A runtime process trust verification system and method for a power grid intelligent agent

Abstract

This invention provides a runtime process trustworthy verification system and method for smart grid agents, relating to smart grid automation, artificial intelligence, multi-agent systems, and industrial process safety. The system includes: generating a full-link governance trajectory of the smart grid through a decision-making agent based on grid dispatch instructions and multimodal perception data; performing real-time runtime auditing of the generated full-link governance trajectory through a verification agent to detect trajectory anomalies and generate a digital signature token for runtime process trustworthy verification; and distributing the real-time audited full-link governance trajectory to the collaborative execution layer based on the digital signature token to complete physical operations and form a closed loop. This invention introduces a verification agent independent of the decision-making loop, which, through a trajectory anomaly detection mechanism, performs real-time logical auditing of the decision-making agent's "perception-thinking-action" full-link trajectory, achieving a leap from "post-event interception at the result side" to "runtime blocking at the process side."

Description

Technical Field

[0001] This invention relates to the fields of smart grid automation, artificial intelligence, multi-agent systems, and industrial process safety, and specifically to a runtime process trust verification system and method for smart grid agents. Background Technology

[0002] With the development of large language models and embodied intelligence technologies, smart grid operation and maintenance is transforming from rule-based "automation" to cognition-based "autonomy." However, directly applying agents with probabilistic generation characteristics to zero-fault-tolerant high-voltage power scenarios faces core challenges: (1) The fundamental contradiction between deterministic security and probabilistic generation: Power grid dispatch and control require absolute determinism, while LLM-based agents are essentially probabilistic models. When dealing with unseen complex operating conditions (such as extreme ice storms combined with equipment failures), agents are prone to hallucinations. Existing Supervisory Control and Data Acquisition (SCADA) systems cannot identify "seemingly reasonable but physically fallacious" instructions generated by agents (e.g., mistakenly disconnecting critical loads to reduce transformer temperature). Such "serious nonsense" is fatal in the physical world.

[0003] (2) "Result compliance" masks "process violation": Current power grid safety verification mainly focuses on the boundaries of actions (e.g., whether the voltage exceeds the limit, whether the switch is allowed to operate). However, the danger of intelligent agents is often hidden in the reasoning path. For example, an intelligent agent executes a correct "trip" action, but its reasoning is wrong (e.g., misjudging the sensor reading). Although this "accidental" decision does not cause an accident at present, it means that the cognitive logic of the intelligent agent has deviated from the real physical environment and may cause a disaster in the next task at any time. Traditional result-based monitoring mechanisms are completely blind to this.

[0004] (3) Cognitive drift and logical degradation in long-term tasks: Power grid disaster management is usually a long-chain process (perception-analysis-planning-scheduling-execution-review). Studies have shown that agents are prone to attentional divergence or goal loss when performing such multi-step tasks. For example, during the de-icing task, the agent may fall into unnecessary interaction loops due to irrelevant environmental noise (such as interference from passersby), leading to timeouts of critical tasks. Currently, there is a lack of an effective runtime mechanism to maintain the logical coherence of agents in long-cycle tasks.

[0005] The aforementioned challenges stem from the lack of process supervision. Existing technologies primarily focus on static checks of "input validity" and "output compliance," lacking dynamic auditing of the agent's intermediate reasoning process. In power scenarios, a single erroneous reasoning step (e.g., pushing the power flow of critical transmission sections to their limits in pursuit of optimal economic dispatch, while ignoring the cascading overload risk following an 'N-1' sudden failure) can lead to catastrophic consequences (causing widespread cascading trips or even grid disconnection). When executing complex, long-range tasks, agents are highly susceptible to anomalies such as process inefficiency, task failure, and unnecessary continuation. In high-voltage grid environments, such uncontrollable processes are unacceptable. Summary of the Invention

[0006] To address the limitations of existing power grid automation systems in effectively monitoring the "black box" decision-making logic of large model-based agents (LLM-based agents), and the technical shortcomings of traditional security verification mechanisms in lacking dynamic auditing of the chain-of-thought process, this invention proposes a runtime process trustworthy verification system and method for power grid agents. It introduces a verification agent independent of the decision-making loop and, through a trajectory anomaly detection mechanism, performs real-time logical auditing of the entire "perception-thinking-action" trajectory of the decision-making agent, achieving a leap from "post-event interception on the result side" to "runtime blocking on the process side."

[0007] According to some embodiments, the present invention adopts the following technical solution: A runtime process trust verification system for power grid intelligent agents includes: The decision-making agent is used to generate the full-link governance trajectory of the smart grid based on the received power grid dispatch instructions and multimodal sensing data, and send it to the verification agent. The verification agent is used to perform real-time runtime auditing of the generated end-to-end governance trajectory based on multimodal perception data that is heterogeneous with the decision-making agent, detect trajectory anomalies, generate a digital signature token for trusted verification of the runtime process, and send it along with the end-to-end governance trajectory to the execution controller. In real-time auditing, if the detected trajectory does not conform to the objective facts of the physical world, the dynamic blocking of the runtime process is immediately triggered, and the virtual rollback mechanism is activated. The execution controller is used to distribute the end-to-end governance trajectory, based on digital signature tokens and real-time auditing, to the collaborative execution layer to complete physical operations and form a closed loop.

[0008] According to some embodiments, the present invention adopts the following technical solution: A runtime process trust verification method for power grid intelligent agents includes: Based on the received power grid dispatch instructions and multimodal sensing data, a full-link governance trajectory for the smart grid is generated; Based on multimodal perception data that is heterogeneous with the decision-making intelligent agent, the generated end-to-end governance trajectory is audited in real time during runtime, trajectory anomalies are detected, and a digital signature token for trusted verification of the runtime process is generated. Based on digital signature tokens, the entire governance process is executed through real-time auditing, completing physical operations and forming a closed loop; In real-time auditing, if the detected trajectory does not conform to the objective facts of the physical world, the dynamic blocking of the runtime process is immediately triggered, and the virtual rollback mechanism is activated.

