Task flow control agent for electric power intelligent monitoring and control method

Abstract

The invention belongs to the field of electric power system regulation and control operation, and provides a task flow control agent for electric power intelligent monitoring and a control method. The intelligent agent comprises a man-machine front-end module, a man-machine interaction compiling module, a scheduling auxiliary strategy library, a scheduling operation constraint target library, a task flow library, a task flow management core, an operation memory module and an operation result display module. A dispatcher natural language instruction is converted into an executable task flow through a large model, automatic construction, intelligent execution and traceable management of a complex power dispatching task are achieved through computing power perception, state tracking and closed-loop memory, the dispatching efficiency and reliability are remarkably improved, and unmanned dispatching of a novel power system is supported.

Description

Technical Field

[0001] This invention belongs to the field of power system regulation and operation, and in particular to a task flow control intelligent agent and control method for intelligent power monitoring. Background Technology

[0002] With the increasing number of new controllable sources and loads in the power system, such as distributed photovoltaics, energy storage, and charging piles, especially the grid connection of massive distributed photovoltaics, the traditional power system has evolved into a "new type of power system" with multi-source random interaction. This has led to prominent grid safety and power quality issues such as grid overload and overvoltage. Furthermore, the rapid increase in distribution network elements has resulted in a rapid increase in the amount of distribution network dispatching and decision-making. Traditional manual dispatching methods, when faced with the new distribution network that includes massive elements, exhibit low monitoring efficiency, slow judgment speed, and crude dispatching strategies, affecting fault response and grid safety.

[0003] New power sources such as distributed photovoltaics, energy storage, and charging piles generally have characteristics such as small capacity, wide distribution, large number, strong randomness of operation, and electronic access to power. Their large-scale access to the distribution network with relatively weak sensing and control capabilities significantly increases the complexity of system operation, becoming one of the key challenges in the construction of a new integrated power system of "source, grid, load, storage, and charging".

[0004] Currently, the coordinated scheduling of massive distributed resources has become one of the core businesses of distribution network production. However, there is a huge shortage of qualified dispatchers, and traditional models are insufficient to guarantee the safe and economical operation of the distribution network. The industry urgently hopes to leverage large-scale modeling technology to achieve intelligent monitoring and automatic scheduling of distributed resources, promote the evolution of dispatching operations from "manually driven" to "unmanned," and support the construction of a new power system based on new energy sources.

[0005] CN120337977A proposes a task understanding and execution system based on multimodal intelligent agents. It can handle multimodal inputs such as voice, text, and images, utilizing a large language model to complete intent recognition, subtask decomposition, and plug-in-level execution. Furthermore, it leverages knowledge memory and self-learning mechanisms to improve the interactive experience in common scenarios such as office work and assistant tasks. However, this system is fine-grained and short-cycle, focusing on single-point function calls and lacking unified scheduling, state tracking, and anomaly recovery capabilities for cross-stage, long-cycle task flows. Moreover, its knowledge base is generic and lacks industry standards and operational constraints, making it difficult to meet the demands of high real-time and high-reliability scenarios.

[0006] In summary, current large-scale modeling technologies mainly focus on semantic understanding and text generation, lacking specific technologies for replacing power dispatching tasks, especially for enabling AI to automatically execute power dispatching tasks item by item. However, power dispatching task flows differ from traditional computer programs; the execution of AI-powered power dispatching task flows typically requires significant computing power and time (on the order of seconds). How to plan and manage power dispatching workflows to replace dispatchers' daily tasks is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a task flow control intelligent agent and control method for intelligent power monitoring. It aims to gradually replace power dispatching business with the intelligent capabilities of a large model and realize intelligent management and automatic execution of task flows for different dispatching businesses.

[0008] The technical solution adopted by this invention to solve the technical problem is: The first aspect of this invention is to provide a task flow control agent based on a large-scale model for intelligent power monitoring, primarily used for controlling and regulating business task flows. As a component of a large-scale model-based intelligent power grid dispatching system, this agent can realize functions such as execution control of dispatching task flows, human-computer interaction, strategy deployment, and memory storage.

[0009] The intelligent agent proposed in this invention includes the following modules: human-computer front-end module, human-computer interaction compilation module, scheduling auxiliary strategy library, scheduling operation constraint target library, task flow library, task flow management core, operation memory module, and operation result display module.

[0010] The human-machine front-end module is implemented by receiving instructions from power dispatchers through text input boxes and voice input modules, and converting them into text commands or text-based questions and answers. These are then passed to the human-machine interaction compilation module. This module also receives the question-and-answer results from the human-machine interaction compilation module to enable interaction with dispatchers.

