Intelligent power grid dispatching fault processing system and method based on intelligent agent self-adaptive arrangement

Abstract

The invention discloses a power grid dispatching fault intelligent processing system based on intelligent agent adaptive arrangement, and the system comprises an alarm clustering module which is used for carrying out the clustering processing of an alarm sample of a power dispatching system through employing a clustering algorithm, and obtaining an alarm cluster set; the fault event judgment module is used for inputting the alarm clusters in the alarm cluster set into the fault detection model and outputting a judgment result about whether the alarm events corresponding to the alarm clusters are fault events or not; the root cause positioning module is used for positioning equipment where a root cause of the fault event is located; and the fault processing module is used for adopting a guide type reward strategy optimization type algorithm in all fault processing strategies obtained through intelligent agent arrangement based on the attribute of the equipment where the root cause of the fault event is located and the running state of the equipment. And finding out the fault processing strategy with the highest reward score in the algorithm of the guide type reward strategy optimization type. According to the method, the manual intervention requirement is reduced, the dependence on advanced operation and maintenance personnel is reduced, and the labor cost is remarkably saved.

Description

Technical Field

[0001] This invention relates to the field of power grid automated dispatching technology, specifically to a power grid dispatching fault intelligent processing system and method based on agent adaptive orchestration. Background Technology

[0002] Power grid dispatch automation systems are crucial for the safe, economical, and reliable operation of power systems. They collect, transmit, and process various monitoring and control data, such as SCADA, EMS, and DMS, to achieve real-time monitoring and control of the power grid status. However, with the continuous expansion and increasing complexity of power systems, the operating environment of dispatch automation systems is becoming increasingly diverse, and the types of abnormal alarms and faults encountered during operation and maintenance have also increased significantly, such as front-end thread blocking, communication link interruption, database deadlock, and host CPU overload.

[0003] Currently, the main problems with the operation and maintenance alarm handling of the dispatch automation system are as follows: 1) When the system experiences an anomaly, it generates a large number of repetitive, invalid or noisy alarms, making it difficult for operation and maintenance personnel to locate critical faults in a timely manner; 2) Existing alarm handling mainly relies on manual intervention or preset rules, resulting in slow response speed and easy errors due to insufficient experience of operation and maintenance personnel; 3) Traditional automated handling solutions are difficult to continuously optimize in complex environments and cannot achieve dynamic evolution and intelligent optimization of strategies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a power grid dispatching fault intelligent processing system based on agent adaptive orchestration, comprising: The alarm clustering module is used to cluster all alarm samples of the power dispatching system within the current time window using a clustering algorithm to obtain an alarm cluster set. An alarm cluster is a set of all alarm samples of alarm events caused by the same root cause. The fault event judgment module is used to input the alarm clusters in the alarm cluster set into the fault detection model to judge whether the alarm event corresponding to the alarm cluster is a fault event, and output the judgment result of whether the alarm event corresponding to the alarm cluster is a fault event; The root cause localization module is used to perform causal reasoning on alarm clusters whose corresponding alarm events are fault events in the judgment results, and to locate the device where the root cause of the fault event is located. The fault handling module defines the sequence obtained by orchestrating all agents in the power dispatching system for handling the fault event as a fault handling strategy. Among all fault handling strategies obtained by agent orchestration, based on the equipment attributes of the root cause of the fault event and the operating status of the equipment, a guided reward strategy optimization algorithm is used to find the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms, and the fault handling strategy with the highest reward score is used to handle the corresponding fault event.

[0005] Furthermore, in the alarm clustering module, the specific method for using a clustering algorithm to cluster all alarm samples of the power dispatching system within the current time window to obtain an alarm cluster set is as follows: Building an alarm sample set ; The alarm sample set was clustered using the DBSCAN density clustering algorithm to obtain an alarm cluster set. In the formula, For the first A cluster of alarms, For the first The first in the alarm cluster Alarm sample The alarm samples are arranged in chronological order.

[0006] Furthermore, the fault event judgment module inputs the alarm clusters from the alarm cluster set into the fault detection model to judge whether the alarm event corresponding to the alarm cluster is a fault event, and outputs the judgment result of whether the alarm event corresponding to the alarm cluster is a fault event. The specific method is as follows: The fault detection model was obtained by training an LSTM-GAN model. alarm cluster Reconstruction loss function in input fault detection model In the formula, For alarm clusters The generator output; alarm cluster Discriminant loss function in input fault detection model In the formula, For alarm clusters The discriminator output; Calculate the anomaly score in the fault detection model In the formula, These are the weighting coefficients; If abnormal scores If the score exceeds the threshold, an alarm cluster will be triggered. The corresponding event is a fault event, if the abnormal score If the score is less than the threshold, an alarm cluster will be triggered. The corresponding event is a normal event.

[0007] Furthermore, the root cause localization module performs causal reasoning on the alarm clusters whose corresponding alarm events are fault events in the judgment results, and the specific method for locating the device where the root cause of the fault event is located is as follows: Construct a Bayesian network, whose nodes include device nodes, device operation status nodes, and alarm nodes; Given a series of alarm nodes, calculate the conditional probability of each node, calculate the posterior probability of all device nodes to be inferred based on the conditional probability of each node, and select the device node with the highest posterior probability as the node where the root cause is located.

[0008] Furthermore, in the fault handling module, the specific method for defining the sequence obtained by orchestrating all agents in the power dispatching system used to handle the fault event as the fault handling strategy is as follows: Fault handling strategy represents an execution plan in which multiple agents work collaboratively in a specific order, denoted as: In the formula, For the k-th fault handling strategy, This represents the j-th agent in the k-th fault handling strategy.

