A size model collaborative empowerment protection control device agent method and system

Abstract

The application discloses a kind of size model collaborative empowerment protection control equipment intelligent agent method and system, method includes: perception intelligent agent gathers one, secondary equipment multi-source electrical and non-electrical quantity parameter and the multilevel state monitoring information of equipment ontology;Analysis and judgment intelligent agent small model realizes the special diagnosis and feature extraction of typical implicit fault, and large model is according to four-stage framework and is verified by digital twin body to carry out interpretable reasoning;Decision-making intelligent agent operates with double closed loop mechanism, and the risk level is divided by improved fault tree analysis algorithm, and linkage search enhancement technology generates disposal scheme;Early warning intelligent agent generates structured report when single station is abnormal, and constructs collaborative early warning network;Operation intelligent agent carries out full-process rehearsal and check after through AR / VR interactive interface, executes encryption and double-checking embodied operation.The application realizes the closed-loop diagnosis of protection control equipment implicit fault and the linkage disposal of associated equipment, provides core technical support for the safe and stable operation of power system.

Description

Technical Field

[0001] This invention relates to the field of power system protection and monitoring technology, and in particular to a method and system for intelligent protection and control equipment that enables collaborative empowerment of large and small models. Background Technology

[0002] As the power system evolves towards a new type of power system that integrates "source, grid, load, and storage," protection and control equipment faces multiple challenges, including diversified data sources, more complex fault types, and real-time response requirements. Scenarios such as the random fluctuations of distributed power sources and flexible interactions on the load side make the grid operating conditions increasingly complex and variable, leading to an exponential increase in the electrical and non-electrical parameters that protection and control equipment needs to handle. Simultaneously, latent faults such as sampling circuit deviation monitoring and degradation aging early warning, caused by long-term equipment operation, are gradually becoming more prominent. These faults have weak initial characteristics and complex evolution paths, making them difficult to detect using traditional monitoring methods and easily triggering chain reactions.

[0003] Traditional protection and control schemes often employ fixed logic criteria and threshold setting modes, resulting in weaknesses such as poor ability to identify latent faults within the protection and control equipment itself, insufficient cross-device fault correlation analysis, and weak adaptability of the system's autonomous decision-making. Traditional fault diagnosis relies solely on surface alarm signals and routine electrical quantity monitoring, lacking deep state perception at the chip and board levels of the equipment itself. This leads to weak identification of latent faults within the protection and control equipment, and makes it impossible to achieve accurate diagnosis and early warning for secondary equipment. Often, a reactive response is only taken after the fault has escalated. Each protection and control module operates independently, lacking a global perception of the fault propagation path across devices and sites, making it difficult to accurately locate the root cause and scope of impact of the fault. Decision-making and execution rely excessively on human experience for fault assessment and handling plan formulation, failing to quickly generate optimal handling strategies adapted to the fault scenario.

[0004] In existing technologies, the application of single AI models in power scenarios has significant limitations: while small models possess real-time processing capabilities, their generalization ability is insufficient, making them unable to handle complex faults across different domains; general-purpose large models lack protection and control expertise, and direct invocation can easily produce inference results that do not conform to industry standards. Furthermore, most protection and control modules operate independently, failing to form a closed-loop collaborative mechanism of "data-analysis-decision-early warning-operation-governance," resulting in low fault handling efficiency and a lack of cross-site collaboration capabilities. When faced with complex faults propagating across different sites, they cannot quickly and accurately isolate power grid faults through protection actions, easily leading to the expansion of the fault range and seriously threatening the safe and stable operation of the power grid. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method and system for intelligent protection and control equipment that integrates professional small models and power adaptation large models, connects data flow and business collaboration links, and enables accurate diagnosis and early warning of hidden faults in secondary equipment.

[0006] Technical solution: The intelligent agent method for protection and control equipment with collaborative empowerment of large and small models as described in this invention includes:

[0007] (1) The sensing agent collects multi-source electrical and non-electrical parameters of primary and secondary equipment and multi-level status monitoring information of the equipment body. The data is uploaded after preprocessing by the protection and control equipment body and invalid data is removed on the intelligent agent side.

[0008] (2) The intelligent agent uses a special small model to perform hidden fault diagnosis and feature extraction on the protection and control equipment side. The protection and control knowledge enhancement large model receives fault features and forms interpretable reasoning conclusions according to the four-stage framework. Combined with the digital twin and real-time data cross-validation, it forms an accurate diagnosis of the power grid operation status.

[0009] (3) The decision-making agent receives the reasoning conclusions in the uplink, calculates the probability of fault propagation and classifies the risk level using the improved fault tree analysis algorithm, and pushes the enhanced retrieval to generate a disposal decision plan that matches the fault type and risk level to the early warning agent; the downlink pushes the coded and converted control information to the operation agent.

[0010] (4) When the early warning intelligent agent makes an abnormal judgment decision at a single station, it generates a structured report and automatically generates a cross-station linkage list based on the power grid topology to build a collaborative early warning network;

[0011] (5) After the intelligent agent performs a full-process rehearsal and verification through the interactive interface, it performs the embodied operation of encryption and double verification;

[0012] (6) Data governance intelligent agents classify and store data to provide iterative data for special small models and protection and control knowledge enhancement large models.

[0013] Furthermore, in step (1), the preprocessing algorithm is deployed on the edge side of the protection and control equipment. For the time-series parameters, the wavelet packet transform algorithm is used for denoising and time-frequency domain feature extraction. For the equipment status parameters, normalization processing is performed. The preprocessed data is encapsulated into standardized data frames and uploaded to the data governance and analysis intelligent agent.

[0014] After the data from the protection and control equipment is uploaded to the intelligent agent, it is processed by an adaptive threshold filtering algorithm and a data validity verification model. The adaptive threshold filtering algorithm filters out abnormal sampled values ​​that exceed the dynamic threshold based on the statistical analysis of the equipment's historical operating data and the reasonable fluctuation range of the data dynamically calculated in real-time operating conditions. The data validity verification model integrates business rules and multi-source data consistency comparison. Data that passes the collaborative verification is marked as high-confidence valid data and input into the subsequent analysis and judgment process.

[0015] Further, step (2) includes:

[0016] (21) Construct a model collaboration architecture, with special small models deployed on the protection and control equipment side, and protection and control knowledge enhancement large models deployed on the intelligent agent side to receive diagnostic information and feature data;

[0017] (22) Based on the operating mechanism, operation and maintenance specifications and fault handling cases of protection and control equipment, a domain-level dataset is constructed. The large model is fine-tuned using low-rank adaptation technology. After the fine-tuning is completed, the backbone network is frozen and the domain adaptation parameters are retained to form an inference model with a professional knowledge baseline of power protection and control.

[0018] (23) Perform collaborative reasoning based on the four-stage framework. The protection and control knowledge enhancement big model follows the four-stage reasoning framework of fault feature extraction, cross-equipment correlation analysis, impact range assessment, and disposal priority ranking. It calls the retrieval enhancement module to perform logical reasoning on the received fault features and form accurate diagnostic conclusions on the power grid operation status of the associated primary equipment.

[0019] (24) Establish a split-state coordination and verification algorithm. When the protection and control equipment is running normally, the small model performs periodic monitoring and the large model integrates the data. When the small model diagnoses a potential hidden fault in the equipment, it immediately triggers the large model to perform rapid reasoning, outputs the diagnostic conclusion, and performs cross-verification with the real-time data through the digital twin. If the deviation exceeds the threshold, it triggers secondary reasoning to optimize the result and outputs the diagnostic conclusion to the decision-making agent.