[0009] According to some embodiments, the present invention adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned runtime process trust verification method for a power grid intelligent agent.

[0010] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned runtime process trusted verification method for a power grid intelligent agent.

[0011] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a runtime process trusted verification method for a power grid intelligent agent.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates trajectory anomaly detection technology. By constructing an objective "environmental truth" model through independent multimodal perception channels that do not reuse decision states, it performs real-time causal auditing on the entire cognitive trajectory generated by the decision-making agent, which includes a "thought-action" sequence. Based on this, a millisecond-level logical firewall is established before physical commands are issued. Once cognitive anomalies such as task failure, inefficient processes, or unnecessary continuation are detected, dynamic blocking is immediately triggered. The agent's state is reset to the moment before the error occurred through virtual rollback, and structured correction prompts are injected to guide it to replan its path. This ensures the logical rigor and operational safety of the AI's decision-making process in complex high-voltage power scenarios. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is the overall logical architecture diagram of Example 1.

[0015] Figure 2 This is a flowchart of the runtime audit process for Example 2.

[0016] Figure 3 This is a timing diagram of dynamic blocking and virtual rollback in Example 2. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0020] Example 1 One embodiment of the present invention provides a runtime process trust verification system for power grid intelligent agents, which aims to achieve the following specific objectives: (1) To resolve the fundamental conflict between the requirement of absolute determinism in power grid dispatch and the probabilistic generation characteristics of large model agents.

[0021] To address the issue of hallucinations that can easily occur when agents handle complex and unknown conditions, and the difficulty of identifying "physical fallacies" in existing systems, this embodiment proposes constructing a "Ground Truth" benchmark independent of the decision loop. By establishing a heterogeneous sensing verification agent, objective physical data is used to provide a "third eye" for logical falsification of the agent's subjective decisions. Simultaneously, a mandatory "digital logic firewall" is deployed before physical commands are issued, ensuring that only commands that pass objective truth verification can be executed. This approach preserves the flexibility of the large model while ensuring deterministic security of power grid control through external constraint mechanisms.

[0022] (2) Break through the limitations of traditional safety verification mechanisms that only focus on action boundaries (such as voltage over-limit), and solve the hidden risk of "correct results stemming from incorrect reasons".

[0023] To address the limitation of existing technologies in auditing the reasoning path of intelligent agents, this embodiment provides a deep cognitive auditing mechanism based on end-to-end trajectory anomaly detection. This mechanism not only examines "what was done" (Action) but also delves deeper into "why it was done" (Chain-of-Thought), enabling real-time identification of logical fallacies hidden behind compliant actions (such as incorrect attribution and causal reversal). This ensures that when an intelligent agent executes a task, not only are the results consistent with electrical constraints, but its cognitive logic also strictly conforms to physical laws and environmental facts.

[0024] (3) Solve the problems of attention divergence, target loss or getting stuck in a dead loop (such as unnecessary interaction) that are easy to occur when intelligent agents perform long chain complex tasks.

[0025] To address the current lack of runtime logic maintenance mechanisms, this embodiment endows the system with adaptive error correction capabilities based on "dynamic blocking - virtual rollback - guided retry". By monitoring abnormal states such as process inefficiency and unwarranted continuation in real time, the system can cut off the erroneous path within milliseconds. Through "virtual rollback" technology, it resets the agent's memory state to the state before the error occurred and injects correction prompts to guide its self-correction. This significantly improves the logical coherence and success rate of long-term unmanned operation and maintenance tasks without human intervention.

[0026] Based on the above objectives, this embodiment proposes a runtime process trust verification system for power grid intelligent agents, comprising: The decision-making agent is used to generate the full-link governance trajectory of the smart grid based on the received power grid dispatch instructions and multimodal sensing data, and send it to the verification agent. The verification agent is used to perform real-time runtime auditing of the generated end-to-end governance trajectory based on multimodal perception data that is heterogeneous with the decision-making agent, detect trajectory anomalies, generate a digital signature token for trusted verification of the runtime process, and send it along with the end-to-end governance trajectory to the execution controller. In real-time auditing, if the detected trajectory does not conform to the objective facts of the physical world, the dynamic blocking of the runtime process is immediately triggered, and the virtual rollback mechanism is activated. The execution controller is used to distribute the end-to-end governance trajectory, based on digital signature tokens and real-time auditing, to the collaborative execution layer to complete physical operations and form a closed loop.

[0027] This embodiment designs a runtime process trustworthy verification framework for a power grid intelligent agent. It adopts a decision-regulatory orthogonal architecture independent of the decision-making loop. While acknowledging the unexplainability and probabilistic risks of large-scale intelligent agents in complex power grid environments, it introduces a verification intelligent agent independent of the decision-making loop. Through a trajectory anomaly detection mechanism, it performs real-time logical auditing of the entire "perception-thinking-action" trajectory of the decision-making intelligent agent, achieving a leap from "post-event interception on the result side" to "runtime blocking on the process side." The specific implementation process is described below: like Figure 1 As shown, the decision-regulatory orthogonal architecture establishes mutually independent and physically isolated regulated objects and regulators in the system logic, and consists of three core logical units: the regulated object (decision agent), the regulator (verification agent), and the intervention controller (execution controller).

[0028] I. Subject under supervision: The embodied decision-making agent (Governor Agent / The Subject) is the "executor" in the framework. It consists of decision-making agents based on large language models (LLM) and is responsible for handling complex situations and generating governance strategies. The system regards it as a "black box" object with probabilistic risks.