[0011] The human-computer interaction compilation module extracts and pre-screens key content from user text commands or questions and answers, and determines whether it is necessary to call the scheduling auxiliary strategy library or the scheduling operation constraint target library to form prompt words for the first call to the large model; Intent recognition is performed: if it is a question and answer, the task flow management core is called to generate a prompt word, which is then processed by the large model to answer the user's question; if it is a command, the prompt word is further input into the large model for another call to generate a workflow execution command in JSON format, which is then transmitted to the task flow management core, and the corresponding task is added to the pending workflow library of the task flow management core.

[0012] The dispatching auxiliary strategy library classifies and organizes dispatching business-related terms to build a keyword matching library, including equipment name library, power terminology library, dual numbering library, dispatching instruction keywords, task flow description keywords, etc.; and uses RAG (retrieval enhancement generation) technology to aggregate dispatching operation instruction tickets, switching operation tickets, dispatching operation procedures, accident handling plans, and operation procedures of transmission and distribution equipment to form a dispatching knowledge auxiliary library.

[0013] The scheduling and operation constraint target library mainly includes two types: safety constraints and optimization targets. The safety constraints mainly summarize the constraints related to grid security in the regional power system, such as the current carrying capacity limit and voltage limit of each device, which are used to support the large model to call the constraints during the calculation process.

[0014] The optimization objectives mainly refer to the optimization objectives of a certain regional power system, such as economic operation, safe operation, and new energy consumption, established by power dispatching experts after analysis. The construction of power grid safe operation objectives facilitates program calls when the large model executes the workflow.

[0015] The task flow library refers to a workflow library that stores the automated execution flow of scheduling tasks. The automated execution flow of scheduling tasks refers to the process where scheduling experts analyze daily scheduling tasks, and workflow developers, in conjunction with these experts, transform these manual tasks into task flows, which are then used to gradually replace manual scheduling tasks with artificial intelligence.

[0016] The task flow management core is the core of intelligent task flow management for scheduling business. It is implemented by calling the task flow in the task flow library through pure software code, and can realize the task flow control function, including workflow input interface, pending workflow management module, server computing power management module, executing workflow management module, executed workflow management module, and execution result storage and retrieval module.

[0017] The runtime memory module refers to the information storage module for all the information during the operation of this intelligent agent. This module stores all task operation logs, including human-computer interaction logs, workflow execution records, workflow execution result records, computing server usage records, real-time and historical operating status of system hardware, real-time and historical data of power dispatching system, etc., to support the closed-loop operation of the entire intelligent agent.

[0018] The execution results display module addresses the issue that large-scale model tasks typically require a certain amount of time (in seconds) to execute. This invention uses a workflow task pool to display the progress percentage of tasks being executed by the task flow management core, pending workflows, and completed workflows, summarizing and displaying the execution results of all workflows. Simultaneously, this module can control the running task flows within the task flow management core.

[0019] Furthermore, the operation results display module can summarize and display statistical data, including all information of executed workflows, number of anomalies, number of generated strategies, and number of reports, to present the operation results of the autonomous cruise of the power system in an intuitive form.

[0020] Furthermore, the expected computing power consumption and workflow introduction are displayed in the pending workflow; the progress percentage and computing power consumed are displayed in the executing workflow; and the actual computing power used and workflow result report are displayed in the executed workflow.

[0021] A second aspect of the present invention is to provide a control method for the above-described intelligent agent, comprising the following steps: The S1 human-machine front-end module receives voice or text instructions from the dispatcher and converts them into text commands / questions and answers; The S2 human-computer interaction compilation module extracts key content and identifies intent from the text, and generates large model prompt words by combining the scheduling auxiliary strategy library and the scheduling operation constraint target library; S3 If the intent is a question-and-answer session, call the large model and return the result to the human-computer front-end module; if the intent is a command, call the large model to generate a JSON-formatted task flow command and send it to the task flow management core. The S4 task flow management core loads the corresponding workflow from the task flow library, performs syntax and semantic verification, adds it to the execution queue according to priority, and starts execution when the computing power meets the threshold, and tracks the progress, computing power consumption and intermediate results in real time. S5 stores the logs, results, and system status during task execution to the runtime memory module; S6 displays the execution status, progress, results, and statistical data of the task flow through the execution results display module.