[0009] Furthermore, in the fault handling module: The guided reward strategy optimization algorithm adopts the GRPO algorithm. Based on the reliability of the fault handling strategy, the resource cost of the fault handling strategy, and the time delay of the fault handling strategy, the reward function used in the GRPO algorithm to calculate the reward score of the fault handling strategy is constructed.

[0010] Furthermore, fault handling strategies Execution reliability Fault handling strategy The product of the success probabilities of all agents in the system handling this type of fault; Fault handling strategy Resource costs of execution Fault handling strategy The sum of the expected resource costs consumed by all agents during execution; Fault handling strategy Execution time delay Fault handling strategy The sum of the expected time delays consumed by all agents during execution. The reward function used to calculate the reward score for the fault handling strategy is the fault handling strategy. The function is obtained by subtracting resource costs and time delays from the reliability, and then balancing the importance of reliability, resource costs, and time delays through weighting coefficients.

[0011] Furthermore, in the reward function construction module, among all fault handling strategies obtained through agent orchestration, based on the device attributes of the root cause of the fault event and the device's operating state, a guided reward strategy optimization algorithm is used to find the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms. The specific method is as follows: The fault handling strategy network in GRPO is denoted as , For strategy parameters, This is a set of operating environment information and fault context features of the power grid automated dispatching system at time t. Fault handling strategy The corresponding intelligent agent in the fault handling strategy network The specific action decisions obtained from the sampling, and the fault handling strategy network Output in the current Select each fault handling strategy below The probability distribution of , where; In the formula, It is a non-linear activation function. This is the attribute feature vector of the device that is the root cause of the fault event. This is the feature vector of the operating state of the equipment at the root cause of the fault event. For bias terms, This is a weight matrix representing the attributes of the device that is the root cause of the failure event. The weight matrix represents the operating state of the equipment at which the root cause of the failure event is located; Calculate the advantage function : In the formula, For the reward function, For state-value functions, a state-value function is defined as a function of state... The following is based on the fault handling strategy. Expected cumulative return from discounts starting at time t The calculation formula is as follows: In the formula, Fault handling strategy At any moment Instant rewards received Discount factor; The GRPO optimization objective function is constructed as follows: in, The cropping threshold, To guide the fault handling strategy, The KL divergence between the current fault handling strategy and the guided fault handling strategy is given. The regularization intensity coefficient is . These are the strategy parameters from the previous iteration; The strategy parameters are iteratively updated by optimizing the objective function using GRPO. until the reward function Convergence occurs, at which point the fault handling strategy network... The output is in the current Select each fault handling strategy below In the probability distribution, the fault handling strategy with the highest probability is the fault handling strategy with the highest reward score.

[0012] A power grid dispatching fault intelligent handling method based on agent adaptive orchestration includes: Clustering algorithms are used to cluster all alarm samples of the power dispatching system within the current time window to obtain an alarm cluster set. An alarm cluster is a set of all alarm samples of alarm events caused by the same root cause. Input the alarm clusters in the alarm cluster set into the fault detection model to determine whether the alarm events corresponding to the alarm clusters are fault events, and output the determination result of whether the alarm events corresponding to the alarm clusters are fault events; For alarm clusters whose corresponding alarm events are fault events in the judgment results, perform causal reasoning to locate the equipment where the root cause of the fault event is located; The sequence obtained by orchestrating all agents in the power dispatching system for handling the fault event is defined as the fault handling strategy. Among all fault handling strategies obtained by agent orchestration, based on the equipment attributes of the root cause of the fault event and the operating status of the equipment, a guided reward strategy optimization algorithm is adopted to find the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms, and the fault handling strategy with the highest reward score is adopted to handle the corresponding fault event.

[0013] A computer program product includes a computer program / instructions that, when executed by a processor, implement the aforementioned intelligent fault handling method for power grid dispatching based on agent adaptive orchestration.

[0014] The beneficial effects of this invention are as follows: 1. Traditional operation and maintenance relies on manual analysis of alarms and formulation of handling plans, resulting in long response delays and a high risk of errors. This invention achieves fault handling within minutes or even seconds through full-process automation. The DBSCAN algorithm automatically filters redundant alarms, compressing thousands of alarms into a small number of event clusters, reducing manual screening time. The root cause localization module based on Bayesian networks can pinpoint the fault source within seconds, avoiding the time-consuming process of traditional line-by-line troubleshooting. The GRPO algorithm adaptively generates optimal strategies based on the real-time power grid dispatch automation system status, replacing manual experience-based decision-making.

[0015] 2. By integrating multiple artificial intelligence algorithms, the false alarm rate and false negative rate of faults are significantly reduced. The LSTM-GAN model utilizes generative adversarial networks to model normal behavior, is sensitive to small deviations, and improves the accuracy of fault prediction. Bayesian networks integrate device status, topology relationships, and alarm information for reasoning, resulting in high root cause localization accuracy.

[0016] 3. The GRPO algorithm automatically selects the most cost-effective strategy through a reward function, balancing reliability, cost, and efficiency to achieve economical operation and maintenance. For example, it prioritizes "resource optimization" over "hardware restart" to reduce business interruption losses. This reduces the need for manual intervention, lowers reliance on senior operation and maintenance personnel, and significantly saves on labor costs.

