively preventing erroneous data from spreading in the task chain.
[0111] S110, after all three checks pass, returns a success status to the master control agent in the large model agent platform, so that the master control agent can schedule the next node.
[0112] S120: If any of the integrity check, consistency check, or semantic check fails, read the handling action corresponding to the check failure type from the node exception handling strategy corresponding to the current standardized capability unit, and execute a retry, switch to a backup standardized capability unit, or rollback process.
[0113] During the scheduling and execution process, end-to-end control ensures that the master control agent's calls to each standardized capability unit are within the security boundaries of trusted identity, matching permissions, compliant parameters, controllable execution, and verifiable results. Simultaneously, through linkage with node anomaly handling strategies, corresponding blocking, reassembly, retry, or degradation measures can be triggered in a timely manner in case of anomalies such as failed identity verification, invalid parameters, response timeout, or failed data verification, preventing the accumulation and propagation of anomalies in the task chain. After a successful call, the verified result drives the master control agent to determine the next node to be scheduled, enabling the large model agent platform to achieve a secure, reliable, and traceable hierarchical scheduling closed loop in complex power grid business tasks.
[0114] Combination Figure 6 As shown in the figure, this embodiment of the invention provides a heterogeneous algorithm scheduling device 60 for a power system, including a capability profiling module 601, a heterogeneous capability object encapsulation module 602, a task parsing module 603, and an algorithm execution plan generation module 604. The capability profiling module 601 is configured to perform capability profiling on multiple heterogeneous capability objects in the power system, generating a structured capability description file for each heterogeneous capability object. The heterogeneous capability objects include heterogeneous algorithms and heterogeneous business services used for power data acquisition, power equipment operation, or grid dispatching business process processing. The heterogeneous capability object encapsulation module 602 is connected to the capability profiling module 601 and is configured to encapsulate the heterogeneous capability objects into standardized capabilities with a unified standardized interface based on the structured capability description file. The system is divided into three modules: a standardized capability unit and a standardized capability unit registration module; a task parsing module 603 connected to a heterogeneous capability object encapsulation module 602, configured to parse the received business task request into a structured task description, and determine a set of candidate standardized capability units from the algorithm capability directory based on the structured task description; and an algorithm execution plan generation module 604 connected to the task parsing module 603, configured to generate an algorithm execution plan including sequential, parallel, or conditional branch structures based on the set of candidate standardized capability units and the orchestration constraint information in the structured task description, so that the master agent in the large model intelligent agent platform can call each standardized capability unit according to the algorithm execution plan; wherein, each node in the algorithm execution plan corresponds to the call of a standardized capability unit.
[0115] Optionally, continue to combine Figure 6As shown, the heterogeneous algorithm scheduling device 60 for power systems also includes a verification module 605. The verification module 605 is configured to read the caller identity information corresponding to the business task request before scheduling each standardized capability unit in sequence according to the algorithm execution plan, and read the permission level and security constraints from the structured capability description file corresponding to the current standardized capability unit, and perform identity authentication and least privilege verification. After the verification is passed, the module performs parameter legality checks on the input parameters to be passed according to the input specifications corresponding to the current standardized capability unit. After the check is passed, the module initiates a call to the current standardized capability unit, and monitors the response latency and execution status of the current standardized capability unit during the call process to trigger the node anomaly handling strategy corresponding to the current standardized capability unit. After the call returns, the module performs integrity verification, consistency verification and semantic verification on the returned data in sequence, and returns a success status to the master control agent after all three verifications are passed, so that the master control agent can schedule the next node.
[0116] Combination Figure 7 As shown, this embodiment of the invention provides a heterogeneous algorithm scheduling device 70 for power systems, including a processor 701 and a memory 702, and may also include a communication interface 703 and a bus 704. The processor 701, communication interface 703, and memory 702 can communicate with each other via the bus 704. The communication interface 703 can be used for information transmission. The processor 701 can call logical instructions in the memory 702 to execute the heterogeneous algorithm scheduling method for power systems described in the above embodiment.