[0029] The term "black box" refers to the framework's approach of not interfering with the internal model training but treating it as a "black box" object that may produce illusions or logical fallacies, and externally monitoring its output trajectory. The decision-making agent's function is to receive power grid dispatch instructions from the dispatch master station or task orchestration system, and based on decision data, generate a full-link governance trajectory T that includes intermediate reasoning steps (Chain-of-Thought). Decision data includes, but is not limited to: (1) Task and business instruction data: human-machine instructions, work order tasks, operation objectives and constraints (such as target area, time limit, allowed operation type, etc.) from the scheduling master station / task orchestration system; (2) Power grid operation and equipment status data: real-time measurements and status quantities from production control systems such as SCADA, including voltage, current, active / reactive power, frequency, power flow, circuit breaker / disconnector status, tap position, alarm information, etc.; when necessary, phasors and dynamic characteristic quantities from PMU / synchronous measurement systems are also included. (3) Power grid model and rule data: structured knowledge such as power grid topology / primary wiring relationship, equipment rated parameters and operating boundaries, protection settings / blocking logic, switching operation ticket rules and typical operation sequences; (4) Task execution side perception data: On-site perception data transmitted back by the execution terminal (e.g., inspection robot, drone, fixed camera, online monitoring device) controlled or scheduled by the decision-making intelligent agent, including images / videos, infrared thermal images, point clouds, positioning information and the execution terminal's own status (battery, attitude, health, etc.). (5) Historical and contextual data: historical operating curves, historical alarm / defect records, maintenance records, operation logs, and recent environmental change information related to the current task (such as load fluctuations, weather changes, etc.).

[0030] The power grid dispatch instructions (e.g., "detect and handle icing in area A"), along with the aforementioned decision data, are input into the large language model for the current round of thinking / planning, based on the generated thought chain. Give the corresponding action instructions (i.e., execute the action) ), and then if the result of this reasoning (i.e., the "thought chain") -action Once the intelligent agent passes the verification test, it is truly implemented on the actual device, and then generates an impact or perception on the environment (i.e., environmental feedback). In other words, the thought chain of the intelligent agent must be generated first. (For example, "Abnormal tension detected, ice suspected, camera needed for confirmation") and execution actions. (For example, "Camera_05.capture()"), after being audited and executed, generates environmental feedback. .

[0031] By continuously collecting decision-making data and reasoning about intermediate governance steps, the power grid governance task is modeled as a sequential decision-making process. The end-to-end governance trajectory T, including intermediate reasoning steps (Chain-of-Thought), is defined as:

[0032] Where I represents the power grid dispatch instruction. For the thought chain at step t, For the action to be performed in step t, This is the environmental feedback for step t.

[0033] Based on the above, the specific steps executed by the decision-making agent are as follows: The decision-making agent receives power grid dispatch instructions and generates a governance plan that includes intermediate reasoning processes based on its internal Large Language Model (LLM), specifically: Task reception: The decision-making agent receives task instructions I in natural language form.

[0034] Thought Chain Reasoning: Intelligent Agent Generates Thought Chains (Reasoning / Thought), explaining the basis for the decision.

[0035] Action formulation: Based on the reasoning results, the agent formulates the physical actions to be performed. (Action).

[0036] Trajectory Generation: The above process forms a governance trajectory containing a sequence of "instruction-thought-action-observation". And submit it to the verification agent for auditing.

[0037] II. Regulatory Entity: The Trajectory Verification Agent (The Supervisor) is the core innovative entity in this embodiment. It is led by a deterministic Verification Agent and operates in parallel as a "digital supervisor".

[0038] To provide an objective verification benchmark for probabilistic subjective decision-making, the verification agent employs source-heterogeneous multimodal sensory data. By integrating objective physical data such as vision, mechanical sensing, and micro-meteorology, a real-time "environmental truth value" independent of the decision-making logic is constructed. This truth value serves as the sole objective standard (Third Eye) for logical falsification, used to audit in real time whether the "thought-action" chain generated by the decision-making agent conforms to the objective facts of the physical world. This identifies and filters out illusory instructions that "seem reasonable but are physically fallacious" (e.g., in a contradictory situation where infrared data drift indicates high temperature but the load current is normal, the multimodal truth value determines that there is no overheating). It not only judges whether the action is illegal but also whether the reasoning is valid (e.g., detecting "erroneous attributions caused by sensor drift"). The specific steps are as follows: 1. Data Collection Verification agents do not share the internal state or context of the decision-making agent, i.e., independent perception (Ground Truth Construction). Multimodal perception data is collected through independent perception channels, including: (1) Independently deployed redundant measurement data: sensing / acquisition sources that are independent of the main business measurement channels read by the decision-making agent, such as independent RTU / IED acquisition links, independent PMU / wave recording devices, edge-side read-only mirror measurement, dual-channel metering / sensor redundancy, etc., used to cross-verify key electrical quantities and status quantities. (2) Heterogeneous multimodal physical sensing data: Multimodal sensing data directly accessed by the verification agent and different from the sensing source of the decision agent, including but not limited to micro-weather stations (temperature, humidity, wind speed, precipitation, etc.), binocular / multi-view vision, infrared thermal imaging, lidar (LiDAR), tension / tilt angle and other mechanical and attitude sensing data, which are used to form objective criteria for risks such as icing, external intrusion, equipment overheating, and abnormal mechanical attitude; (3) Equipment event and timing evidence data: traceable evidence from SOE (Sequence of Event) sequence event records, protection action / alarm records, waveform files, communication and control command logs, etc., used to causally align whether the action occurred, when it occurred, and whether the triggering cause chain is consistent; (4) Physical consistency results obtained by independent calculation: Verification results obtained based on power grid mechanism model and digital simulation calculation, such as power flow verification, constraint verification, N-1 safety verification, dynamic stability / limit exceedance risk assessment, and evolution prediction results such as icing growth and wind-induced sway, are used as a deterministic verification benchmark for the reasoning premise and action necessity of decision-making agents.

[0039] Furthermore, the aforementioned multimodal perception data remains independent or heterogeneous from the decision data of the decision-making agent in the acquisition device, upstream system, or communication link. This enables the verification agent to cross-falsify the "thinking-action" chain without reusing the internal state and context of the decision-making agent, thereby identifying and filtering abnormal planning and control instructions that "seem reasonable but are physically fallacious".