[0020] Furthermore, the four-stage framework includes: fault feature extraction and service mapping stage, service matching and collaborative invocation stage, cross-service correlation analysis and comprehensive judgment stage, and conclusion encapsulation and feedback tracing stage.

[0021] Furthermore, in step (3), the decision-making agent data parsing module extracts and integrates the received fault features and interpretable reasoning conclusions;

[0022] For decisions that can be remotely operated or automatically simulated in a digital twin environment, the decision-making agent will encode the operable parts into control command information through standardized protocols and push them to the operating agent through an encrypted channel. Before the command is issued, it must be verified for compliance by a verification engine.

[0023] After the operating agent executes the instruction, it feeds back the operation result status to the decision-making agent. The decision-making agent compares and verifies the feedback result with the expected status. If there is a deviation or execution failure, a secondary retrieval and decision optimization process is triggered to regenerate or adjust the handling plan.

[0024] Furthermore, the improved fault tree analysis algorithm includes:

[0025] Dynamic root node adaptation uses the main fault cause clearly identified in the reasoning conclusion of the knowledge-enhanced big model as the dynamic root node of the fault tree, replacing the traditional fixed root node setting, and adapting to different types of hidden fault scenarios of protection and control equipment.

[0026] Dynamic fault tree structure construction combines real-time feature parameters provided by the small model to dynamically generate the fault tree structure. Intermediate events are updated with real-time operating conditions, covering the entire propagation path of faults in the protection and control secondary circuit, associated primary equipment and power grid topology.

[0027] Boolean algebra is used to simplify the fault tree to obtain the minimum cut set. The fault propagation probability is calculated based on the minimum cut set method, and four risk levels are determined according to the fault propagation probability.

[0028] Further, step (4) includes:

[0029] (41) The early warning agent calls the data validity verification model to verify the integrity of the received protection and control equipment fault data, fault type, risk propagation probability, associated primary equipment and power grid operation status and the logical consistency of protection and control.

[0030] (42) Based on the received data, the intelligent agent automatically retrieves historical operation and maintenance records and inspection procedures, and executes single-station early warning output in a state-by-state manner;

[0031] (43) Automatic generation of cross-site linkage list and collaborative early warning: When the probability of fault propagation calculated by the decision-making agent exceeds the preset threshold, the early warning agent automatically triggers cross-site linkage.

[0032] Further, step (5) includes:

[0033] (51) Synchronize the three-dimensional geometric coordinates, board layout, terminal block wiring relationship and real-time operating status of the protection and control equipment body to the AR / VR terminal to form a virtual and real operation interface that integrates the physical coordinates of the equipment;

[0034] (52) Through the AR / VR terminal, according to the control information instructions, perform full-process pre-drill and logic verification of protection control business, call the protection control logic rule library, simulate and verify the secondary circuit state change and protection function activation / deactivation logic chain relationship caused by the pre-drill operation, and verify its logical correctness and impact on the secondary circuit business chain.

[0035] (53) The intelligent agent encodes the verified embodied operation instructions into control command frames according to the standardized protocol, processes them through the permission and logic verification engine and encryption algorithm, and sends them out through the station control layer communication interface;

[0036] (54) During the operation, the status change signal fed back by the physical device is compared with the expected status of the virtual device in the digital twin through real-time and high-speed connection. If the deviation exceeds the safety threshold, an early warning is immediately triggered and subsequent operations are automatically blocked. At the same time, the abnormal status is fed back to the decision-making intelligent agent to start the optimization process.

[0037] Further, step (6) includes:

[0038] (61) Based on the business logic of protection and control equipment operation, fault diagnosis and handling, establish a data classification and labeling system for protection and control business, and divide the collected multi-source heterogeneous data into real-time operation data, historical case data and domain knowledge data;

[0039] (62) The data agent receives data requests from each agent through the API interface, and matches and extracts data from the corresponding data pool based on the business scenario tags carried in the request;

[0040] (63) The newly generated data, such as the status feedback after the operation of the intelligent agent, the disposal report generated by the early warning intelligent agent, the protection and control equipment after the operation, the associated primary equipment, and the status of the power grid operation, along with their corresponding original fault data and diagnostic conclusions, are classified and stored as new historical case data after the governance process, forming a data pool that supports the application of special models and protection and control knowledge enhancement big models.

[0041] A protection and control device intelligent agent system with collaborative empowerment of large and small models, comprising:

[0042] Sensing intelligent agent: used to collect electrical and non-electrical parameters through multi-source sensors of primary and secondary equipment, and at the same time collect multi-level status monitoring information of the equipment body at the chip level, board level, and whole machine level;

[0043] The analysis and judgment intelligent agent is connected to the sensing intelligent agent. A specialized small model is used to realize the specialized diagnosis and feature extraction of typical hidden faults. The protection and control knowledge is used to enhance the large model to perform interpretable reasoning based on a preset reasoning framework. After verification by a digital twin, a power grid operation status diagnosis conclusion is generated.

[0044] Decision-making intelligent agent: It communicates with the analysis and judgment intelligent agent, classifies risk levels through an improved fault tree analysis algorithm, and generates a handling plan through linkage retrieval enhancement technology. It pushes the plan to the early warning intelligent agent and pushes control commands to the operation intelligent agent.

[0045] Early warning agent: Communicatively connected to the decision-making agent, used to generate structured reports and update the early warning ledger when a single station is abnormal, and automatically generate a cross-station linkage list based on the power grid topology to build a cross-station collaborative early warning network;

[0046] Operational intelligent agent: Communicatively connected to the decision-making intelligent agent, used to perform embodied operations of encryption and double verification after full-process pre-playing and verification through the AR / VR interactive interface;

[0047] Data governance agents: These agents communicate with each of the aforementioned agents to store data categorized by real-time, historical, and knowledge types, and match the needs of each agent using tags.

[0048] Beneficial effects: Compared with existing technologies, this invention has the following significant advantages: 1. Effectively reduces the risk of misjudgment of faults, making the diagnostic results more consistent with the actual operation scenario of the power grid; 2. Solves the bottleneck of incomplete equipment status monitoring information in traditional technologies, providing a high-quality data foundation for subsequent diagnostic decisions; 3. Shortens the response time from fault perception to operation execution, and quickly curbs the spread of faults through differentiated early warning and cross-station linkage mechanisms, building a comprehensive power grid operation safety protection network and improving overall operation stability; 4. Avoids the risks of violations and misoperation, forming a closed loop of continuous empowerment, helping the system to continuously upgrade with operational practice and adapt to the long-term development needs of the power grid. Attached Figure Description

[0049] Figure 1 This is a general diagram of a multi-agent architecture;

[0050] Figure 2 A schematic diagram illustrating the fusion of large and small models of intelligent agents;

[0051] Figure 3 This is a schematic diagram of the split-state collaborative workflow;

[0052] Figure 4 This is a schematic diagram of a data governance intelligent agent. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, the intelligent agent method for protection and control equipment with collaborative empowerment of large and small models according to the present invention includes:

[0055] (1) Operation of the perceptual intelligent agent

[0056] Electrical and non-electrical parameters are collected by multiple source sensors in primary and secondary equipment. At the same time, multi-level status monitoring information of the equipment body at the chip level, board level, and whole machine level is collected. The data is preprocessed in the protection and control equipment body and uploaded to the data governance and analysis intelligent agent through the MQTT protocol. On the intelligent agent side, invalid data is removed through adaptive threshold filtering algorithm and data validity verification model, providing high-confidence data support for upper-level analysis and judgment.