[0017] 4. The GRPO algorithm optimizes the policy network based on the handling results, making the system increasingly intelligent with use. Bayesian network parameters can be automatically corrected using historical data to adapt to changes in the power grid topology. A well-trained LSTM-GAN model can quickly adapt to new connected devices, reducing the cost of repetitive modeling. Attached Figure Description

[0018] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0019] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0020] Glossary Power Grid Automated Dispatch System: A power grid automated dispatch system is a complex set of hardware and software systems that utilizes computer technology, communication technology, automatic control technology, and sensor measurement technology to monitor, coordinate, control, make decisions, and manage the entire process of power generation, transmission, distribution, and consumption in real time. Its ultimate goal is to ensure the safe and stable operation of the power grid, optimize resource allocation, improve power supply reliability, and enhance operational efficiency.

[0021] Intelligent Agent: In agent-based adaptive orchestration scheduling automation systems, an intelligent agent refers to an autonomous software module that encapsulates specific fault handling functions and possesses perception, analysis, decision-making, and execution capabilities. Each intelligent agent is a "digital expert" specifically designed to handle a particular type of power grid operation and maintenance task. For example, a communication fault handling intelligent agent is specifically designed to diagnose and handle faults such as network link interruption, message loss, and communication timeout. Another example is a database optimization intelligent agent, which handles faults such as database deadlock, connection pool exhaustion, and slow response.

[0022] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a classic density-based clustering algorithm that effectively discovers clusters of arbitrary shapes and identifies noise points. DBSCAN clusters samples by analyzing their density distribution in the feature space. Its core idea is that clusters are composed of high-density regions, and samples in low-density regions are considered noise. Unlike distance-based clustering algorithms (such as K-Means), DBSCAN does not rely on a preset number of clusters; instead, it automatically forms cluster structures based on the density relationships of samples within their neighborhoods.

[0023] LSTM-GAN Model: LSTM-GAN is a hybrid generative deep learning model designed for anomaly detection and prediction in time series data. Its name comes from its two core components: LSTM (Long Short-Term Memory), a special type of recurrent neural network adept at capturing long-term dependencies in time series data; and GAN (Generative Adversarial Network), a framework that learns through an adversarial game between a generator and a discriminator. Therefore, LSTM-GAN can be defined as: a generative adversarial network built around an LSTM network, used to learn the normal distribution pattern of time series data and thereby detect anomalies deviating from this pattern or predict future states. The LSTM-GAN model contains two sub-networks trained together through an adversarial process: a generator and a discriminator. The generator aims to learn the distribution patterns of normal time series data, receiving random noise or the original sequence as input, and attempting to generate time series data that is as realistic as possible. The generator is composed of LSTM layers, which enables it to effectively understand and memorize dynamic patterns and long-term dependencies in time series data, thus generating realistic data with reasonable temporal relationships. The goal of a discriminator is to distinguish whether the input data comes from a real, normal time series or from fake data generated by a generator. Discriminators are typically composed of LSTM or CNN layers, used to extract features from the input sequence and make a judgment on whether the data is real or fake.

[0024] The GRPO (Guided Reward Policy Optimization) algorithm is an advanced reinforcement learning algorithm that cleverly integrates data-driven reinforcement learning with knowledge-driven expert experience to solve policy optimization problems in complex environments. GRPO belongs to the policy gradient algorithm family and is a method in the field of deep reinforcement learning. It is an improvement on the PPO (Proximal Policy Optimization) algorithm, and its core innovation lies in introducing a guiding policy as a constraint in the optimization process. Traditional reinforcement learning algorithms (such as PPO) often explore complex environments like "blind men feeling an elephant," requiring a large number of trial and error to find a good policy, which is inefficient and may lead to dangerous operations. The core idea of ​​GRPO is to equip this "exploration" process with an "expert mentor." During the optimization process, the algorithm not only pursues high rewards but also constantly refers to the expert policy (guiding policy) to ensure that the new policy does not deviate too far from expert experience. This achieves guided and safe exploration. Its core goal is to efficiently find the optimal policy that obtains the highest overall reward while ensuring the safety and reliability of policy optimization.

[0025] Example 1 A power grid dispatching fault intelligent processing system based on agent adaptive orchestration includes: The alarm clustering module is used to cluster all alarm samples of the power dispatching system within the current time window using a clustering algorithm to obtain an alarm cluster set. An alarm cluster is a set of all alarm samples of alarm events caused by the same root cause. The fault event judgment module is used to input the alarm clusters in the alarm cluster set into the fault detection model to judge whether the alarm event corresponding to the alarm cluster is a fault event, and output the judgment result of whether the alarm event corresponding to the alarm cluster is a fault event; The root cause localization module is used to perform causal reasoning on alarm clusters whose corresponding alarm events are fault events in the judgment results, and to locate the device where the root cause of the fault event is located. The fault handling module defines the sequence obtained by orchestrating all agents in the power dispatching system for handling the fault event as a fault handling strategy. Among all fault handling strategies obtained by agent orchestration, based on the equipment attributes of the root cause of the fault event and the operating status of the equipment, a guided reward strategy optimization algorithm is used to find the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms, and the fault handling strategy with the highest reward score is used to handle the corresponding fault event.

[0026] (1) Construct an alarm sample set , In the formula, For the first Alarm sample For the first The timestamp of the alarm sample To generate the first Devices that generate alarm samples (such as front-end servers). For the first The types of alarm samples (such as communication timeout, database deadlock, thread blocking, etc.). For the first The status of each alarm sample (reflecting the current lifecycle status of the alarm, indicating whether the alarm needs to be handled immediately, has been resolved, or has been confirmed) is raw alarm data collected in real time from multiple sources such as data acquisition and monitoring control module, energy management module, power distribution management module, front-end unit, communication link and database.