[0040] 2. Edge feature fusion Edge computing nodes run multimodal fusion algorithms. For icing scenarios, a correlation-based fusion network is used to construct a covariance matrix of visual texture and mechanical parameters to identify transparent ice or mixed ice conditions. For external damage scenarios, a spatiotemporal attention mechanism is used to align 2D visual targets with LiDAR's 3D point cloud and calculate the target's precise three-dimensional coordinates (x, y, z).

[0041] 3. Evolutionary trend prediction Run lightweight time-series evolution models, such as GM(1,N) or Kalman filtering, to predict the development trend of potential hazards within future time windows (such as ice thickness in the next 3 hours or the movement trajectory of a crane boom in the next 15 seconds).

[0042] 4. Environmental Truth Generation The verification agent generates the current environmental truth based on the aforementioned independent perception and inference results. As the sole objective benchmark for subsequent logical auditing, the specific generation steps are as follows: 1) Time alignment and unified calibration: Data from micro-weather stations, visual / infrared, LiDAR, tension / tilt angle and power grid operation measurements are uniformly clock-aligned and sampled and re-normalized to obtain a set of synchronous observations at the same time / within the same sliding window.

[0043] 2) Spatial registration and target alignment: Coordinate system transformation and registration are performed on visual targets, point clouds, equipment poses, and geographic coordinates to map "2D image target - 3D point cloud - tower / line GIS location" to a unified spatial reference system, so as to obtain the calculable object position, distance, attitude and relative relationship.

[0044] 3) Anomaly removal and credibility assessment: Based on sensor health, data consistency and historical statistical characteristics, anomaly detection and filtering are performed on the observations (e.g., outlier removal, drift identification, missing data completion), and credibility weights are assigned to each type of observation; 4) Multi-source fusion and state estimation: The processed multi-modal observations are fused to obtain the task-related environmental state estimation results; in different implementations, the fusion can adopt methods such as weighted consensus voting, Bayesian fusion, Kalman filtering or evidence theory fusion to output state variables with definite values ​​or interval values; 5) Mechanism consistency verification and constraint derivation: Input the fused state variables into the power grid mechanism model or physical rule base for consistency verification (e.g., power flow / limit violation / N-1 verification, icing and temperature / humidity / tension relationship verification, wind speed and flight safety rule verification), and derive the set of executable / non-executable safety constraints from this. 6) Structured encapsulation generation The above "state estimation results + risk assessment + constraint set + evidence chain" are encapsulated into a structured environmental truth value. Wherein, the environmental truth value It must include at least one or more of the following elements: Subset of truth values ​​for power grid operation : A snapshot of the status after consistency verification of voltage, current, active / reactive power, frequency, power flow, topology connectivity, switch / disconnector status, protection / alarm key events, etc. of the relevant sections / areas; Physical environment truth subset The set of mission-related environmental facts, including micro-meteorological conditions such as temperature, humidity, wind speed, precipitation, and visibility, as well as the presence / thickness range of icing, the temperature rise range of equipment surface, the three-dimensional coordinates (x,y,z) of the external target and its movement trend. Device and actuator truth subset Verifiable equipment status includes: tower / conductor stress and tilt angle, status of online monitoring devices, and drone / robot pose and health (battery level, load, fault codes); Risk and Evolutionary Truth Subset : Predictions of risk evolution within future time windows (e.g., the icing growth trend in the next 3 hours, extrapolation of the target trajectory for breaching in the next 15 seconds), and the corresponding risk levels; Safety constraints and permitted action sets Constraints and permissible action ranges derived from mechanism verification and procedural rules (e.g., "Wind speed > threshold prohibits takeoff", "A certain switch is locked and cannot be operated", "Upper limit of current flow at a certain section"). Chain of evidence and confidence level The source identifier, timestamp, key observation summary, and confidence level / confidence interval corresponding to the above truth value entries are used to support the traceability of subsequent audits.

[0045] Therefore, environmental truth value It is not a single sensor reading, but a structured truth snapshot of the task-related "power grid status - physical environment - equipment status - risk trends - security constraints" after heterogeneous observation fusion and mechanism consistency verification, used to perform runtime falsification and auditing of the reasoning premises and action necessity of the decision-making agent.

[0046] 5. Real-time auditing Implementing a deep causal audit with both logical and physical dimensions differs from existing technologies that only target... This embodiment performs threshold checks on the action and also checks the thresholds for the action. → (Inferring causality from actions) and → (Observing the responsiveness of reasoning) Two-way verification is performed to ensure that every physical action is not only compliant in terms of electrical parameters, but also has sufficient necessity and rationality in terms of cognitive logic, thereby achieving a technological leap from "outcome-side compliance inspection" to "process-side cognitive auditing," specifically: (1) Consistency verification: Verify that the agent's reasoning premises in the governance trajectory T are consistent with the constructed environmental truth. The comparison shall include at least the following processes: 1) Premise extraction and structuring On the trajectory of governance The thought chain in ,action and the observations it cites Semantic parsing is performed to extract a computable set of reasoning premises. That is, the set of assertions. In some implementations, the set of reasoning premises... It is represented as a number of structured assertion units, each of which includes at least: object identifier (device / region / executor), attribute items (e.g., wind speed, ice thickness, on / off status, load rate, etc.), assertion type (numerical value / range / category / logical relationship), assertion value or threshold, timestamp, and confidence marker.

[0047] 2) Object mapping and spatiotemporal alignment Will Object identifiers in the mapping to environment truth values Object indexes (e.g., GIS points, line / interval IDs, execution unit IDs, etc.), and based on timestamps and Perform alignment at the same time or within the same sliding window; for assertions with spatial relationships (such as "the external target is inside / outside the protected area"), perform coordinate system consistency processing between the position description in the assertion and the 3D coordinates / attitude results in Gtenv or Gtasset.