[0022] Furthermore, S4 specifically refers to: The workflow input interface validates the syntax and semantics of the received workflow. The workflow management module caches the verified tasks in a Redis (remote dictionary service) queue and schedules them according to priority, starting execution when the server's GPU / CPU remaining computing power is ≥60%; The server computing power management module monitors and provides feedback on computing load in real time. The workflow management module is currently tracking the task execution progress through a state machine and providing real-time feedback to the human-machine front-end interaction module on the percentage of the workflow being executed and intermediate results. The executed workflow management module archives the status of completed tasks and computing power consumption to the time-series database; The execution result storage and retrieval module writes structured execution results into the database, supporting multi-dimensional queries by task ID and time range.

[0023] The advantages and positive effects of this invention are: 1. This invention application provides a task flow control intelligent agent based on a large model for intelligent power monitoring, so as to realize the intelligence of the large model to achieve the item-by-item substitution of power dispatching business, and realize intelligent management and automatic execution of task flows for different dispatching businesses.

[0024] 2. This invention further realizes the specific path for large-scale model-enabled intelligent replacement of power system dispatching services.

[0025] 3. This invention further realizes the intelligent monitoring of massive regulatory data by a large model and the real replacement of scheduling business under limited computing power by coordinating the computing power consumption of the workflow and adjusting the execution frequency of the workflow.

[0026] 4. The functional modules of this invention are clear, the logic is closed-loop, and it is easy to implement. It has already been successfully applied in the field. Attached Figure Description

[0027] Figure 1 This is a structural block diagram of the present invention.

[0028] Figure 2 This is a flowchart of the process when a specific workflow is executed. Detailed Implementation

[0029] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0030] The task flow control agent based on a large model for intelligent power monitoring provided by this invention is mainly used for controlling and regulating business task flows. As a component of a large model-based intelligent power grid dispatching system, this agent can realize functions such as execution control of dispatching task flows, human-computer interaction, strategy deployment, and memory storage.

[0031] like Figure 1 As shown, the intelligent agent proposed in this invention includes the following modules: human-computer front-end module, human-computer interaction compilation module, scheduling auxiliary strategy library, scheduling operation constraint target library, task flow library, task flow management core, operation memory module, and operation result display module.

[0032] The aforementioned human-machine front-end module mainly refers to the front-end human-machine interaction interface deployed in the dispatcher's workstation, providing two input methods: one is voice input, which uses a voice recognition model (such as an ASR engine) to convert the voice commands of the distribution network dispatcher into text; the other is text input, where the front-end interaction module receives text instructions and then sends them to the human-machine interaction compilation module in the back-end.

[0033] The human-computer interaction compilation module, combined with the keyword matching library (such as the device dual name library) in the scheduling auxiliary strategy library, re-recognizes and converts voice commands, and combines the intent analysis capabilities of the language big model.

[0034] Furthermore, in the case of a question-and-answer session, the system runtime memory module and the large model are invoked to answer the user's question. The system runtime information memory module can be invoked in RAG format, and historical data can be invoked in NL-to-SQL (Natural Language to SQL) format. Prompt words are constructed for the large model to analyze and then provide feedback to the user.

[0035] Furthermore, if the command is a scheduler command, it is transformed into a rule-based model instruction by combining the scheduling assistance strategy library and the scheduling execution constraint target library. A large model prompt containing contextual constraints is generated through preset rules. The prompt is parsed by the large model, and a task sequence in JSON format conforming to a predetermined specification is output and pushed to the task queue of the task flow management core.

[0036] The implementation of the aforementioned scheduling assistance strategy library is relatively simple. This module is based on Elasticsearch (an elastic search, distributed search and analytics engine) to build a Retrieval Enhanced Generation (RAG) architecture, and integrates internal enterprise scheduling knowledge documents (such as unstructured texts like instruction tickets, contingency plans, and procedures). After preprocessing (word segmentation and vectorization), the knowledge documents are stored in a vector database (such as FAISS). At runtime, the system retrieves relevant strategy clauses in real time based on keywords extracted from the front end and the operational context, embedding them into large model prompts to enhance the accuracy and compliance of the generated content.

[0037] The implementation of the aforementioned scheduling and operation constraint target library mainly involves transforming the constraints of power system operation into specific equipment limits, including equipment limit tables (such as line current carrying capacity, transformer capacity, and voltage upper and lower limits). It also quantifies the power grid operation targets into specific equipment formulas, such as the minimum network loss mode and the maximum renewable energy absorption target, forming quantifiable targets for use in specific workflows.

[0038] The task flow library refers to a workflow library that stores and schedules tasks for automatic execution. It is developed by workflow developers in collaboration with scheduling experts, who transform routine manual tasks into task flows for core task flow management.