[0057] (1.1) To construct a multi-source sensing system covering the status of the protection and control equipment itself, the operating status of associated equipment, and the environmental status:

[0058] The system provides multi-level status monitoring information for the protection and control equipment, including chip-level, board-level, and system-wide levels. This status monitoring information covers core chip operating current and junction temperature, board power supply voltage ripple, system power consumption, and cooling fan speed, and integrates event sequence data such as device self-tests and setting modifications.

[0059] The status quantities of circuit breakers and disconnectors, relay protection actions and alarms, primary equipment auxiliary status quantities, operation and interlocking status quantities, etc., monitored by protection and control equipment;

[0060] The system synchronously collects electrical parameters such as bus voltage / current and line voltage / current provided by primary and secondary equipment sensors, as well as non-electrical parameters such as transformer oil temperature, and transmits them to the sensing agent through standardized interfaces or sensor networks.

[0061] (1.2) Precise preprocessing

[0062] For time-series parameters such as current and voltage, wavelet packet transform algorithm is used for denoising and time-frequency domain feature extraction. Multi-scale time-frequency features are obtained through three-level wavelet packet decomposition to separate effective signal and noise components. The VisuShrink thresholding method is used to suppress the noise components. The threshold calculation formula is as follows:

[0063]

[0064] Where σ is the noise standard deviation (estimated using the first 100 sampling points of the signal), and N is the data sampling length. After processing, the data is reconstructed using inverse wavelet transform, and the reconstructed data is further processed using 3×3 window midpoint filtering to eliminate impulse noise.

[0065] For the status parameters of the device at the chip level, board level, and overall system level, the min-max normalization method is used to eliminate the influence of dimensions. The normalization formula is as follows:

[0066]

[0067] Where x represents the original data, x min x max These are the historical extreme values ​​of this parameter.

[0068] (1.3) Encapsulate the preprocessed data into standardized data frames and upload them to the data governance and analysis agent via the MQTT protocol:

[0069] Valid data is encapsulated using a standard format. The encapsulation structure includes fixed fields and extended fields. Fixed fields include device ID, collection timestamp, parameter type code, original value, and checksum. Extended fields add exclusive parameters based on device type to ensure unique data identification and complete information.

[0070] The MQTT protocol is used for transmission configuration, and QoS levels are set to ensure reliable message delivery. The client ID consists of the site code, the type of protection and control device, and the sensor ID of the associated device. A topic hierarchical mechanism is established. If a connection is interrupted during transmission, the local cache is automatically started, and the message is retransmitted in timestamp order after the connection is restored.

[0071] (1.4) After the data from the protection and control equipment is uploaded to the intelligent agent, the invalid data is accurately removed through an adaptive threshold filtering algorithm and a data validity verification model:

[0072] Based on the statistical analysis of historical equipment operation data and the reasonable fluctuation range of real-time operating condition dynamic calculation data, a dynamic update mechanism for parameter fluctuation thresholds is established to filter abnormal sampled values ​​that exceed the dynamic thresholds, ensuring the rationality of the data.

[0073] By integrating business rules with multi-source data consistency comparison, a data validity verification model is formed. Business rule verification includes core rules such as parameter range and rate of change. Multi-source data consistency comparison ensures the authenticity of data by cross-validating the correlation parameters of data from different dimensions.

[0074] (1.5) Data that passes the collaborative verification is marked as high-confidence valid data and input into the subsequent analysis and judgment process. At the same time, a verification log is recorded, including verification pass / fail indicators, details of abnormal data, and other information.

[0075] (2) An analytical and judgmental intelligent agent is realized through a collaborative approach of specialized small models and overall coordination of large models. Specialized small models realize the diagnosis and feature extraction of typical hidden faults such as sampling loop deviations on the protection and control equipment side. The protection and control knowledge-enhanced large model receives fault features and forms interpretable reasoning conclusions according to the four-stage framework. After cross-validation with digital twins and real-time data, a precise diagnosis of the power grid operation status of related primary equipment is formed. Figure 2 , 3 The process is as follows:

[0076] (2.1) Construct a deployment architecture for distributed monitoring and centralized inference, and clarify the functional boundaries and collaboration mechanisms of each model:

[0077] Specialized mini-models are deployed on the protection and control equipment side, divided into sub-domains according to the type of latent faults, and deployed specialized mini-models such as sampling loop deviation monitoring, deterioration and aging early warning, and setpoint drift detection. Each mini-model completes standardized service encapsulation, and clarifies the service ID, input parameter specifications, output protocol and health check mechanism. The input of the specialized model is the core state quantity of the protection and control equipment, and the output is a standardized fault description and characteristic parameter set.

[0078] Based on the Qwen-3 model, after being fine-tuned by protection and control knowledge, it is deployed on servers with large model inference capabilities in substations and central control stations. An API gateway is built to complete the unified registration and routing management of small model services, and to receive fault characteristic information output by small models through APIs or interfaces.

[0079] (2.2) The general large model was transformed into a protection and control vertical domain using low-rank adaptation technology:

[0080] We constructed a question-answer pair dataset covering typical areas such as protection and monitoring principles, device operation mechanisms, equipment operation and maintenance specifications, power grid topology information, and fault handling cases. The format is "question-answer-feature parameter annotation". The annotation information includes fault type, core feature parameters, diagnostic basis, etc., injecting professional knowledge of power protection and control into the large model.

[0081] Freeze the Qwen3-7B backbone network parameters and only fine-tune the LoRA adapter parameters:

[0082] (2.2.1) Let the weight matrix of the pre-trained model be... (where d is the input dimension and k is the output dimension), LoRA introduces two low-rank matrices. (rank r) and During fine-tuning, only A and B are updated. With the introduction of LoRA, the forward propagation process of the model is as follows:

[0083]

[0084] x represents the input features of the large model, and r represents the LoRA rank. This is the scaling factor. For the fine-tuned weight matrix, we choose r=16. =16, which is much smaller than d=4096 and k=4096;

[0085] (2.2.2) The cosine annealing algorithm is used for learning rate scheduling, with an initial learning rate of The minimum learning rate is set to The annealing cycle is 50 rounds, and the AdamW optimizer is used. The training strategy uses a batch size of 32 and a gradient accumulation step of 4.

[0086] (2.2.3) After fine-tuning, freeze the Transformer layer parameters of the backbone network of the large model, retain only the updated LoRA adapter parameters, and solidify the core logic parameters of service calls. Generate a dedicated model file for the fault diagnosis of protection and control equipment, and improve the inference speed of the knowledge-enhanced large model through INT8 quantization technology.