[0048] 3) Item-by-item consistency determination For each assertion unit p∈ Based on the assertion type, the corresponding comparison rules and verification methods are invoked: Numerical / Interval Assertion Comparison: When the assertion value is numerical or an interval, the verification agent verifies the assertion from... , or Extract the corresponding truth value entry (definite value or confidence interval), calculate the difference and compare it with the tolerance threshold; if the truth value is an interval, determine whether the assertion value falls into the interval or has an intersection with the interval; if not, it is recorded as inconsistent. Categorical / State-based assertion comparison: When the assertion is a category label or state description (e.g., "no icing / icing present", "wind speed meets takeoff requirements", "switch is open / closed"), verify the agent's... Extract the corresponding state truth value and confidence level. If the categories are inconsistent and the truth value confidence level is higher than the preset threshold, the assertion is determined to be inconsistent. Topology / Procedure / Constraint Assertion Comparison: When assertions involve topological connectivity, locking logic, procedural constraints, or executable scope, verify agent calls. Safety constraints and permitted action sets are used to check the preconditions of the actions supported by the assertions; if the action triggering conditions caused by the assertion are not met... (For example, if the lockout is inoperable, the wind speed exceeds the limit and takeoff is prohibited, or the cross-sectional tidal current constraint is not met, etc.), then it is judged as inconsistent and marked as "hard inconsistency"; Evidence chain consistency comparison: If the inference premise claims in the trajectory that it is "derived from a certain observation / log / alarm", the verification agent further verifies the evidence chain consistency. The system retrieves the corresponding evidence summary, source identifier, and timestamp to verify whether the cited evidence exists, whether it is traceable from the same source, and whether it is valid within the same time window. If any information is missing or does not match, it is considered inconsistent.

[0049] 4) Consistency scoring and pass determination: In some implementations, the verification agent calculates a consistency score for each assertion unit and fuses them according to the importance weight of the assertions (e.g., safety-related assertions have higher weights) to obtain a trajectory-level consistency index. ; if and only if (i) there is no “hard inconsistency” (i.e., violation) (ii) When the consistency index meets the preset threshold, the governance trajectory T is determined to pass the consistency check, the agent is verified to issue a digital signature token, and the instruction is allowed to be sent to the physical device; otherwise, it is determined to fail, and a list of inconsistencies and their corresponding truth entries and evidence chain summaries are output to provide a basis for subsequent anomaly classification detection and dynamic blocking.

[0050] To facilitate the controller's handling of anomalies, and considering potential cognitive biases in large-scale intelligent agents during long-term power grid tasks, anomalies are categorized into three types: Category 1: Task Failure & Safety Violation Identify fatal logical fallacies in intelligent agents that "violate common sense in physics" or "violate safety procedures," including reasoning errors (such as drawing counterintuitive conclusions based on incorrect sensor readings, for example, inferring severe icing when micrometeorological data shows normal temperature) and execution errors (such as forcibly scheduling drones to take off under extreme weather conditions such as typhoons).

[0051] By comparing thought chains With environmental truth value This directly falsifies the reasoning premises of the intelligent agent, preventing accidental successes.

[0052] Category 2: Process Inefficiency and Loops Identify "cognitive stagnation" or "resource waste" behaviors of intelligent agents in the reasoning space, and detect redundant operations (e.g., having already obtained high-definition images to confirm the target, but still scheduling the robot to repeat on-site observation) and logical dead loops (e.g., repeatedly requesting invalid / offline sensor data without triggering the timeout exit mechanism).

[0053] By auditing the historical state sequence of the trajectory, it can be determined whether the decision-making process has fallen into a lossless oscillation or cycle, thereby avoiding the ineffective consumption of computing power and electricity resources.

[0054] Category 3: Unwarranted Continuation Identify "inertial loss of control" or "over-boundary execution" of intelligent agents in the task-terminated state, including execution even when the target is lost (e.g., the intelligent agent mechanically executes the expulsion process even though the target has been detected to have left) and inability to refuse (e.g., failing to report an error when faced with instructions that are beyond its capabilities, and instead generating a series of invalid attempts).

[0055] By evaluating the match between the task objective (Goal) and the current environmental state (State) in real time, subsequent operations that are not practically meaningful are forcibly interrupted.

[0056] After the audit is completed, the audit results will be sent to the intervention controller. The audit results include the trajectory. Digital signature tokens.

[0057] III. Intervention Controller: Dynamic Blocking & Rollback System. This is a "safety valve" connecting the decision-making level and physical equipment, and a "digital logic firewall" before deploying physical commands. It deploys a mandatory logical interception barrier between the decision-making intelligent agent and the physical execution mechanism (such as circuit breakers and transformer taps).

[0058] The firewall employs a "zero-trust" mechanism, which allows the output trajectory of the decision-making agent to be converted into a physical control signal only after the output trajectory of the decision-making agent has passed the consistency check based on the environmental truth value and obtained a digital execution token issued by the verification agent.

[0059] For any probabilistic output that fails truth verification, the firewall will implement deterministic physical blocking to ensure that the power grid control system retains the flexibility of large models in handling complex and unknown operating conditions while strictly meeting the deterministic security constraints of industrial-grade control. At the same time, it will activate the virtual rollback mechanism to reset the state of the decision-making agent to the node before the error occurred and inject correction prompts.

[0060] Specifically, addressing "cognitive drift" and logical degradation in long-term tasks, an adaptive closed-loop error correction mechanism based on "dynamic blocking-virtual rollback-guided retry" is proposed. To solve the problems of attentional divergence, target loss, or getting stuck in infinite loops that easily occur when agents execute long-chain complex tasks, and to endow the system with "self-repair" capabilities without human intervention, this embodiment establishes an adaptive closed-loop error correction mechanism based on virtual rollback. This mechanism is not a simple "error-reporting stop," but rather dynamically adjusts the agent's cognitive state to enable it to recover from errors, specifically including the following three core steps: (1) Establish a second-level "logic fuse" and dynamic blocking mechanism: Before physical commands are issued, a real-time intervention barrier (Logic Fuse) is constructed. Once the digital signature is verified and the verifying agent detects an abnormal trajectory, the logic fuse immediately triggers a dynamic blocking mechanism to intercept the current physical action command (such as intercepting a drone takeoff signal or a circuit breaker tripping signal), controlling the risk in the digital space. Specifically: Blocking "mission failure / safety violation": When an instruction that violates physical common sense or safety procedures is detected (such as "takeoff due to excessive wind speed"), the system directly intercepts the physical action to prevent irreversible equipment damage or safety accidents.