[0027] The alarm sample set was clustered using the DBSCAN density clustering algorithm. The specific process is as follows: Density calculation and neighborhood analysis: for each alarm sample Calculate its - Neighborhood ( The neighborhood radius defines the number of similar alarms within the neighborhood of a sample point. Similarity is based on a distance metric (such as Euclidean distance or a custom distance function) of alarm features (such as time, location, type). Determining core points, boundary points, and noise points: Core points: If ( The minimum number of points threshold represents the minimum number of samples required to form a high-density region (used to distinguish between high-density and low-density regions). The core point is marked as the center of a high-density region; boundary points: if ,but Located at a certain core point - If a sample is within the neighborhood of a core point, it is marked as a boundary point; Noise point: Samples that neither meet the core point condition nor are within the neighborhood of any core point are judged as noise points and are removed.

[0028] Cluster formation and expansion: Starting from any core point, recursively expand it through density reachability. - All samples within the neighborhood (including other core points and boundary points) are grouped into the same cluster until the cluster can no longer be expanded.

[0029] Obtain alarm cluster set In the formula, For the first A cluster of alarms, For the first The first in the alarm cluster Alarm sample The alarm samples in the data are arranged in chronological order. Each cluster It represents a set of alarm events that are highly correlated in time and characteristics, triggered by the same underlying cause.

[0030] DBSCAN can automatically identify and remove noise points (such as isolated alarms), retaining only alarm clusters in high-density areas. This significantly reduces the interference of invalid alarms on subsequent processing. It requires no preset cluster count and can automatically generate clusters of arbitrary shapes based on the inherent distribution of alarm data, adapting to the complexity and dynamism of power system alarms. Through the fusion of multiple features such as time, device, and type, it ensures that alarms with the same root cause are accurately aggregated, providing high-quality input for fault event judgment. This lays the foundation for data preprocessing, solves the problems of large amounts of repetitive and invalid alarms that are difficult to locate in traditional methods, and improves the processing efficiency of subsequent modules.

[0031] (2) The specific method for the fault event judgment module to input the alarm clusters in the alarm cluster set into the fault detection model to judge whether the alarm event corresponding to the alarm cluster is a fault event and output the judgment result of whether the alarm event corresponding to the alarm cluster is a fault event is as follows: The fault detection model is obtained by training an LSTM-GAN model. The training of LSTM-GAN is a dynamic game process consisting of a generator and a discriminator, as follows; The objective function for optimizing the training process is: In the formula, Discriminator The goal is to maximize the objective function , Represents generator The goal is to minimize the objective function , Discriminator The target, namely the discriminator The generator is considered The data provided comes from real data. The probability (between 0 and 1, the closer to 1 is the correct judgment, the closer to 0 is the wrong judgment). This indicates the distribution of real data. Real data from mid-sample Seeking expectations, This indicates that the discriminator is effective against real data. The output of the judgment is defined as follows: Real data is defined as a standard normal time-series sample formed after preprocessing historical operation and maintenance data generated by the power grid automation dispatching system under normal and healthy operating conditions. This real data originates from performance indicators, logs, and low-level events collected by multiple sources of equipment, including data acquisition and monitoring control modules, energy management modules, distribution management modules, front-end processors, communication links, and databases, during the fault-free operation of the power grid automation dispatching system. This data is constructed into standardized time-series samples. For example, it could be a sequence of indicators such as CPU utilization, memory usage, network traffic, and database query response time arranged in chronological order. These data samples do not contain any fault or abnormal alarms; they represent the normal behavioral patterns that the system is expected to learn and reproduce. Represents generator With discriminator The confrontation Indicates the distribution of prior noise Find the expectation of the random noise z sampled in the middle. This indicates that the generator will generate noise. The generated fake data through mapping This indicates that the discriminator is responsible for the generated data. The output of the judgment (i.e., the probability that the discriminator considers it to be real data); The training process is performed alternately, rather than simultaneously, as shown in the following loop: a) Training the discriminator D: Sample a batch of real time-series data from a real dataset. Sample a batch of data from random noise (or latent space). Input generator , to obtain generated data Using this batch of real data (Label 1) and data (Label 0) together serve as discriminators The training set. Calculate the loss function of the discriminator D (i.e., the loss function of the objective function mentioned above). The relevant parts are updated only, and the parameters of the discriminator D are updated (using gradient descent), with the goal of maximizing its classification accuracy.

[0032] b) Training generator G: Sample another batch of data from the random noise. Input generator , to obtain generated data .Will Input into the newly updated discriminator In the middle, obtain the discrimination result. Calculate the generator's loss function (i.e., the part of the objective function related to G mentioned above), and update only the parameters of the generator G, with the goal of making... Get as close to 1 as possible (i.e., deceive the discriminator).

[0033] This process is repeated iteratively until the quality of the data generated by the generator is good enough that the discriminator cannot effectively distinguish between them (i.e., the accuracy of the discriminator is close to 50%).

[0034] After training is completed, the alarm cluster will be activated. Reconstruction loss function in input fault detection model In the formula, For alarm clusters The generator output; alarm cluster Discriminant loss function in input fault detection model In the formula, For alarm clusters The discriminator output; Calculate the anomaly score in the fault detection model In the formula, These are the weighting coefficients; If abnormal scores If the score exceeds the threshold, an alarm cluster will be triggered. The corresponding event is a fault event, if the abnormal score If the score is less than the threshold, an alarm cluster will be triggered. The corresponding event is a normal event.