[0087] (2.3) The protection and control knowledge enhancement model follows a four-stage reasoning framework: fault feature extraction, cross-fault correlation analysis, impact range assessment, and handling priority ranking. It calls the retrieval enhancement (RAG) module to perform interpretable logical reasoning on the received fault features, and finally forms an accurate diagnostic conclusion on the grid operation status of the associated primary equipment:

[0088] (2.3.1) To achieve interpretable and traceable accurate diagnosis, a four-stage logical framework for collaborative reasoning of the large-scale model enhanced by solidifying protection and control knowledge is established:

[0089] Fault Feature Extraction and Service Mapping Stage: The large model receives the preprocessed sensing data, analyzes the data patterns based on the built-in knowledge of power protection and control, determines the potential types of fault features, and transforms the abstract data problem into specific, callable professional service requests.

[0090] Service matching and collaborative invocation phase: The system verifies the real-time online status and load of the target special small model service by querying the service registry. After confirming availability, it initiates a standardized invocation request with parameter signature and time stamp.

[0091] Cross-service correlation analysis and comprehensive judgment stage: When a fault may involve multiple loops or state variables, the large model will aggregate the diagnostic results of multiple related small models, and combine them with the built-in knowledge base of related primary equipment and power grid topology to perform logical reasoning and judgment on the correlation relationship and impact path of the fault.

[0092] Conclusion encapsulation and feedback traceability stage: The final diagnostic conclusion is encapsulated according to the standard data format, including the inferred fault type, model confidence, and service call logs, and the structured conclusion is synchronously fed back to the decision-making agent.

[0093] (2.3.2) The Retrieval Enhancement (RAG) module constructs a knowledge base of protection and control fault cases and stores historical call logs, including related data such as fault scenarios, feature parameters, called services, and response results. When the large model receives the return results from the small model, it uses the Sentence-BERT model to convert the text data into a 512-dimensional vector. After vector normalization, the vector is stored in the Milvus vector database. The index type adopts the retrieval precision priority method, and the number of cluster centers is set to 1024.

[0094] (2.3.3) Combining BM25 keyword retrieval with Sentence-BERT semantic retrieval, 20 similar fault cases are matched from the fault case database. The BM25 keyword retrieval formula is as follows:

[0095]

[0096] Where Q is the set of keywords in the query text, and D is the target document in the case knowledge base; The word frequency of keyword q in document D; The length of document D (number of characters); The average length of all documents in the knowledge base; term frequency saturation coefficient. Set to 1.2; document length normalization factor b is set to 0.75; IDF is inverse document frequency.

[0097] (2.3.4) Taking into account the fault scenario, feature parameter similarity, and topology adaptability of associated equipment, the matching results are re-ranked using the Cross-BERT model, and three optimal cases are selected as the basis.

[0098] (2.3.5) The knowledge-enhanced large model integrates case information and call experience to output conclusions. The content includes fault type, confidence level, reasoning basis, call service ID and status. When the model confidence level is lower than 90%, the service expansion mechanism is triggered to increase the resource quota of the target small model and re-call to obtain more accurate features.

[0099] (2.4) Fractal coordination and verification

[0100] Establish a state-based coordination and verification algorithm to dynamically adjust the model's working mode based on the equipment's operating status, and ensure accurate diagnostic conclusions through cross-validation between digital twins and real-time data.

[0101] (2.4.1) Fragmented cooperative mechanism:

[0102] When the protection and control equipment is running normally, the small model monitors the state variables involved on time according to their characteristics. It only actively reports the status to the large model when the parameters are close to the threshold, thus avoiding invalid calls and resource consumption. The large model performs a full service status inspection once an hour, calls the diagnostic results of the small model as needed, integrates the data of each small model to generate a report containing the operation score of the protection and control equipment, the risk points of related equipment and power grid operation, maintenance suggestions, etc., and synchronizes it to the data governance intelligent body.

[0103] When the small model detects abnormal parameters, it triggers the small model emergency call mechanism to increase the monitoring frequency. It pushes the data to the large model through priority marking. The large model starts the resource exclusive strategy, closes unnecessary APIs, and dedicates resources to hidden fault diagnosis and reasoning. At the same time, it calls the locally cached topology diagram and wiring diagram to complete the risk assessment of related equipment and power grid operation.

[0104] (2.4.2) Cross-validation and secondary inference: After the large model outputs diagnostic conclusions, cross-validation is performed using a digital twin and real-time data. The digital twin, based on the equipment's 3D model and physical parameters, simulates the equipment's operational performance under fault conditions, and the simulation results are compared with the real-time acquired data.

[0105] After the large model outputs the preliminary diagnostic conclusions, the system immediately starts the digital twin for simulation verification. Based on the equipment type, fault mode and current operating conditions corresponding to the fault conclusions, the digital twin loads the corresponding high-precision three-dimensional model and physical parameters of the equipment, and simulates and calculates the expected change curves and characteristic values ​​of relevant electrical and non-electrical quantities, as well as associated equipment and power grid operation under the fault state.

[0106] The system performs high-precision, multi-dimensional comparisons between the expected data sequence generated by the digital twin simulation and the real-time data collected from the same time segment obtained from the sensing agent. The comparison dimensions mainly include waveform similarity, feature amplitude deviation, consistency of change trend, and key event sequence.

[0107] If the deviations of all key features are within the preset threshold, the cross-validation is deemed successful, the diagnostic conclusion is marked as high confidence, and it is directly output to the decision-making agent; if the feature deviations exceed the threshold, it indicates that there is a non-negligible difference between the preliminary diagnostic conclusion and the actual behavior of the device, and the secondary inference optimization process is immediately triggered.

[0108] In the second inference, the large model will re-analyze by integrating real-time data, digital twin simulation results, and deviation reports. The system will increase the weight of real-time data features and guide the large model to search for cases with similar deviation patterns in the historical case library that were ultimately successfully diagnosed. The conclusions generated by the second inference need to be verified again by the aforementioned digital twin, forming a logical closed loop of diagnosis, simulation verification, and feedback optimization.

[0109] (3) The dual-loop decision-making agent realizes dual-loop decision-making through a mechanism of receiving inference conclusions and generating auxiliary suggestions through the uplink, and executing and feeding back control commands through the downlink:

[0110] (3.1) The data parsing module of the decision-making agent extracts and integrates the received fault characteristics and interpretable reasoning conclusions. The module receives the fault characteristics and interpretable reasoning conclusions of the protection and control equipment output by the analysis and judgment agent through the communication interface. The parsing module extracts and integrates the data, including format, integrity and semantic verification, to ensure the validity and consistency of the data. After the verification is passed, the core information such as the main cause of the fault, characteristic parameters, and list of affected equipment is extracted to form a standardized data set.

[0111] (3.2) An improved fault tree analysis algorithm is adopted, using the main fault cause clearly identified in the reasoning conclusion of the protection and control knowledge-enhanced large model as the root node of the logic tree. Combined with the real-time feature parameters provided by the small model, the probability of fault propagation in the secondary circuit of protection and control and the associated primary equipment is dynamically calculated, and the risk level is accurately classified according to the preset rules:

[0112] The improved fault tree analysis method includes:

[0113] Dynamic root node adaptation uses the main fault cause clearly identified in the reasoning conclusion of the knowledge-enhanced big model as the dynamic root node of the fault tree, replacing the traditional fixed root node setting, and adapting to different types of hidden fault scenarios of protection and control equipment.

[0114] Dynamic fault tree structure construction combines real-time feature parameters provided by the small model to dynamically generate a fault tree structure that includes intermediate events such as secondary circuit anomalies and protection device malfunctions. The intermediate events are updated with real-time operating conditions, covering the entire propagation path of the fault in the protection and control secondary circuit, associated primary equipment and power grid topology.