[0039] The core of task flow management is implemented using a microservice architecture, including the following sub-services: Workflow Input Interface: Receives JSON task flows based on a RESTful (Representational State Transition) API, performing syntax and semantic validation; Pending Workflow Management Module: Uses Redis to cache task queues, supports priority scheduling and concurrency control, and plans execution workflows based on server computing power, initiating pending workflows when computing power remains above 60%; Executing Workflow Management Module: Tracks task execution progress through a state machine, providing real-time feedback to the human-machine interface module on the percentage of executing workflows and intermediate results; Executed Workflow Management Module: Archives the status of completed tasks and consumed computing power to a time-series database (such as the InfluxDB open-source time-series database), and stores execution results in the execution result module; Execution Result Storage and Retrieval Module: Writes structured results to the database, interfaces with the human-machine interface module, and supports multi-dimensional queries by task ID, time range, etc.; Server Computing Power Management Module: Monitors GPU / CPU resource usage status and provides computing power load feedback to pending workflows.

[0040] The runtime memory module is primarily used to store the full runtime logs, including human-computer interaction records, task flow execution snapshots, task flow execution results, server resource snapshots, and full power grid telemetry and telecontrol data. Data is stored according to dimensions such as agent runtime data, power grid historical data, workflow execution snapshots and results, and human-computer interaction data. Text information is stored in vector knowledge base format, and data information is stored in relational database format. During front-end human-computer interaction, large-scale natural language to SQL (NLtoSQL) commands and RAG technology are used to perform closed-loop access to all memory information, achieving intelligent closed-loop operation within the entire agent.

[0041] The execution result display module mainly refers to the front-end interface deployed in the scheduler's workstation. This interface provides real-time feedback on the execution results of each workflow, using a progress bar-like approach to show the progress of the workflow in execution, workflows to be executed, and completed workflows.

[0042] This invention also provides a control method for the aforementioned intelligent agent, such as... Figure 2 As shown, it includes the following steps: The S1 human-machine front-end module receives voice or text instructions from the dispatcher and converts them into text commands / questions and answers; The S2 human-computer interaction compilation module extracts key content and identifies intent from the text, and generates large model prompt words by combining the scheduling auxiliary strategy library and the scheduling operation constraint target library; S3 If the intent is a question-and-answer session, call the large model and return the result to the human-computer front-end module; if the intent is a command, call the large model to generate a JSON-formatted task flow command and send it to the task flow management core. The S4 task flow management core loads the corresponding workflow from the task flow library, performs syntax and semantic verification, adds it to the execution queue according to priority, and starts execution when the computing power meets the threshold, and tracks the progress, computing power consumption and intermediate results in real time. S5 stores the logs, results, and system status during task execution to the runtime memory module; S6 displays the execution status, progress, results, and statistical data of the task flow through the execution results display module.

[0043] Furthermore, S4 specifically refers to: The workflow input interface validates the syntax and semantics of the received workflow. The workflow management module caches the verified tasks in a Redis queue and schedules them according to priority, starting execution when the server's GPU / CPU remaining computing power is ≥60%; The server computing power management module monitors and provides feedback on computing load in real time. The workflow management module is currently tracking the task execution progress through a state machine and providing real-time feedback to the human-machine front-end interaction module on the percentage of the workflow being executed and intermediate results. The executed workflow management module archives the status of completed tasks and computing power consumption to the time-series database; The execution result storage and retrieval module writes structured execution results into the database, supporting multi-dimensional queries by task ID and time range.

[0044] This invention has been pre-developed in the laboratory. Under the premise of existing large model technology, the effect meets the expectations of dispatchers and can replace 30% of the current work of distribution network dispatchers. It can be applied in the field.

[0045] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the inventive concept, and these all fall within the protection scope of the present invention.

Claims

1. A task flow control intelligent agent for intelligent power monitoring, characterized in that, include: The human-machine front-end module is used to receive voice or text instructions from power dispatchers, convert them into text commands or text Q&A, and provide feedback on the interaction results. The human-computer interaction compilation module is used to extract key content and recognize intent from user text commands / questions and answers, generate prompt words for the large model, and call the large model to generate task flow commands; The dispatching auxiliary strategy library is constructed by classifying and building a matching library of equipment names, power terms, and dispatching instruction keywords, and aggregating instruction tickets, procedures, and contingency plan texts to form a dispatching knowledge auxiliary library; The scheduling and operation constraint target library summarizes power grid security constraints for use in large model calculations. A task flow library is used to store standardized task flows built by scheduling business experts; The task flow management core is used to receive and manage task flow commands, and control the execution, scheduling and result storage of task flows; The runtime memory module is used to record task execution logs, interaction records, system status, and power grid data. The execution results display module is used to show the execution status, progress, results, and statistical data of the task flow.