[0117] Furthermore, the logic instructions in the aforementioned memory 702 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.
[0118] The memory 702, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of the present invention. The processor 701 executes functional applications and data processing by running the program instructions / modules stored in the memory 702, thereby implementing the heterogeneous algorithm scheduling method for power systems in the above method embodiments.
[0119] The memory 702 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 702 may include high-speed random access memory and may also include non-volatile memory.
[0120] Processor 701 is a circuit with signal processing capabilities. In one implementation, processor 701 can be a circuit with instruction read and execute capabilities, such as a central processing unit (CPU), microprocessor, graphics processing unit (GPU) (which can be understood as a type of microprocessor), or digital signal processor (DSP). In another implementation, processor 701 can implement certain functions through the logical relationships of hardware circuits. These logical relationships of hardware circuits are fixed or reconfigurable. For example, processor 701 can be a hardware circuit implemented as an ASIC or a programmable logic device (PLD), such as an FPGA. In a reconfigurable hardware circuit, the process of the processor loading a configuration document and configuring the hardware circuit can be understood as the process of the processor loading instructions to implement the functions of some or all of the above modules. Furthermore, it can also be a hardware circuit designed for artificial intelligence, which can be understood as a type of ASIC, such as a neural network processing unit (NPU), tensor processing unit (TPU), or deep learning processing unit (DPU). The processor 701 is used to execute related programs to implement the functions required by the units in the heterogeneous algorithm scheduling device for power systems according to the embodiments of the present invention, or to execute the heterogeneous algorithm scheduling method for power systems according to the method embodiments of the present invention.
[0121] As can be seen, each module in the above device can be one or more processors (or processing circuits) configured to implement the above methods, such as: CPU, GPU, NPU, TPU, DPU, microprocessor, DSP, ASIC, FPGA, or a combination of at least two of these processor types.
[0122] Furthermore, the modules in the above devices can be integrated in whole or in part, or they can be implemented independently. In one implementation, these modules are integrated together as a system-on-a-chip (SOC). The SOC may include at least one processor for implementing any of the above methods or for implementing the functions of the modules of the device. The at least one processor may be of different types, such as CPU and FPGA, CPU and artificial intelligence processor, CPU and GPU, etc.
[0123] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.
Claims
1. A heterogeneous algorithm scheduling method for power systems, applied to a large-scale intelligent agent platform, characterized in that, include: Capability profiles are created for multiple heterogeneous capability objects in the power system, and a structured capability description file is generated for each heterogeneous capability object. The heterogeneous capabilities include heterogeneous algorithms and heterogeneous business services used for power data acquisition, power equipment operation, or grid dispatch business process processing. Based on the structured capability description file, heterogeneous capability objects are encapsulated into standardized capability units with a unified standardized interface, and the standardized capability units are registered to the algorithm capability directory. The received business task request is parsed into a structured task description, and based on the structured task description, a set of candidate standardized capability units is determined from the algorithm capability catalog. Based on the candidate standardized capability unit set and the orchestration constraint information in the structured task description, an algorithm execution plan including sequential, parallel, or conditional branch structures is generated so that the master agent in the large model agent platform can call each standardized capability unit according to the algorithm execution plan; wherein, each node in the algorithm execution plan corresponds to the call of a standardized capability unit.
2. The heterogeneous algorithm scheduling method according to claim 1, characterized in that, Capability profiles are created for multiple heterogeneous capability objects in the power system, generating a structured capability description file for each heterogeneous capability object, including: A structured format is used to create an initial capability description file for each heterogeneous capability object; the initial capability description file includes the functional description, input and output specifications, and data dependencies of the heterogeneous capability object; Obtain and, based on a preset set of power grid business tags, determine the functional types and applicable power grid business scenarios for the initial capability description file, write the semantic tags into the initial capability description file, and generate an intermediate capability description file. Write the interface protocol type, performance indicators, permission level and security constraints of the heterogeneous capability objects into the intermediate capability description file to generate the final structured capability description file.