[0115] Boolean algebra is used to simplify the fault tree to obtain the minimum cut set. The fault propagation probability is calculated based on the minimum cut set method, and four risk levels are determined according to the fault propagation probability.

[0116] (3.2.1) Using the fault cause as the root node and combining real-time feature parameters, a fault tree model containing intermediate events is constructed. The intermediate events cover key nodes on the fault propagation path, including secondary circuit anomalies, protection device malfunctions, etc. The bottom-level events are the operating parameters of the protection and control equipment and the specific parameters monitored.

[0117] (3.2.2) The design uses the minimum cut set method to calculate the fault propagation probability, as shown in the following formula:

[0118]

[0119] The minimum number of cut sets, n, is obtained by simplifying the fault tree using Boolean algebra. i Let be the probability of the i-th cut set occurring. Based on the calculation results, four risk levels are defined: low risk (P < 5%), medium risk (5% ≤ P < 15%), high risk (15% ≤ P < 30%), and emergency risk (P ≥ 30%).

[0120] (3.3) Real-time linkage early warning intelligent agent, synchronously start RAG technology, in the protection and control equipment fault handling decision library, use the protection and control equipment fault type, associated equipment risk level, substation circuit and equipment / grid topology as search keywords, and through the search vector generation, matching case screening and reordering described in step (2), select 3 optimal solutions, including key information such as operation steps, target equipment ID, parameter threshold, and estimated time consumption.

[0121] (3.4) For decisions that can be remotely operated or automatically simulated in a digital twin environment, the decision-making agent will encode the operable parts into specific control command information using standardized protocols and push them to the operating agent through an encrypted channel. Before the commands are issued, they must be verified for compliance by a verification engine.

[0122] (3.4.1) The operable part of the RAG optimal solution is converted into control information, and the Protobuf 3.0 protocol encoding method is adopted. The structure includes instruction ID, target device ID, instruction type, timestamp and extensible fields;

[0123] (3.4.2) The AES-256-GCM algorithm is used to encrypt the encoded instructions. The key is stored in the hardware encryption module (HSM). The transmission channel is protected by TLS1.3 and dedicated optical fiber. Before the channel is established, the device identity must be authenticated in both directions. After the decision-making agent sends the instruction, it waits for 200ms. If no ACK is received, it will resend. If the ACK is successfully received, the "push successful" log is recorded.

[0124] (3.4.3) Compliance verification: The built-in verification engine presets protection control compliance rules, the core of which include operation logic verification, consistency with protection action logic verification, operation device permission verification, parameter range verification, topology conflict verification, etc. If the verification fails, the command interception is triggered, and an interception report containing "violation item and suggested correction value" is generated and synchronously fed back to the operation and maintenance terminal.

[0125] (3.5) After the operation agent executes the instruction, it feeds back the operation result status to the decision agent. The decision agent compares and verifies the feedback result with the expected status. If there is a deviation or execution failure, a secondary retrieval and decision optimization process is triggered to regenerate or adjust the handling plan:

[0126] (3.5.1) Feedback Data Reception and Deviation Calculation: The system receives instruction execution status, real-time parameters of protection and control equipment, and operating conditions of associated equipment and the power grid uploaded by the operational agent via the WebSocket protocol. The deviation between the actual parameters and the expected decision parameters is calculated using the following formula:

[0127]

[0128] in, This is the deviation value. y represents the actual value, and y represents the expected value. If the difference exceeds the threshold, it is determined as "instruction execution deviation," triggering a secondary optimization process.

[0129] (3.5.2) RAG secondary search and scheme update: After the deviation is triggered, the "deviation type, fault type, decision opinion, current protection and control equipment and associated primary equipment, and power grid operation status snapshot" are automatically added to the RAG search conditions, and the search process is re-executed. The focus is on matching "deviation correction" cases. The new scheme must be verified before the correction instruction can be generated and pushed back to the operation agent.

[0130] (4) The early warning intelligent agent relies on the risk information pushed by the decision intelligent agent to realize single-station early warning and cross-station linkage. When an anomaly occurs in the single-station judgment and decision, a structured report containing protection and control equipment failure, related equipment and power grid operation risks, and inspection suggestions is generated. The report is pushed to the station operation and maintenance terminal through digital twin and the early warning ledger is updated. At the same time, a cross-station linkage list is automatically generated according to the power grid topology, and the risk level and protection suggestions are synchronized to the related stations to build a cross-station collaborative early warning network.

[0131] (4.1) The early warning intelligent agent invokes the data validity verification model to verify the completeness and logical consistency of the received protection and control equipment fault data, fault type, risk propagation probability, and diagnostic conclusions of associated equipment and power grid operation status with the protection and control logic:

[0132] (4.1.1) Extract core data such as fault type code, associated equipment information list, risk assessment level, early warning time, and fault propagation path from the decrypted risk information;

[0133] (4.1.2) Ensure that the core fields are not null, and only allow the core fields such as L1-L4 to be validated for risk levels;

[0134] (4.1.3) Logical verification includes ensuring that the list of affected devices matches the power grid topology to avoid invalid device IDs, and that the fault propagation path conforms to the power system operation logic. If the verification fails, the data is marked as "abnormal" and pushed to the decision-making agent through the feedback channel to request retransmission.

[0135] (4.2) Based on the received data, the intelligent agent automatically retrieves historical operation and maintenance records and inspection procedures of related equipment, and executes single-station early warning output in a state-dependent manner:

[0136] (4.2.1) When the risk level is low or medium, the preventive early warning mode shall be activated, including:

[0137] Automatically generate standardized reports containing "Protection and Control Fault Equipment ID - Risk Assessment Level - Trend Analysis - Related Equipment Risk - Inspection Recommendations". The reports support PDF / Excel export. The trend analysis uses a linear regression algorithm to predict parameter changes. The inspection recommendations are generated by combining the equipment's historical operation and maintenance records and procedure clauses.

[0138] The location, impact range, and spread path of the fault are visualized through a digital twin model. Simultaneously, the structured report is pushed to the large screen in the operation and maintenance room of the substation and central control station to display panoramic early warning information, on the duty officer's computer and the mobile terminal of the operation and maintenance personnel.

[0139] Record the warning time, risk level, processing status, list of push terminals, and digital twin display link, and update the warning ledger in real time. The ledger supports multi-dimensional retrieval by device type, time range, and risk level.

[0140] (4.2.2) When the risk level is high or urgent, the emergency warning mode will be activated immediately. In addition to generating and pushing the detailed structured report mentioned above, the warning intelligent agent will directly link with the station monitoring system, drive the audible and visual alarm devices through standard protocols, and automatically retrieve the three-dimensional coordinate information of the fault protection cabinet and related secondary circuits. The virtual and real fusion image of the faulty equipment will be focused and displayed on the AR / VR terminal or monitoring screen, providing real-time visual support for emergency decision-making by operation and maintenance personnel.

[0141] (4.3) When the fault propagation probability calculated by the decision-making agent exceeds the preset threshold, indicating that the fault may affect other sites through electrical connections or protection coordination, the early warning agent automatically triggers cross-site linkage, automatically generates a cross-site linkage list, and realizes collaborative early warning:

[0142] (4.3.1) Fault propagation threshold setting and judgment: A preset configurable propagation threshold is used to adapt to different power grid topologies. When any state quantity exceeds the threshold, cross-site linkage is triggered.