2. The intelligent agent according to claim 1, characterized in that, The human-machine interface module provides two input methods. First, voice input: a voice recognition model is used to convert the voice commands of the distribution network dispatcher into text. Second, text input: after receiving a text command, it is sent to the human-computer interaction compilation module.

3. The intelligent agent according to claim 1, characterized in that, The human-computer interaction compilation module extracts and pre-screens key content from text commands or questions and answers from the human-computer front-end module, and determines whether to call the scheduling auxiliary strategy library or the scheduling operation constraint target library to generate the first large model prompt words. If the intent is to ask and answer questions, the runtime memory module is invoked to generate prompts, the large model generates the answer, and the answer is returned to the human-computer front-end module. If the intent is a command, the prompt word is input into the large model again, the workflow execution task command is output, and it is sent to the pending workflow library in the task flow management core.

4. The intelligent agent according to claim 1, characterized in that, The scheduling assistance strategy library is built on an Elasticsearch-based search enhancement and generation RAG architecture. It integrates with the enterprise's internal scheduling knowledge documents. After preprocessing, the knowledge documents are stored in a vector database. At runtime, the system retrieves relevant strategy clauses in real time based on keywords extracted from the front end and the operation context, and embeds them into the prompts in the large model.

5. The intelligent agent according to claim 1, characterized in that, The aforementioned scheduling and operation constraint target library transforms the constraints of power system operation into limits for specific equipment, quantifies the power grid operation targets into specific equipment formulas, and forms quantifiable targets for workflow invocation.

6. The intelligent agent according to claim 1, characterized in that, The core of the task flow management system includes: Workflow input interface, validating workflow syntax and semantics; The pending workflow management module is used to cache verified tasks in a Redis queue and schedule them according to priority, starting execution when the server's GPU / CPU remaining computing power is ≥60%; The server computing power management module is used to monitor and provide feedback on computing load in real time. The workflow management module is currently executing, tracking the task execution progress through a state machine and providing real-time feedback to the human-machine front-end interaction module on the percentage of workflow execution and intermediate results. The workflow management module has been executed to archive the status of completed tasks and computing power consumption to the time-series database; The execution result storage and retrieval module is used to write structured execution results into the database and supports multi-dimensional queries by task ID and time range.

7. The intelligent agent according to claim 1, characterized in that, The runtime memory module is used to store human-computer interaction logs, workflow execution records and results, computing server usage records, real-time and historical status of system hardware, and real-time and historical data of the power dispatching system. It is organized in the form of vector library and / or relational library to provide data support for task tracing, model prompting optimization and intelligent agent closed-loop operation.

8. The intelligent agent according to claim 1, characterized in that, The aforementioned execution result display module displays the progress percentage of the task flow being executed by the task flow management core, the pending workflows, and the completed workflows in the form of a workflow task pool, and summarizes and displays the execution results of all workflows.

9. The control method for an intelligent agent according to any one of claims 1-8, characterized in that, Includes the following steps: The S1 human-machine front-end module receives voice or text instructions from the dispatcher and converts them into text commands / questions and answers; The S2 human-computer interaction compilation module extracts key content and identifies intent from the text, and generates large model prompt words by combining the scheduling auxiliary strategy library and the scheduling operation constraint target library; S3 If the intent is question-and-answer, call the large model and return the result to the human-computer front-end module; If the intent is a command, invoke the large model task flow command and send it to the task flow management core; The S4 task flow management core loads the corresponding workflow from the task flow library, performs syntax and semantic verification, adds it to the execution queue according to priority, and starts execution when the computing power meets the threshold, and tracks the progress, computing power consumption and intermediate results in real time. S5 stores the logs, results, and system status during task execution to the runtime memory module; S6 displays the execution status, progress, results, and statistical data of the task flow through the execution results display module.

10. The control method according to claim 9, characterized in that, S4 specifically refers to: The workflow input interface validates the syntax and semantics of the received workflow. The workflow management module caches the verified tasks in a Redis queue and schedules them according to priority, starting execution when the server's GPU / CPU remaining computing power is ≥60%; The server computing power management module monitors and provides feedback on computing load in real time. The workflow management module is currently tracking the task execution progress through a state machine and providing real-time feedback to the human-machine front-end interaction module on the percentage of the workflow being executed and intermediate results. The executed workflow management module archives the status of completed tasks and computing power consumption to the time-series database; The execution result storage and retrieval module writes structured execution results into the database, supporting multi-dimensional queries by task ID and time range.

Images