3. The heterogeneous algorithm scheduling method according to claim 2, characterized in that, Based on structured capability description files, heterogeneous capability objects are encapsulated into standardized capability units with unified standardized interfaces. This includes employing a three-layer encapsulation structure (bottom layer, middle layer, and top layer) to perform layer-by-layer encapsulation of heterogeneous capability objects; whereby... At the underlying level, based on the interface protocol type in the structured capability description file, a corresponding adapter is matched for the heterogeneous capability object to convert the raw response data output by the heterogeneous capability object in the native interface protocol into a unified first intermediate data format. In the middle layer, for the core objects in the first intermediate data format, field mapping is performed according to the unified structure expression, and professional attribute normalization is performed according to the general model of the power industry to generate a second intermediate data format with unified semantics; At the upper layer, based on the second intermediate data format and the functional descriptions, input / output specifications, and security constraints in the structured capability description file, standardized capability units for large-scale intelligent agent platforms are generated.
4. The heterogeneous algorithm scheduling method according to claim 1, characterized in that, The structured task description includes: task target keywords characterizing the intent of power dispatch operations or grid equipment objects; data requirement fields characterizing the measurement data needed to perform power dispatch operations; and constraints characterizing the grid operation safety rules and dispatch response time requirements corresponding to the power dispatch operations. Based on the structured task description, a set of candidate standardized capability units is determined from the algorithm capability catalog, including: The task target keywords and data requirement fields are assembled into functional semantic query conditions and data requirement query conditions according to the preset query semantic template; Perform semantic matching between the functional semantic query conditions and the semantic tags corresponding to each standardized capability unit in the algorithm capability catalog to obtain an initial set of standardized capability units; Based on the data requirement query conditions, availability filtering is performed on each standardized capability unit in the initial standardized capability unit set to obtain an intermediate standardized capability unit set. Based on the semantic tags, historical reliability scores, and constraints corresponding to each standardized capability unit in the intermediate standardized capability unit set, the professional suitability score of the standardized capability unit is determined. Standardized capability units whose professional suitability scores meet the preset threshold in the intermediate standardized capability unit set are identified as candidate capability unit sets.
5. The heterogeneous algorithm scheduling method according to claim 4, characterized in that, Based on semantic tags, historical reliability scores, and constraints, the professional fit score of the standardized capability unit is determined, including: Semantic labels are mapped to semantic matching vectors, historical reliability scores are mapped to reliability feature vectors, and constraints are encoded into constraint satisfaction vectors. Nonlinear transformations are performed on the semantic matching vector, reliability feature vector, and constraint satisfaction vector respectively to obtain the semantic latent vector, reliability latent vector, and constraint latent vector; among them, the semantic latent vector and the reliability latent vector have the same vector dimension. Calculate the attention weight between the semantic latent vector and the reliability latent vector, and use the attention weight to perform weighted fusion of the semantic latent vector and the reliability latent vector to obtain the fused latent vector; The fused latent vector and the constrained latent vector are concatenated to obtain a concatenated vector. A fully connected transformation and activation process are then performed on the concatenated vector to obtain a professional suitability score that represents the comprehensive matching degree of the standardized capability unit to the current business task.
6. The heterogeneous algorithm scheduling method according to claim 4, characterized in that, Based on the structured task description, a set of candidate capability units is determined from the algorithm capability catalog, which also includes: When the professional suitability scores of each standardized capability unit in the intermediate standardized capability unit set are all lower than the preset threshold, semantic expansion is performed on the functional semantic query conditions to obtain expanded query conditions. Based on the extended query conditions, semantic matching and availability filtering are re-executed on each standardized capability unit in the algorithm capability catalog to obtain a supplementary set of standardized capability units. After merging the supplementary standardized capability unit set with the intermediate standardized capability unit set, the professional suitability scoring and threshold screening are re-executed to obtain the candidate standardized capability unit set.