[0143] (4.3.2) The initiating station early warning intelligent agent encapsulates the verified protection and control equipment fault information, risk assessment level, related equipment and power grid operation diagnosis conclusions into a standardized linkage request. The request body encapsulates fault information, coordination requirements and data sharing scope, etc.

[0144] (4.3.3) After receiving the request, the regional collaborative node (or master station intelligent agent) automatically analyzes the electrical connection relationship between the fault point and the upstream and downstream substations and the coordination logic of the protection configuration based on the power grid topology model, identifies the related sites that may be affected and the key protection and control equipment and primary equipment in the station, and generates a "cross-station linkage list" containing the related site ID, the list of equipment to be monitored, monitoring parameters, risk level and protection recommendations.

[0145] (4.3.4) The regional node will synchronize the list and protection recommendations to the early warning intelligent agents of each associated site through an encrypted channel. After receiving the information, each associated site will automatically start key monitoring of the designated equipment, integrate the early warning information into the local early warning ledger, and synchronize the risk level and protection recommendations to the operation and maintenance terminal of the station, thereby constructing a collaborative early warning network centered on the fault source site and covering related substations.

[0146] (5) With AR / VR interaction, standardized protocols, and dual verification as reliable guarantees, embodied operations are achieved through the operation of intelligent agents. Through AR / VR terminals, maintenance personnel can conduct full-process pre-rehearsal and verification in the virtual and real interface of the fusion equipment's three-dimensional geometric coordinates to verify the logical correctness and the impact on the secondary circuit business chain, and then perform embodied operations such as virtual pressure plate activation / deactivation, setting area switching, and soft pressure plate operation:

[0147] (5.1) Synchronize the three-dimensional geometric coordinates, board layout, terminal block wiring relationship, and real-time operating status of the protection and control equipment body to the AR / VR terminal to form a virtual and real operation interface that integrates the physical coordinates of the equipment:

[0148] (5.1.1) Import the three-dimensional geometric model of the protection and control equipment, including detailed structures such as the equipment shell, internal boards, terminal blocks, and chips, to ensure that the model matches the physical equipment 1:1;

[0149] (5.1.2) Establish the connection between the AR / VR terminal and the digital twin and physical device through the WebSocket protocol, synchronize the real-time running status of the device to the virtual and real interface, and ensure that the displayed status of the interface is completely consistent with the actual status of the physical device;

[0150] (5.1.3) Permission verification configuration: The AR / VR terminal has a built-in operation permission verification module. Operation and maintenance personnel need to pass two-factor authentication of account password and dynamic token. Only after the verification is passed can they obtain operation permissions. The permission scope is configured according to the operation and maintenance position.

[0151] (5.2) Using an AR / VR terminal, based on control information instructions, perform a full-process pre-drill and logic verification of protection control services. Call the protection control logic rule base to simulate and verify the secondary circuit state transitions and protection function activation / deactivation logic chain relationships caused by the pre-drill operation, and verify its logical correctness and impact on the secondary circuit service chain:

[0152] (5.2.1) Maintenance personnel simulate virtual pressure plate activation / deactivation, setting zone switching, soft pressure plate operation, and other specific operation processes in the virtual-real interface. The interface displays the operation steps and equipment status changes in real time.

[0153] (5.2.2) Call the protection control logic rule library, which mainly includes secondary circuit wiring logic, protection function activation / deactivation conditions, setting value activation rules, etc., to perform simulation calculations on the secondary circuit current / voltage changes, protection control equipment action logic, and associated equipment status response caused by the pre-operation simulation.

[0154] (5.2.3) If there is a logical conflict in the pre-rehearsal operation or it may cause an abnormality in the secondary loop business chain, the "verification failed" prompt will be output immediately, indicating the location and cause of the problem; if the verification passes, a "pre-rehearsal qualified" report will be generated, allowing entry into the actual operation stage.

[0155] (5.3) The intelligent agent encodes the verified embodied operation instructions into specific control command frames according to the standardized protocol, processes them through the permission and logic verification engine and AES encryption algorithm, and sends them out through the station control layer communication interface;

[0156] (5.4) During the operation, the status change signals fed back by the physical device are compared with the expected status of the virtual device in the digital twin through real-time, high-speed connection. If the deviation exceeds the safety threshold, an early warning is immediately triggered and subsequent operations are automatically blocked. At the same time, the abnormal status is fed back to the decision-making intelligent agent to start the optimization process.

[0157] (5.4.1) Collect the status change signals of the protection and control equipment in real time during the operation through the WebSocket protocol, and compare them with the expected status of the virtual device in the digital twin in milliseconds;

[0158] (5.4.2) If the deviation exceeds the safety threshold, the early warning mechanism will be triggered immediately, including AR / VR terminal pop-up alarm, station sound and light prompts, etc., automatically blocking subsequent operations, and recording abnormal status information including the deviation value of the protection and control equipment, the time of occurrence, the associated equipment and the power grid operating conditions;

[0159] (5.4.3) Feed back the abnormal status information to the decision-making intelligent agent and start the optimization process. After the abnormality is eliminated, the pre-rehearsal and verification process must be executed again before the operation command can be initiated again.

[0160] (5.4.4) After the operation is completed, the intelligent agent collects the final status data of the protection and control equipment, and after encryption processing, feeds it back to the AR / VR terminal, the decision intelligent agent and the early warning intelligent agent. The AR / VR terminal displays the "operation completed" result, the decision intelligent agent updates the decision closed-loop log, and the early warning intelligent agent updates the early warning ledger to the "handled" status.

[0161] (6) The data governance intelligent agent stores protection and control equipment and related equipment and power grid operation data according to the classification of "real-time / historical / knowledge". It matches the needs of each intelligent agent through tags, provides iterative data for special small models and protection and control knowledge-enhanced large models, supports the optimization of the collaborative framework, and combines with Figure 4 The data intelligence agent architecture has the following process:

[0162] (6.1) Based on the business logic of protection and control equipment operation, fault diagnosis and handling, establish a data classification and labeling system for protection and control business, and divide the collected multi-source heterogeneous data into three categories: real-time operation data, historical case data and domain knowledge data:

[0163] (6.1.1) Establish a distributed data storage architecture and adopt a differentiated storage solution:

[0164] The storage and sensing agent collects real-time parameters of the device, interaction instructions between agents, and real-time status feedback from the operating agent, including electrical parameters, non-electrical parameters and multi-level status monitoring information of the device body as described in step (1). It is stored in a memory database to ensure high read and write speed.

[0165] The storage device's historical parameters, fault diagnosis records, operation execution logs, early warning event ledgers, fault handling cases, etc., including complete fault characteristics, reasoning conclusions, handling plans and execution feedback records after processing in steps (2) and (3), are stored in a time-series database, with an index built by timestamp, supporting fast retrieval by time range, device type, fault type and other dimensions;

[0166] The system stores protection and control principles, equipment / device operation and maintenance specifications, electrical drawings and power grid topology, parameters of large and small models, RAG knowledge base, protection and control logic rule base, typical fault handling case base, etc. It uses a combination of relational database and vector database for storage, which ensures both the orderly management of structured knowledge and supports the efficient retrieval of unstructured knowledge.