[0061] To prevent "inefficient processes / infinite loops": When the system detects that the agent is stuck in repetitive operations or invalid interactions, it forcibly pauses the current decision-making thread, breaks the agent's cognitive deadlock, and prevents the continuous waste of computing power and time resources.

[0062] Interruption of "unnecessary continuation": When it is determined that the task objective has been achieved or is no longer achievable (such as the objective being lost), but the agent continues to perform subsequent actions out of inertia, the system immediately terminates the task flow to avoid unnecessary interference to the environment.

[0063] (2) Implement the “Cognitive Rollback” strategy: An innovative "soft reset" method is proposed, operating solely in the digital cognitive space, to replace traditional physical device resets. Upon triggering a blockage, the system initiates an adaptive repair process, guiding the decision-making agent to correct errors, including: Memory state reset does not require the physical device to return to its initial position. Instead, it directly manipulates the decision agent's context memory, rewinding its state pointer to the moment before the error occurred. loc-1 ), clearing short-term memory fragments that cause errors, that is, verifying the specific steps involved in the agent's output anomaly index L loc Instead of physically resetting, the memory state of the decision-making agent is reverted to the moment before the error occurred. loc-1 .

[0064] Physical compensation, only when physical actions have been partially performed due to inertia (such as when a drone has taken off the ground), coordinates the execution of necessary "safe hovering" or "attitude reset" commands to ensure that the physical state remains synchronized with the cognitive state after the rollback.

[0065] (3) Construct a bootstrap retry mechanism based on "Critique Injection": To prevent the agent from retracing the same erroneous path after a rollback, the verification agent generates structured correction prompts based on the exception type and injects them into the decision agent's context window, guiding it to replan its path, including: Correcting "cognitive biases": For mission failure anomalies, inject prompts containing environmental truth values ​​(e.g., "The actual wind speed is 18 m / s, which does not meet the takeoff conditions. Please find an alternative solution") to correct the agent's illusionary premise.

[0066] Breaking "Logical Deadlock": For process inefficiency anomalies, inject hints containing historical operation patterns (e.g., "Operational oscillation detected, line L2 has impact load, please try to plan line L3") to force the agent to escape the local optimum trap and explore new solution space.

[0067] Terminate "Invalid Continuation": For non-essential continuation exceptions, inject explicit status prompts such as "Task objective has been updated" or "Objective unreachable" to guide the agent to end the task normally or switch to standby mode.

[0068] Finally, based on the rolled-back state and correction prompts, the decision-making agent re-infers and generates a new compliant trajectory T'.

[0069] For compliance instructions that pass the audit, they are issued to the collaborative execution layer to complete the physical operations and form a closed loop, specifically as follows: (1) Coordinated actions: Inspection robots, drones and various online governance devices (such as DC de-icing devices and sound and light de-icing devices) execute verified instructions.

[0070] (2) State closed loop: The execution result is fed back to the perception layer, the environmental state is updated, and the next round of perception and decision-making cycle is entered until the task goal is completely achieved.

[0071] Example 2 One embodiment of the present invention provides a runtime process trust verification method for power grid intelligent agents, comprising: Step S1: Based on the received power grid dispatch instructions and multimodal sensing data, generate the full-link governance trajectory of the smart grid; Step S2: Based on multimodal perception data that is heterogeneous with the decision-making intelligent agent, perform real-time runtime auditing on the generated end-to-end governance trajectory, detect trajectory anomalies, and generate a digital signature token for trusted verification of the runtime process. Step S3: Based on the digital signature token, execute the end-to-end governance process through real-time auditing to complete the physical operation and form a closed loop; In real-time auditing, if the detected trajectory does not conform to the objective facts of the physical world, the dynamic blocking of the runtime process is immediately triggered, and the virtual rollback mechanism is activated.

[0072] Taking local power grid switching operations—solving process inefficiencies—as a specific implementation scenario, this paper provides a concrete example. In a power grid self-healing scenario, it detects and breaks down "load oscillations" and "logic deadlocks" caused by short-sighted decision-making by the intelligent agent, ensuring the efficiency of switching operations. Figure 2 , Figure 3 As shown, the specific steps are as follows: Step 1: Multimodal Heterogeneous Perception and Ground Truth Construction This method not only focuses on instantaneous cross-sectional data, but also on the "full-dimensional load truth" constructed by the validation agent, which includes temporal features, including: (1) Full-dimensional data acquisition: The end-side SCADA system transmits in real time the load rate of 10kV line L1 is 98% (close to overload) and the current load rate of line L2 is 30% (light load).

[0073] (2) Edge feature mining (key): Verify that the agent does not only look at the current "light load" appearance, but independently retrieves the historical load curves of edge nodes.

[0074] (3) Truth model generation: Constructing the environmental truth G t This reveals a hidden fact: Line L2 connects to a large electric vehicle charging station and exhibits strong characteristics of "intermittent impact load" (multiple instantaneous changes in 0% to 80% occurred within the past hour). Furthermore, the operation log shows that the system had just performed a "switch back from L2 to L1" operation 3 minutes prior.

[0075] Step 2: The decision-making agent generates the end-to-end governance trajectory. The decision-making agent (Governor Agent) is limited by short-term memory or a single time-slice perspective, resulting in short-sighted governance strategies, specifically: (1) Task reception: Instruction I = “Relieve the risk of overload on line L1 and perform load transfer”.

[0076] (2) Reasoning by thought chain r t The agent reasoned that "L1 is currently overloaded, and scanning nearby lines reveals that L2's load is only 30%, indicating it has the capacity to transfer power. The optimal solution is to transfer the load downstream of L1 to L2." Here, the agent falls into the local trap of seeking the 'optimal at the current moment' and ignores the risk of historical oscillations.