[0035] LSTM components can capture long-term dependencies in alarm sequences, while GAN adversarial training enables the model to learn normal pattern distributions, making it sensitive to small deviations and achieving early fault prediction. It requires only normal historical data for training and does not need labeled fault samples, making it suitable for power system scenarios with many fault types and few samples. Anomaly scores are calculated by weighting reconstruction loss and discriminative loss, taking into account both data distribution deviations and model confidence, thus reducing false alarm rates. It achieves intelligent determination from alarm clusters to fault events, solving the pain point of poor prediction accuracy in traditional methods and providing reliable input for root cause localization.

[0036] (3) The root cause localization module performs causal reasoning on the alarm clusters whose corresponding alarm events are fault events in the judgment results, and the specific method for locating the equipment where the root cause of the fault event is located is as follows: A Bayesian network is constructed, and its nodes include device nodes, device status nodes, and alarm nodes. Device nodes include front-end processors, communication links, databases, and protection devices. Device status nodes include CPU utilization, memory leaks, and lost heartbeats. Alarm nodes include communication timeouts and thread blocking. Device nodes often serve as potential root causes of faults and are the physical carriers of those faults (e.g., database server crashes). Device status nodes represent the abnormal operating states of devices, are the result of the root causes, and may also be the direct cause of alarms (e.g., database connection exhaustion). Alarm nodes are directly observable evidence in the power grid automation dispatching system and represent the final manifestation of the fault (e.g., a "database connection failed" alarm).

[0037] It is important to note the alarm sample. equipment in The alarm signal is the device or location that directly generates it, representing the point of failure. The root cause of the failure, on the other hand, refers to the device that initially malfunctions, triggering the entire chain reaction; it is the root cause of the failure.

[0038] Given a series of alarm nodes, calculate the conditional probability of each node, calculate the posterior probability of all device nodes to be inferred based on the conditional probability of each node, and select the device node with the highest posterior probability as the node where the root cause is located.

[0039] Calculate the conditional probability of each node: In the formula, For the first There are several nodes, with the current node representing the result. Therefore... It can be a status node or an alarm node. Example: If If it's a heartbeat loss, then it's a state node. If it's a communication timeout, then it's an alarm node; For the first Nodes The set of parent nodes; in causal reasoning, parent nodes usually represent causes, therefore... Often, these are device nodes or status nodes. For example, when calculating P(heartbeat lost | communication link interrupted), This means the communication link is interrupted (this is a device node). For a known set of parent nodes In the case of values, the first Nodes The probability of taking a certain value, set of parent nodes Take the total number of occurrences of a specific combination of values. For the first Nodes Take a specific value, and its parent node set Also takes a specific combination value, which is the number of times both of these cases occur simultaneously; In this system, device nodes typically serve as root or intermediate nodes. If a device node is the root node (without a parent node), its prior probability is used instead of its conditional probability calculation. Device status nodes act as a bridge between device nodes and alarm nodes; their parent nodes may include device nodes (e.g., "communication link interruption" leading to "heartbeat loss"). Alarm nodes are leaf nodes or observation nodes; their parent nodes are typically status nodes or device nodes (e.g., "heartbeat loss" leading to "communication timeout").

[0040] To calculate the posterior probability of a device node, the conditional probabilities of all the following nodes are needed for the following reasons: The conditional probabilities of alarm nodes are used to calculate the likelihood function, since these are observational evidence.

[0041] The conditional probability of the device operation status node is used to propagate evidence, as the device operation status node connects the device node and the alarm node.

[0042] For the conditional probability (or prior probability) of a device node, the device node is considered as a potential root cause, and its prior probability and conditional probability (if there is a parent node) are used to calculate the prior distribution.

[0043] The formula for root cause reasoning is: In the formula, The first one to be inferred Each device node acts as a root node in a Bayesian network. A root node is defined as a node without a parent node in a Bayesian network. The set of observed alarm nodes. This indicates traversing all device nodes to be inferred. Calculate the posterior probability of each root cause given the observed evidence O, and then select the root causes that result in a posterior probability of O. The device node with the largest value The equipment that caused the failure was located.

[0044] By leveraging conditional probability networks of devices, status, and alarm nodes, the most likely root cause device is inferred in reverse, avoiding misjudgments caused by reliance on human experience. Based on a probabilistic model, it can fuse multi-source evidence (such as alarm sets) and output the probability value of the root cause, supporting the interpretability of decisions. Bayesian network parameters can be updated through historical data to adapt to changes in system topology, improving the robustness of long-term applications. This solves the problem of difficult root cause location, elevating operations and maintenance from symptom analysis to essential diagnosis, providing precise targets for intelligent handling.

[0045] (4) In the fault handling module, the specific method for defining the sequence obtained by orchestrating all intelligent agents in the power dispatching system used to handle the fault event as the fault handling strategy is as follows: Fault handling strategy represents an execution plan in which multiple agents work collaboratively in a specific order, denoted as: In the formula, For the k-th fault handling strategy, This represents the j-th agent in the k-th fault handling strategy.

[0046] Building an intelligent agent pool Among them, intelligent agents For the first One fault handling unit; intelligent agent Attributes , This includes fault type, response time, supported device types, and execution permission level, which will affect the intelligent agent. Attributes Match the attributes of the fault event from the agent pool. Select a set of candidate agents .

[0047] By defining task flows through intelligent agent sequences (such as diagnosis, repair, and verification), the orderly and complete nature of handling actions is ensured. Each intelligent agent encapsulates specific functions (such as communication recovery and database optimization), supporting flexible combinations to adapt to diverse fault scenarios. This transforms the handling process from "manual intervention" to an "automated pipeline," improving response speed and standardization.