[0167] (6.1.2) Add multi-dimensional tags to each type of data. The tag content includes data source, data type, associated device, application scenario, data level, etc., to ensure that the data is clearly identified and can be accurately matched.

[0168] (6.2) The data agent receives data requests from various agents through the API interface, and matches and extracts data from the corresponding data pool based on the business scenario tags carried in the request:

[0169] (6.2.1) Configure a dedicated API interface for each agent, with a built-in permission control module to clearly define the data access permissions for each agent:

[0170] The sensing agent can only upload data by calling the real-time running data storage interface;

[0171] The analysis and judgment of intelligent agents can call historical case data and domain knowledge data interfaces for model reasoning and retrieval;

[0172] The decision-making agent can access real-time operational data and historical case data interfaces for risk calculation and solution generation;

[0173] The early warning intelligent agent can call real-time operational data and historical case data interfaces for generating early warning information and updating ledgers;

[0174] The operating agent can only read device status information from real-time operating data to avoid data abuse and unauthorized access.

[0175] (6.2.2) After receiving a data request, the data governance intelligent agent parses the business scenario tags in the request, matches the data that meets the conditions from the corresponding data pool, and after format conversion and integrity verification, pushes the data to the requesting intelligent agent through the API interface to ensure accurate and efficient data supply.

[0176] (6.3) The newly generated data, including the status feedback after the operation of the intelligent agent in step (5), the handling report generated by the early warning intelligent agent in step (4), the status of the protection and control equipment, the associated equipment, and the power grid operation after the operation, together with their corresponding original fault data and diagnostic conclusions, are classified and stored as new historical case data after the governance process in step (6.2), forming a data pool to support the application of special models and protection and control knowledge enhancement big models:

[0177] (6.3.1) For real-time operation data, priority and low latency push to the operation agent and the early warning agent to support the real-time verification of the specific operation in step (5) and the linkage early warning in step (4); the features related to typical hidden faults of protection and control equipment such as sampling loop deviation and component aging in the real-time data are diverted and injected into the special small model monitoring data pool deployed on the equipment side for periodic monitoring and feature extraction; for historical case data, after data validity verification and model cleaning, it is classified into two categories: "exclusive fault cases in the protection and control field" and "reasoning samples that support cross-fault correlation analysis". The former is used for domain optimization and parameter fine-tuning of the special small model, and the latter is supplied to the protection and control knowledge enhancement big model in step (2) for iterative upgrading of its thinking chain reasoning ability.

[0178] (6.3.2) The newly collected data is cleaned and labeled. After the permission verification and format conversion in step (6.2), it is classified and stored in the historical case data pool.

[0179] (6.3.3) Regularly extract new data from the data pool, classify them according to fault type to generate special small model training datasets and large model iteration datasets containing cross-fault correlation analysis samples, provide data support for special small model parameter updates and protection control knowledge enhancement large model LoRA adapter fine-tuning, and provide data basis for communication protocol optimization and data transmission path adjustment of the collaborative framework, continuously improving system collaborative performance.

[0180] This invention also provides a protection and control device intelligent agent system with collaborative empowerment of large and small models, comprising:

[0181] The sensing agent is used to collect electrical and non-electrical parameters through multi-source sensors of primary and secondary equipment, and at the same time collect multi-level status monitoring information of the equipment body at the chip level, board level, and whole machine level. After completing edge-side preprocessing, it is uploaded and invalid data is removed.

[0182] The analysis and judgment of the communication connection between intelligent agents and sensing intelligent agents adopts a collaborative approach of specialized small models and overall planning of large models. Specialized small models are used to achieve specialized diagnosis and feature extraction of typical hidden faults, while large models enhanced by protection and control knowledge are used to perform interpretable reasoning according to a four-stage framework and verified by digital twins to form accurate diagnostic conclusions on the power grid operation status.

[0183] The decision-making agent and the analysis and judgment agent communicate with each other and operate in a dual closed-loop mechanism. The risk level is divided by an improved fault tree analysis algorithm, and the response plan is generated by linkage retrieval enhancement technology. The plan is pushed to the early warning agent and the control command is pushed to the operation agent.

[0184] The early warning agent communicates with the decision-making agent to generate structured reports and update the early warning ledger when a single station is abnormal. At the same time, it automatically generates a cross-station linkage list based on the power grid topology to build a cross-station collaborative early warning network.

[0185] The operational intelligent agent and the decision-making intelligent agent communicate with each other, and are used to perform embodied operations with encryption and double verification after full-process pre-playing and verification through the AR / VR interactive interface;

[0186] The data governance agent communicates with each of the aforementioned agents to store data according to real-time, historical, and knowledge categories, and matches the needs of each agent with labels to provide data support for model iteration.

[0187] This invention effectively overcomes the bottlenecks in traditional protection and control systems, such as incomplete equipment status monitoring information, reliance on manual experience for operation and maintenance analysis, insufficient precision in fault location, and weak multi-station collaboration, through the collaborative architecture of intelligent agents, thus providing core technical support for the safe and stable operation of power systems.

Claims

1. A method for intelligent agents in protection and control equipment that leverages collaborative empowerment of large and small models, characterized in that, include: (1) The sensing agent collects multi-source electrical and non-electrical parameters of primary and secondary equipment and multi-level status monitoring information of the equipment body. The data is uploaded after preprocessing by the protection and control equipment body and invalid data is removed on the intelligent agent side. (2) The intelligent agent uses a special small model to perform hidden fault diagnosis and feature extraction on the protection and control equipment side. The protection and control knowledge enhancement large model receives fault features and forms interpretable reasoning conclusions according to the four-stage framework. Combined with the digital twin and real-time data cross-validation, it forms an accurate diagnosis of the power grid operation status. (3) The decision-making agent receives the reasoning conclusions in the uplink, calculates the probability of fault propagation and classifies the risk level using the improved fault tree analysis algorithm, and pushes the enhanced retrieval to generate a disposal decision plan that matches the fault type and risk level to the early warning agent; the downlink pushes the coded and converted control information to the operation agent. (4) When the early warning intelligent agent makes an abnormal judgment decision at a single station, it generates a structured report and automatically generates a cross-station linkage list based on the power grid topology to build a collaborative early warning network; (5) After the intelligent agent performs a full-process rehearsal and verification through the interactive interface, it performs the embodied operation of encryption and double verification; (6) Data governance intelligent agents classify and store data to provide iterative data for special small models and protection and control knowledge enhancement large models.

2. The intelligent agent method for protection and control equipment according to claim 1, characterized in that, In step (1), The preprocessing algorithm is deployed at the edge of the protection and control equipment. For time-series parameters, wavelet packet transform algorithm is used for denoising and time-frequency domain feature extraction. For equipment status parameters, normalization processing is performed. The preprocessed data is encapsulated into standardized data frames and uploaded to the data governance and analysis intelligent agent. After the data from the protection and control equipment is uploaded to the intelligent agent, it is processed by an adaptive threshold filtering algorithm and a data validity verification model. The adaptive threshold filtering algorithm is based on the statistical analysis of the equipment's historical operating data and the dynamic calculation of the reasonable fluctuation range of the data under real-time operating conditions. It filters out abnormal sampled values ​​that exceed the dynamic threshold. The data validity verification model integrates business rules and multi-source data consistency comparison. Data that passes the collaborative verification is marked as high-confidence valid data and input into the subsequent analysis and judgment process.