[0077] (3) Action planning a tGenerate the instruction: "Switch_S1.close(); Switch_K1.open()" (Close the tie switch S1 and open the sectionalizing switch K1).

[0078] (4) Trajectory Submission: Generate trajectory T and submit it for auditing.

[0079] Step 3: Runtime Logic Causal Auditing Based on Agent Trajectory Anomaly Detection The Verifier Agent performs historical consistency audits on the trajectory, including: (1) Consistency check: Compare the trajectory T with the environmental truth G t The comparison was performed between the "Historical Operation Log" and the "L2 Impact Load Characteristics" in the database, specifically as follows: 1) Extraction and structuring of inference premises: from the inference chain r of trajectory T t Extracting from the middle to form an assertion set P t At least including: Assertion p1: Line L1 is currently close to overload (e.g., 98%), and there is an overload risk; Assert p2: Line L2 currently has a low load factor (e.g., 30%) and has a capacity for power transfer; Assertion p3 (implicit premise): The load capacity of line L2 is "stable and available" within the decision time window, and transferring the load to L2 will not trigger secondary over-limit or back-cut requirements in the short period of time; Assertion p4: Action to be performed a t ={Switch_S1.close();Switch_K1.open()} can effectively alleviate L1 overload and will not cause operational oscillation.

[0080] 2) From the environmental truth value G t Retrieve the corresponding truth entry in G: Verify the agent's position in G. t The search for the truth subset corresponding to the above assertions includes at least: From G t grid Read the current load rate and topology connectivity of L1 and L2 (used to verify the instantaneous facts of p1 and p2). From G t evid Read historical operation logs / SOE events (e.g., past τ) log Within minutes, S1, K1, and L1 L2 transfer-related operation records, timestamps, and directions); From G t risk or G t grid Read L2 from historical sequence evidence in the past τhist Load curve segments within a time window (e.g., the past hour) are used to calculate the "impact load characteristics"; short-term forecasts (e.g., future τ) are read when necessary. pred (Probability or range of load surge within minutes).

[0081] 3) Instantaneous state consistency comparison (verifying whether the "surface reason" is valid): The verification agent will assert that p1, p2 and G t grid The current state snapshot is used for numerical consistency comparison: for example, to check whether the L1 load rate is indeed 98%, the L2 load rate is indeed 30%, and whether the current topology allows the transfer path to be realized through S1 and K1. If this comparison passes, it only indicates that "the instantaneous fact is true", and does not directly determine that the plan is reasonable.

[0082] 4) Consistency comparison of "L2 impact load characteristics" (verify whether "implicit premise p3" is true): The verification agent calculates the impact load index based on the L2 historical load sequence, and judges whether the implicit premise of the decision agent regarding "L2 stability and availability" is consistent with the true value. Specifically, this includes: Calculation of impact characteristic quantities from L2 historical load curves (example of implementation method): Number of load transitions N spike : In τ hist The number of abrupt changes in load from low to high within the window; Maximum transition amplitude ΔP max The maximum magnitude of a single jump (e.g., 0% → 80%). Fluctuation intensity CV: The coefficient of variation or peak value of the rate of change per unit time of the load sequence; The above characteristic values ​​are compared with preset thresholds to obtain the "impact load label" or "stability level" (e.g., if N...). spike With ΔP max If the threshold is exceeded, L2 is determined to have significant impact load characteristics.

[0083] Consistency determination rule (scenario-based): If the decision-making agent's reasoning is based solely on "current load 30%" without introducing any stability constraints / historical fluctuation evidence, then its implicit premise p3 and G... t The "L2 impact load characteristics" are inconsistent (due to insufficient premises / contradiction with the true value), and the inconsistent item is output: Example of an inconsistency: "Treat L2 as a stable bearer line" Truth: "L2 may experience charging station load surges, potentially reaching high load levels within a short period."

[0084] 5) Consistency comparison of "Historical Operation Logs" (verifying whether it will cause operational fluctuations): Verify the agent in τ log Search the operation log within the time window to identify whether there are operation patterns that are opposite to or repetitive with the currently intended action, specifically including: Historical event E was retrieved. t Δ For example, a "switch back from L2 to L1" operation occurred 3 minutes ago (including timestamp, involving switches, and direction); The currently planned action a t Pattern matching with historical event directions: If there is a combination of "short interval reverse cut + repeated transfer", it is determined that there is a risk of operational oscillation; Consistency Judgment Rule (Contextualized): When this oscillation pattern is detected, the judgment assertion p4 ("This action can effectively alleviate the oscillation and will not cause oscillation") is consistent with G. t evid The historical evidence is inconsistent, and the inconsistent items are output: Example of an inconsistency: "The current transfer to L2 is valid and stable." Truth value: "A reverse cutback has occurred within the last 3 minutes; repeating this will result in L1..." L2 reciprocating switching.

[0085] 6) Formulate constraint conclusions and provide trajectory-level judgments: The verification agent writes the inconsistencies obtained in steps 4) and 5) into G. t cons The constraint derivation results (e.g., marking L2 as a candidate line that is "short-term unavailable / requires additional condition verification", or setting "τ") osc The procedure constraint of "prohibiting repeated supply transfers within the time window" is used to form a trajectory-level consistency conclusion for trajectory T. If either "insufficient / contradictory premises caused by impact load characteristics" or "evidence of short-cycle reciprocating operation" is met, trajectory T will fail the consistency check, and this inconsistency result will be used as the direct basis for "Type 2: Process Inefficiency & Loops" in the subsequent anomaly classification, which will be used to trigger logical circuit breaking and dynamic blocking.