[0048] (5) In the fault handling module, the algorithm of the guided reward strategy optimization type adopts the GRPO algorithm. Based on the reliability of the fault handling strategy, the resource cost of the fault handling strategy and the time delay of the fault handling strategy, the reward function used to calculate the reward score of the fault handling strategy in the GRPO algorithm is constructed.

[0049] Fault handling strategy Execution reliability Fault handling strategy The reliability of the fault handling strategy execution is the product of the success probabilities of all agents in handling this type of fault. The calculation method is as follows: In the formula, Fault handling strategy The Middle A smart agent The probability of successfully handling this type of fault It is a fault handling strategy The number of agents included.

[0050] This can be obtained from historical execution records, including the number of successes and attempts. This is the success rate of a single agent based on historical data or expert experience, representing the probability of success when the agent performs a specific fault handling task alone.

[0051] For example, if the agent is a "database restart agent", the reliability of its fault handling strategy execution. A value of 0.95 indicates that the agent has a 95% success rate when performing database restart operations in the past. During fault handling, the reliability of each candidate strategy is calculated, and the optimal strategy is selected by combining other factors. For example, if a strategy involves 3 agents with success probabilities of 0.9, 0.8, and 0.95, the strategy has an overall success probability of 68.4%.

[0052] Fault handling strategy Resource costs of execution Fault handling strategy The sum of the expected resource costs consumed by all agents during execution, and the resource cost of executing the fault handling strategy. The calculation method is as follows: In the formula, , Fault handling strategy The Middle A smart agent Resource costs (communication volume, computing volume, number of manual interventions, etc.). For intelligent agents The resource cost vector of the first Each component (such as CPU and memory consumption). The weight of this resource component reflects its importance and is specified by the operation and maintenance strategy or learned from historical data. During fault handling, the system calculates the resource cost for each candidate strategy and makes a comprehensive decision based on reliability and time delay.

[0053] Fault handling strategy Execution time delay Fault handling strategy The sum of the expected time delays consumed by all agents during execution, and the time delay of the fault handling strategy execution. The calculation method is as follows: In the formula, Fault handling strategy The Middle A smart agent The estimated execution time for handling this fault (can be estimated from the average value of equipment operation logs or from the protocol / equipment specifications).

[0054] The reward function used to calculate the reward score for the fault handling strategy is the fault handling strategy. The reward function is obtained by subtracting the penalties for resource cost and time delay from the reliability, and then balancing the importance of reliability, resource cost, and time delay using weighted coefficients. The reward function is expressed as follows: In the formula, These are reliability weight, resource cost weight, and time delay weight, respectively. The sum is 1. Fault handling strategy Bonus points, bonus points Used to measure fault handling strategies The overall benefits obtained after execution.

[0055] Reliability weight Emphasizing the importance of successful fault handling by the strategy. A higher value indicates a greater tendency for the algorithm to choose strategies with higher success rates; values ​​are typically set between 0.4 and 0.6 to prioritize safety. Resource cost weight. The impact of controlling resource economy. A larger value means the algorithm penalizes strategies with high resource consumption; it's typically set between 0.2 and 0.3 to balance performance and resource usage. Time delay weight. Priority affects the speed of strategy execution. A higher value indicates a greater preference for strategies with faster response times. Values ​​are typically set between 0.2 and 0.3 to minimize downtime. These coefficients need to be set by the system administrator based on specific operational needs or optimized using historical data. For example, during peak business hours, they can be increased to prioritize faster recovery.

[0056] The reward function takes into account reliability, resource cost, and time delay, avoiding the one-sidedness caused by single-objective optimization. This makes the reward function computable and optimizable, providing an accurate evaluation benchmark for the GRPO algorithm.

[0057] (6) In the reward function construction module, among all fault handling strategies obtained through agent orchestration, based on the device attributes of the root cause of the fault event and the device's operating state, the specific method for finding the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms is as follows: The fault handling strategy network in GRPO is denoted as , For strategy parameters, This is a set of operating environment information and fault context features of the power grid automated dispatching system at time t. This includes the current operating status of the scheduling system, alarm information characteristics, historical actions, and processing results. For example... ={0.82, 120, "Bus voltage limited", 3, "Switching to standby bus - successful"}, where the parameters are system load rate, communication link delay, alarm type, alarm severity, and historical action feedback, respectively. Fault handling strategy The corresponding intelligent agent in the fault handling strategy network The specific action decisions obtained from the sampling, and the fault handling strategy network Output in the current Select each fault handling strategy below The probability distribution of , where; In the formula, It is a non-linear activation function. This is the attribute feature vector of the device that is the root cause of the fault event. This is the feature vector of the operating state of the equipment at the root cause of the fault event. For bias terms, This is a weight matrix representing the attributes of the device that is the root cause of the failure event. The weight matrix represents the operating state of the equipment at which the root cause of the failure event is located; Calculate the advantage function : In the formula, For state-value functions, a state-value function is defined as a function of state... The following is based on the fault handling strategy. Expected cumulative return from discounts starting at time t The calculation formula is as follows: In the formula, Fault handling strategy At any moment Instant rewards received As a discount factor, Used to measure the relative importance of immediate rewards versus future rewards; The GRPO optimization objective function is constructed as follows: in, The threshold is set to prevent fault handling strategies from being updated too quickly. To guide fault handling strategies (primarily derived from existing expert fault handling strategies, knowledge graphs, and historical behavior distributions, such as historical success strategies and expert rules). The KL divergence between the current fault handling strategy and the guiding fault handling strategy (serving as a penalty term for the fault handling strategy deviating from the prior). This is the regularization strength coefficient (used to balance the dependence between exploring the next fault handling strategy and guiding the fault handling strategy). These are the strategy parameters from the previous iteration.