7. The heterogeneous algorithm scheduling method according to claim 1, characterized in that, The scheduling constraint information includes task timing relationships, preconditions, postconditions, and power grid operation safety constraints; Based on the candidate capability unit set and the orchestration constraints in the structured task description, an algorithm execution plan is generated, including sequential, parallel, or conditional branch structures, including: Based on the task sequence, each standardized capability unit in the candidate capability unit set is mapped to a node in the algorithm execution plan, and the initial dependencies between nodes are determined. Based on the preconditions and postconditions, determine the data dependencies between nodes, mark nodes that have data dependencies as sequential execution branches, and mark nodes that do not have data dependencies as parallel execution branches. Based on power grid operation safety constraints, a safety verification node is inserted before the node where high-risk equipment operation exists, and branch triggering conditions are configured for the node where conditional branches exist. The initial dependencies, sequential and / or parallel flags, security check nodes, and branch triggering conditions are assembled into a structured algorithm execution plan.
8. The heterogeneous algorithm scheduling method according to claim 7, characterized in that, Based on the candidate capability unit set and the orchestration constraint information in the structured task description, an algorithm execution plan is generated, including sequential, parallel, or conditional branch structures. This also includes: For each node in the algorithm execution plan, based on the functional requirements corresponding to the node, standardized capability units that meet the functional requirements are selected from the candidate capability unit set to obtain a subset of candidate units for the node. When the candidate unit subset includes multiple standardized capability units that can perform the same function, obtain the data quality assessment value and computing resource utilization rate of the data source corresponding to each standardized capability unit. Based on the data quality assessment value and computing resource utilization rate, the standardized capability units in the candidate unit subset are prioritized, the standardized capability unit with the highest priority is determined as the primary standardized capability unit of the node, and the standardized capability units in the candidate unit subset other than the highest priority standardized capability unit are written into the node's exception handling strategy as backup standardized capability units. The exception handling strategy is used to select backup standardized capability units to initiate retry in the event that the primary standardized capability unit call fails.
9. The heterogeneous algorithm scheduling method according to any one of claims 1 to 8, characterized in that, Also includes: Before scheduling each standardized capability unit in sequence according to the algorithm execution plan, read the caller identity information corresponding to the business task request, and read the permission level and security constraints from the structured capability description file corresponding to the current standardized capability unit, and perform identity authentication and least privilege verification. After the verification is passed, the parameter validity check is performed on the input parameter to be input according to the input specification corresponding to the current standardized capability unit; After the check is passed, a call to the current standardized capability unit is initiated, and the response latency and execution status of the current standardized capability unit are monitored during the call process to trigger the node exception handling strategy corresponding to the current standardized capability unit. After the call returns, integrity verification, consistency verification, and semantic verification are performed on the returned data in sequence. After all three verifications pass, a success status is returned to the master control agent so that the master control agent can schedule the next node.
10. A heterogeneous algorithm scheduling device for power systems, applied to a large-scale intelligent agent platform, characterized in that, include: The capability profiling module is configured to profile multiple heterogeneous capability objects in the power system and generate a structured capability description file for each heterogeneous capability object. The heterogeneous capabilities include heterogeneous algorithms and heterogeneous business services used for power data acquisition, power equipment operation, or grid dispatch business process processing. The heterogeneous capability object encapsulation module, connected to the capability profiling module, is configured to encapsulate heterogeneous capability objects into standardized capability units with a unified standardized interface based on structured capability description files, and register the standardized capability units to the algorithm capability directory. The task parsing module, connected to the heterogeneous capability object encapsulation module, is configured to parse the received business task request into a structured task description, and determine a set of candidate standardized capability units from the algorithm capability catalog based on the structured task description. The algorithm execution plan generation module, connected to the task parsing module, is configured to generate an algorithm execution plan, including sequential, parallel, or conditional branch structures, based on the candidate standardized capability unit set and the orchestration constraint information in the structured task description. This enables the master agent in the large model agent platform to call each standardized capability unit according to the algorithm execution plan. Each node in the algorithm execution plan corresponds to the call of a standardized capability unit.
Citation Information
Patent Citations
Universal model reasoning service framework supporting multi-type heterogeneous algorithm packaging
CN121599108A