3. The intelligent agent method for protecting and controlling equipment according to claim 1, characterized in that, Step (2) includes: (21) Construct a model collaboration architecture, with special small models deployed on the protection and control equipment side, and protection and control knowledge enhancement large models deployed on the intelligent agent side to receive diagnostic information and feature data; (22) Based on the operating mechanism, operation and maintenance specifications and fault handling cases of protection and control equipment, a domain-level dataset is constructed. The large model is fine-tuned using low-rank adaptation technology. After the fine-tuning is completed, the backbone network is frozen and the domain adaptation parameters are retained to form an inference model with a professional knowledge baseline of power protection and control. (23) Perform collaborative reasoning based on the four-stage framework. The protection and control knowledge enhancement big model follows the four-stage reasoning framework of fault feature extraction, cross-equipment correlation analysis, impact range assessment, and disposal priority ranking. It calls the retrieval enhancement module to perform logical reasoning on the received fault features and form accurate diagnostic conclusions on the power grid operation status of the associated primary equipment. (24) Establish a split-state coordination and verification algorithm. When the protection and control equipment is running normally, the small model performs periodic monitoring and the large model integrates the data. When the small model diagnoses a potential hidden fault in the equipment, it immediately triggers the large model to perform rapid reasoning, outputs the diagnostic conclusion, and performs cross-verification with the real-time data through the digital twin. If the deviation exceeds the threshold, it triggers secondary reasoning to optimize the result and outputs the diagnostic conclusion to the decision-making agent.

4. The intelligent agent method for protecting and controlling equipment according to claim 3, characterized in that, The four-stage framework includes: fault feature extraction and service mapping stage, service matching and collaborative invocation stage, cross-service correlation analysis and comprehensive judgment stage, and conclusion encapsulation and feedback traceability stage.

5. The intelligent agent method for protection and control equipment according to claim 1, characterized in that, In step (3), the decision-making agent data parsing module extracts and integrates the received fault features and interpretable reasoning conclusions; For decisions that can be remotely operated or automatically simulated in a digital twin environment, the decision-making agent will encode the operable parts into control command information through standardized protocols and push them to the operating agent through an encrypted channel. Before the command is issued, it must be verified for compliance by a verification engine. After the operating agent executes the instruction, it feeds back the operation result status to the decision-making agent. The decision-making agent compares and verifies the feedback result with the expected status. If there is a deviation or execution failure, a secondary retrieval and decision optimization process is triggered to regenerate or adjust the handling plan.

6. The intelligent agent method for protection and control equipment according to claim 1, characterized in that, The improved fault tree analysis algorithm includes: Dynamic root node adaptation uses the main fault cause clearly identified in the reasoning conclusion of the knowledge-enhanced big model as the dynamic root node of the fault tree, replacing the traditional fixed root node setting, and adapting to different types of hidden fault scenarios of protection and control equipment. Dynamic fault tree structure construction combines real-time feature parameters provided by the small model to dynamically generate the fault tree structure. Intermediate events are updated with real-time operating conditions, covering the entire propagation path of faults in the protection and control secondary circuit, associated primary equipment and power grid topology. Boolean algebra is used to simplify the fault tree to obtain the minimum cut set. The fault propagation probability is calculated based on the minimum cut set method, and four risk levels are determined according to the fault propagation probability.

7. The intelligent agent method for protection and control equipment according to claim 1, characterized in that, Step (4) includes: (41) The early warning agent calls the data validity verification model to verify the integrity of the received protection and control equipment fault data, fault type, risk propagation probability, associated primary equipment and power grid operation status and the logical consistency of protection and control. (42) Based on the received data, the intelligent agent automatically retrieves historical operation and maintenance records and inspection procedures, and executes single-station early warning output in a state-by-state manner; (43) Automatic generation of cross-site linkage list and collaborative early warning: When the probability of fault propagation calculated by the decision-making agent exceeds the preset threshold, the early warning agent automatically triggers cross-site linkage.

8. The intelligent agent method for protection and control equipment according to claim 1, characterized in that, Step (5) includes: (51) Synchronize the three-dimensional geometric coordinates, board layout, terminal block wiring relationship and real-time operating status of the protection and control equipment body to the AR / VR terminal to form a virtual and real operation interface that integrates the physical coordinates of the equipment; (52) Through the AR / VR terminal, according to the control information instructions, perform full-process pre-drill and logic verification of protection control business, call the protection control logic rule library, simulate and verify the secondary circuit state change and protection function activation / deactivation logic chain relationship caused by the pre-drill operation, and verify its logical correctness and impact on the secondary circuit business chain. (53) The intelligent agent encodes the verified embodied operation instructions into control command frames according to the standardized protocol, processes them through the permission and logic verification engine and encryption algorithm, and sends them out through the station control layer communication interface; (54) During the operation, the status change signal fed back by the physical device is compared with the expected status of the virtual device in the digital twin through real-time and high-speed connection. If the deviation exceeds the safety threshold, an early warning is immediately triggered and subsequent operations are automatically blocked. At the same time, the abnormal status is fed back to the decision-making intelligent agent to start the optimization process.

9. The intelligent agent method for protection and control equipment according to claim 1, characterized in that, Step (6) includes: (61) Based on the business logic of protection and control equipment operation, fault diagnosis and handling, establish a data classification and labeling system for protection and control business, and divide the collected multi-source heterogeneous data into real-time operation data, historical case data and domain knowledge data; (62) The data agent receives data requests from each agent through the API interface, and matches and extracts data from the corresponding data pool based on the business scenario tags carried in the request; (63) The newly generated data, such as the status feedback after the operation of the intelligent agent, the disposal report generated by the early warning intelligent agent, the protection and control equipment after the operation, the associated primary equipment, and the status of the power grid operation, along with their corresponding original fault data and diagnostic conclusions, are classified and stored as new historical case data after the governance process, forming a data pool that supports the application of special models and protection and control knowledge enhancement big models.

10. A protective control equipment intelligent agent system with collaborative empowerment of large and small models, characterized in that, include: Sensing intelligent agent: used to collect electrical and non-electrical parameters through multiple sources of sensors in primary and secondary equipment, and at the same time collect multi-level status monitoring information of the equipment body at the chip level, board level, and whole machine level; The analysis and judgment intelligent agent is connected to the sensing intelligent agent. A specialized small model is used to realize the specialized diagnosis and feature extraction of typical hidden faults. The protection and control knowledge is used to enhance the large model to perform interpretable reasoning based on a preset reasoning framework. After verification by a digital twin, a power grid operation status diagnosis conclusion is generated. Decision-making intelligent agent: It communicates with the analysis and judgment intelligent agent, classifies risk levels through an improved fault tree analysis algorithm, and generates a handling plan through linkage retrieval enhancement technology. It pushes the plan to the early warning intelligent agent and pushes control commands to the operation intelligent agent. Early warning agent: Communicatively connected to the decision-making agent, used to generate structured reports and update the early warning ledger when a single station is abnormal, and automatically generate a cross-station linkage list based on the power grid topology to build a cross-station collaborative early warning network; Operational intelligent agent: Communicatively connected to the decision-making intelligent agent, used to perform embodied operations of encryption and double verification after full-process pre-playing and verification through the AR / VR interactive interface; Data governance agent: It communicates and connects with each of the above agents to store data according to real-time, historical and knowledge categories, and matches the needs of each agent through tags.

Images