[0058] The strategy parameters are iteratively updated by optimizing the objective function using GRPO. This refers to the training of GRPO. The GRPO training process employs a fault-handling strategy and gradient update rules to ensure that while obtaining positive rewards, the absorption and utilization of expert knowledge are preserved. This continues until the reward function... Convergence occurs, at which point the fault handling strategy network... The output is in the current Select each fault handling strategy below In the probability distribution, the fault handling strategy with the highest probability is the fault handling strategy with the highest reward score.

[0059] Optimize the parameters in the objective function of GRPO The process continues, and ultimately the system selects the fault handling strategy with the highest expected reward. This enables the adaptive selection and orchestration of intelligent agents.

[0060] The GRPO algorithm ensures the convergence and safety of policy optimization, enabling the system to dynamically generate policies with the highest reward scores and achieve intelligent operation and maintenance through autonomous learning and evolution.

[0061] This invention proposes an intelligent fault and anomaly handling method for scheduling automation systems based on agent-based adaptive orchestration. It integrates multiple stages, including fault prediction, root cause localization, policy generation, and adaptive orchestration, comprehensively improving the system's response efficiency and intelligent handling capabilities for abnormal events. By introducing an LSTM-GAN structure for accurate prediction of operational states and combining it with a Bayesian network for accurate fault root cause localization, a GRPO-based policy generation mechanism is introduced, enabling rapid response and accurate decision-making by the agent when facing diverse abnormal events in the scheduling automation system. Compared to traditional fixed-rule or statically mapped policy formulation methods, GRPO can incorporate expert prior knowledge during policy optimization, effectively improving policy interpretability and convergence efficiency while avoiding overfitting.

[0062] This method features strong model generalization ability, fast response and decision-making speed, and self-learning optimization of execution feedback. While improving the level of automation in fault handling, it significantly reduces human intervention and enhances the intelligence, adaptability, and operation and maintenance efficiency of the scheduling automation system.

[0063] Example 2 A power grid dispatching fault intelligent handling method based on agent adaptive orchestration includes: Clustering algorithms are used to cluster all alarm samples of the power dispatching system within the current time window to obtain an alarm cluster set. An alarm cluster is a set of all alarm samples of alarm events caused by the same root cause. Input the alarm clusters in the alarm cluster set into the fault detection model to determine whether the alarm events corresponding to the alarm clusters are fault events, and output the determination result of whether the alarm events corresponding to the alarm clusters are fault events; For alarm clusters whose corresponding alarm events are fault events in the judgment results, perform causal reasoning to locate the equipment where the root cause of the fault event is located; The sequence obtained by orchestrating all agents in the power dispatching system for handling the fault event is defined as the fault handling strategy. Among all fault handling strategies obtained by agent orchestration, based on the equipment attributes of the root cause of the fault event and the operating status of the equipment, a guided reward strategy optimization algorithm is adopted to find the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms, and the fault handling strategy with the highest reward score is adopted to handle the corresponding fault event.

[0064] Example 3 A computer program product includes a computer program / instructions that, when executed by a processor, implement the intelligent power grid dispatch fault handling method based on agent adaptive orchestration in Embodiment 2.

[0065] The contents not described in detail in this specification are prior art known to those skilled in the art. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] 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.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] 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.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A power grid dispatching fault intelligent processing system based on agent adaptive orchestration, characterized in that, include: The alarm clustering module is used to cluster all alarm samples of the power dispatching system within the current time window using a clustering algorithm to obtain an alarm cluster set. An alarm cluster is a set of all alarm samples of alarm events caused by the same root cause. The fault event judgment module is used to input the alarm clusters in the alarm cluster set into the fault detection model to judge whether the alarm event corresponding to the alarm cluster is a fault event, and output the judgment result of whether the alarm event corresponding to the alarm cluster is a fault event; The root cause localization module is used to perform causal reasoning on alarm clusters whose corresponding alarm events are fault events in the judgment results, and to locate the device where the root cause of the fault event is located. The fault handling module defines the sequence obtained by orchestrating all agents in the power dispatching system for handling the fault event as a fault handling strategy. Among all fault handling strategies obtained by agent orchestration, based on the equipment attributes of the root cause of the fault event and the operating status of the equipment, a guided reward strategy optimization algorithm is used to find the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms, and the fault handling strategy with the highest reward score is used to handle the corresponding fault event.

2. The intelligent power grid dispatching fault handling system based on agent adaptive orchestration according to claim 1, characterized in that, In the alarm clustering module, the specific method for using a clustering algorithm to cluster all alarm samples of the power dispatching system within the current time window to obtain an alarm cluster set is as follows: Building an alarm sample set ; The alarm sample set was clustered using the DBSCAN density clustering algorithm to obtain an alarm cluster set. In the formula, For the first A cluster of alarms, For the first The first in the alarm cluster Alarm sample The alarm samples are arranged in chronological order.

3. The intelligent power grid dispatching fault handling system based on agent adaptive orchestration according to claim 2, characterized in that, The fault event judgment module inputs alarm clusters from the alarm cluster set into the fault detection model to determine whether the alarm events corresponding to the alarm clusters are fault events, and outputs the judgment results of whether the alarm events corresponding to the alarm clusters are fault events. The specific method is as follows: The fault detection model was obtained by training an LSTM-GAN model. alarm cluster Reconstruction loss function in input fault detection model In the formula, For alarm clusters The generator output; alarm cluster Discriminant loss function in input fault detection model In the formula, For alarm clusters The discriminator output; Calculate the anomaly score in the fault detection model In the formula, These are the weighting coefficients; If abnormal scores If the score exceeds the threshold, an alarm cluster will be triggered. The corresponding event is a fault event, if the abnormal score If the score is less than the threshold, an alarm cluster will be triggered. The corresponding event is a normal event.