[0086] (2) Anomaly classification and detection: The judgment result is "Trajectory Anomaly Type 2: Process Inefficiency and Infinite Loop", and the audit logic is as follows: Although the "transfer to L2" is currently compliant in terms of electrical parameters, historical data shows that a reverse operation occurred just 3 minutes ago. If this instruction is executed, the system will fall into "oscillation" of L1→L2→L1. This is an invalid reciprocating operation, wasting switch life and threatening grid stability.

[0087] Step 4: Logical circuit breaking and dynamic blocking before physical commands (1) Blocking trigger: The logic fuse recognizes the “process inefficient” label and immediately blocks the physical operation commands for switches S1 and K1.

[0088] (2) Risk control: The physical switch remains in place, avoiding unnecessary arc impact and mechanical wear of equipment, and preventing the power grid from falling into an unstable oscillation state.

[0089] Step 5: Cognitive Rollback and Boot Error Correction Initiating adaptive repair breaks the logical deadlock of the agent, specifically as follows: (1) Error location: The problem was located in the "target route selection" step of the decision-making logic. loc .

[0090] (2) Virtual rollback: The memory state of the decision agent is rolled back to the moment when "L1 overload has been detected" but "no transfer path has been selected".

[0091] (3) Critique Injection: The agent injection includes a historical perspective correction criterion, such as, “Criticism: A potential risk of oscillation operation has been detected. Line L2 has intermittent impact load (charging station) and a back-cut occurred 3 minutes ago, which is an unstable power source.”

[0092] Guidance: Mark line L2 as 'unavailable'. Please recalculate the topology, assess the carrying capacity of line L3, or consider initiating demand-side response (DR). (4) Path replanning: After receiving the prompt, the decision agent abandons L2 and searches for a new path: "Exclude L2. Calculate that the current load of line L3 is 50% and the load curve is stable. Generate a new plan: transfer the power to line L3." Step Six: Collaborative Execution and Closed-Loop Feedback (1) Coordinated action: The new trajectory T (transfer to L3) passed the truth check (L3 has no impact characteristics and no oscillation history), and the system issued an execution token. The smart switch was activated to complete the transfer.

[0093] (2) State closed loop: The load rate of line L1 dropped to 75%, while the load rate of line L3 rose to 70%, and the power grid operated smoothly.

[0094] The system successfully broke the logical loop and achieved truly effective overload elimination.

[0095] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned runtime process trust verification method for a power grid intelligent agent.

[0096] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions. When the computer instructions are executed by a processor, they implement the aforementioned runtime process trust verification method for a power grid intelligent agent.

[0097] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a runtime process trusted verification method for a power grid intelligent agent.

[0098] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A runtime process trust verification system for power grid intelligent agents, characterized in that, include: The decision-making agent is used to generate the full-link governance trajectory of the smart grid based on the received power grid dispatch instructions and multimodal sensing data, and send it to the verification agent. The verification agent is used to perform real-time runtime auditing of the generated end-to-end governance trajectory based on multimodal perception data that is heterogeneous with the decision-making agent, detect trajectory anomalies, generate a digital signature token for trusted verification of the runtime process, and send it along with the end-to-end governance trajectory to the execution controller. In real-time auditing, if the detected trajectory does not conform to the objective facts of the physical world, the dynamic blocking of the runtime process is immediately triggered, and the virtual rollback mechanism is activated. The execution controller is used to distribute the end-to-end governance trajectory, based on digital signature tokens and real-time auditing, to the collaborative execution layer to complete physical operations and form a closed loop.

2. The runtime process trust verification system for power grid intelligent agents as described in claim 1, characterized in that, The multimodal perception data of the decision-making agent includes power grid operation and equipment status data, power grid model and rule data, task execution side perception data, and historical and contextual data; The multimodal perception data of the verification agent includes redundant measurement data, multimodal physical perception data, device event and timing evidence data, and physical consistency results.

3. The runtime process trust verification system for power grid intelligent agents as described in claim 1, characterized in that, The end-to-end governance trajectory is defined as follows: Where I represents the power grid dispatch instruction. For the thought chain at step t, For the action to be performed in step t, This is the environmental feedback for step t.

4. The runtime process trust verification system for power grid intelligent agents as described in claim 3, characterized in that, By verifying the intelligent agent to perform real-time runtime auditing of the generated end-to-end governance trajectory, trajectory anomalies are detected, specifically: Collect independently perceived full-dimensional data; Multimodal fusion is performed on the collected full-dimensional data; Run a lightweight time-series evolution model to predict the development trend of potential hazards within future time windows; Based on the above independent perception and inference results, the true environmental value at the current moment is generated. ; By comparing thought chains With environmental truth value This directly falsifies the reasoning premises of the intelligent agent.

5. The runtime process trust verification system for power grid intelligent agents as described in claim 1, characterized in that, The dynamic blocking refers to intercepting the current physical action commands in the entire governance trajectory, thus controlling the risk in the digital space.

6. The runtime process trust verification system for power grid intelligent agents as described in claim 1, characterized in that, The activation of the virtual rollback mechanism resets the state of the decision-making agent to the node before the anomaly occurred and injects a correction prompt.

7. A method for verifying the reliability of runtime processes of a power grid intelligent agent, characterized in that, Based on the runtime process trust verification system as described in any one of claims 1-6, the specific steps are as follows: Based on the received power grid dispatch instructions and multimodal sensing data, a full-link governance trajectory for the smart grid is generated; Based on multimodal perception data that is heterogeneous with the decision-making intelligent agent, the generated end-to-end governance trajectory is audited in real time during runtime, trajectory anomalies are detected, and a digital signature token for trusted verification of the runtime process is generated. Based on digital signature tokens, the entire governance process is executed through real-time auditing, completing physical operations and forming a closed loop; In real-time auditing, if the detected trajectory does not conform to the objective facts of the physical world, the dynamic blocking of the runtime process is immediately triggered, and the virtual rollback mechanism is activated.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the runtime process trust verification method for a power grid intelligent agent as described in claim 7.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the runtime process trusted verification method for a power grid intelligent agent as described in claim 7.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a runtime process trusted verification method for a power grid intelligent agent as described in claim 7.

Images