4. The intelligent power grid dispatching fault handling system based on agent adaptive orchestration according to claim 1, characterized in that, The root cause localization module performs causal reasoning on alarm clusters whose corresponding alarm events are fault events in the judgment results, and the specific method for locating the device where the root cause of the fault event is located is as follows: Construct a Bayesian network, whose nodes include device nodes, device operation status nodes, and alarm nodes; Given a series of alarm nodes, calculate the conditional probability of each node, calculate the posterior probability of all device nodes to be inferred based on the conditional probability of each node, and select the device node with the highest posterior probability as the node where the root cause is located.

5. The intelligent power grid dispatching fault handling system based on agent adaptive orchestration according to claim 1, characterized in that, In the fault handling module, the specific method for defining the sequence obtained by orchestrating all agents in the power dispatching system used to handle the fault event as the fault handling strategy is as follows: Fault handling strategy represents an execution plan in which multiple agents work collaboratively in a specific order, denoted as: In the formula, For the k-th fault handling strategy, This represents the j-th agent in the k-th fault handling strategy.

6. The intelligent power grid dispatching fault handling system based on agent adaptive orchestration according to claim 5, characterized in that, In the fault handling module: The guided reward strategy optimization algorithm adopts the GRPO algorithm. Based on the reliability of the fault handling strategy, the resource cost of the fault handling strategy, and the time delay of the fault handling strategy, the reward function used in the GRPO algorithm to calculate the reward score of the fault handling strategy is constructed.

7. The intelligent power grid dispatching fault handling system based on agent adaptive orchestration according to claim 6, characterized in that: Fault handling strategy Execution reliability Fault handling strategy The product of the success probabilities of all agents in the system handling this type of fault; Fault handling strategy Resource costs of execution Fault handling strategy The sum of the expected resource costs consumed by all agents during execution; Fault handling strategy Execution time delay Fault handling strategy The sum of the expected time delays consumed by all agents during execution. The reward function used to calculate the reward score for the fault handling strategy is the fault handling strategy. The function is obtained by subtracting resource costs and time delays from the reliability, and then balancing the importance of reliability, resource costs, and time delays through weighting coefficients.

8. The intelligent power grid dispatching fault handling system based on agent adaptive orchestration according to claim 7, characterized in that, In the reward function construction module, among all fault handling strategies obtained through agent orchestration, based on the device attributes of the root cause of the fault event and the device's operating state, the specific method for finding the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms is as follows: The fault handling strategy network in GRPO is denoted as , For strategy parameters, This is a set of operating environment information and fault context features of the power grid automated dispatching system at time t. Fault handling strategy The corresponding intelligent agent in the fault handling strategy network The specific action decisions obtained from the sampling, and the fault handling strategy network Output in the current Select each fault handling strategy below The probability distribution of , where; In the formula, It is a non-linear activation function. This is the attribute feature vector of the device that is the root cause of the fault event. This is the feature vector of the operating state of the equipment at the root cause of the fault event. For bias terms, This is a weight matrix representing the attributes of the device that is the root cause of the failure event. The weight matrix represents the operating state of the equipment at which the root cause of the failure event is located; Calculate the advantage function : In the formula, For the reward function, For state-value functions, a state-value function is defined as a function of state... The following is based on the fault handling strategy. Expected cumulative return from discounts starting at time t The calculation formula is as follows: In the formula, Fault handling strategy At any moment Instant rewards received Discount factor; The GRPO optimization objective function is constructed as follows: in, The cropping threshold, To guide the fault handling strategy, The KL divergence between the current fault handling strategy and the guided fault handling strategy is given. The regularization intensity coefficient is . These are the strategy parameters from the previous iteration; The strategy parameters are iteratively updated by optimizing the objective function using GRPO. until the reward function Convergence occurs, at which point the fault handling strategy network... The output is in the current Select each fault handling strategy below In the probability distribution, the fault handling strategy with the highest probability is the fault handling strategy with the highest reward score.

9. A method for intelligent fault handling in power grid dispatching based on agent adaptive orchestration, characterized in that, include: Clustering algorithms are used to cluster all alarm samples of the power dispatching system within the current time window to obtain an alarm cluster set. An alarm cluster is a set of all alarm samples of alarm events caused by the same root cause. Input the alarm clusters in the alarm cluster set into the fault detection model to determine whether the alarm events corresponding to the alarm clusters are fault events, and output the determination result of whether the alarm events corresponding to the alarm clusters are fault events; For alarm clusters whose corresponding alarm events are fault events in the judgment results, perform causal reasoning to locate the equipment where the root cause of the fault event is located; The sequence obtained by orchestrating all agents in the power dispatching system for handling the fault event is defined as the fault handling strategy. Among all fault handling strategies obtained by agent orchestration, based on the equipment attributes of the root cause of the fault event and the operating status of the equipment, a guided reward strategy optimization algorithm is adopted to find the fault handling strategy with the highest reward score among the guided reward strategy optimization algorithms, and the fault handling strategy with the highest reward score is adopted to handle the corresponding fault event.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the intelligent power grid dispatching fault handling method based on agent adaptive orchestration as described in claim 9.

Images