AI-based power fault diagnosis and processing method and system

Abstract

The invention relates to the technical field of electric power system automation and artificial intelligence crossing, and particularly discloses an AI-based electric power fault diagnosis and processing method and system. The system comprises a local rapid diagnosis module which is used for matching real-time fault features with a historical case library and realizing second-level response of common faults; the deep analysis and diagnosis module is used for carrying out deep reasoning by fusing physical characteristics obtained by power grid load flow calculation with an AI model when rapid matching fails, and generating a candidate fault list; the simulation verification and decision module is used for sequentially verifying the candidate list through digital twin simulation, determining final diagnosis and generating a personalized treatment scheme; and the closed-loop optimization module is used for synchronous updating and self-learning according to the field treatment feedback. According to the method, a whole-process closed loop from fault perception, intelligent analysis, physical verification, decision generation to knowledge evolution is realized, the accuracy of power fault diagnosis and the safety and efficiency of disposal are improved, and the method has a continuous self-optimization capability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power system automation and artificial intelligence, and specifically relates to a method and apparatus for intelligent diagnosis, handling scheme generation and system self-optimization of power grid faults. Background Technology

[0002] The safe and stable operation of modern power systems highly depends on the rapid and accurate diagnosis and effective handling of faults. Traditional fault diagnosis methods primarily rely on alarm information from Supervisory Control and Data Acquisition (SCADA) systems, Service Operation Reports (SOEs) of protective devices, and fault waveform data, with operators manually analyzing and making decisions. This method has significant limitations: First, the diagnostic process is highly dependent on the personal experience and technical skills of the operators. Faced with massive amounts of alarm information, it is difficult for humans to quickly correlate multi-source data and accurately locate the fault point. Especially for complex or newly occurring faults, the analysis is time-consuming and prone to misjudgment, leading to delays in fault handling.

[0003] Secondly, the development of fault handling plans lacks standardized and systematic support. Existing methods mostly rely on operating procedures and personal memory, which can easily lead to omissions, incorrect sequences, or inadequate safety measures due to negligence or lack of experience, posing safety hazards. Furthermore, excellent handling experience is difficult to accumulate and share effectively.

[0004] Furthermore, existing technologies utilize fault information in a relatively isolated manner. Most diagnostic methods are either based on simple logical rules, analyze only electrical quantity waveforms, or focus solely on protection action logic. They fail to deeply integrate and collaboratively analyze electrical transient characteristics, protection and control system behavior, power grid flow changes, and network topology, thus limiting the depth and accuracy of diagnostics.

[0005] Finally, existing systems lack self-improvement capabilities. Traditional expert systems or decision support tools have fixed knowledge bases and cannot automatically learn from actual fault cases and handling feedback. They struggle to adapt to dynamic changes in power grid structure, equipment, and operating modes, and diagnostic and handling strategies cannot be continuously optimized.

[0006] Therefore, there is an urgent need for a power fault diagnosis and treatment technology that can automatically integrate multi-source information, make intelligent reasoning decisions, and have the ability to continuously learn and evolve. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an AI-based power fault diagnosis and processing method and system to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an AI-based power fault diagnosis and processing system, comprising: The local rapid diagnosis module is used to quickly match the real-time collected fault waveform data and event sequence records with the historical fault case library to calculate the comprehensive similarity. If the highest comprehensive similarity exceeds the first preset threshold, the corresponding pre-stored diagnosis results and handling plans in the historical cases will be output. The deep analysis and diagnosis module is activated when the highest comprehensive similarity does not exceed the first preset threshold. This module is configured to: perform power flow calculations before and after the fault to generate physical features characterizing power transfer and voltage anomalies, and input the physical features, features extracted from the fault waveform data and event sequence records, and power grid topology data into an AI analysis model to obtain a list of candidate fault root causes sorted by probability of occurrence. The simulation verification and decision-making module is used to simulate each fault hypothesis in the candidate fault root cause list by calling the corresponding digital twin model in the simulation fault library in probability order, and calculate the consistency score between the simulation data and the real-time data. The first fault hypothesis whose consistency score exceeds the second preset threshold is determined as the final diagnosis conclusion, and fault handling operation instructions are generated based on the conclusion and the real-time parameters. The closed-loop optimization module is used to receive on-site feedback after the execution of the fault handling operation command, evaluate the accuracy of the diagnosis based on the actual fault cause in the feedback, and synchronously update and optimize the historical fault case library, the simulated fault library and the AI ​​analysis model based on the evaluation results.

[0009] Furthermore, when the local rapid diagnosis module calculates the comprehensive similarity, it specifically calculates the first similarity of the current fault and historical cases in terms of waveform features and the second similarity in terms of action logic features, and then performs a weighted sum of the first similarity and the second similarity, wherein the weight coefficient of the second similarity is higher than the weight coefficient of the first similarity.

[0010] Furthermore, the physical features generated by the deep analysis and diagnosis module specifically include a set of lines whose active power changes exceed a set threshold as identified through power flow calculation, and a set of buses whose voltage changes exceed a set threshold.

[0011] Furthermore, when the simulation verification and decision-making module calculates the consistency score, it specifically performs a weighted fusion calculation of the waveform similarity between the simulated waveform and the measured waveform, as well as the timing consistency between the simulated protection action time and the measured protection action time.

[0012] Furthermore, the closed-loop optimization module optimizes the AI ​​analysis model through incremental learning, specifically including: constructing new cases from the correct fault diagnosis process data, merging them with historical correct cases into an incremental training set, and fine-tuning the parameters of the AI ​​analysis model accordingly.

[0013] Furthermore, the closed-loop optimization module is also used to perform adaptive adjustment: when the same misjudgment occurs a preset number of times for a specific type of fault, the second preset threshold of the corresponding fault scenario in the simulated fault library is automatically increased.

[0014] Furthermore, each historical case in the historical fault case library is associated with a dynamic confidence score; the confidence score is dynamically updated based on the diagnostic correctness feedback after the case is matched and referenced; when performing matching, the local rapid diagnosis module prioritizes recommending historical cases with higher confidence scores.

[0015] Furthermore, the system also includes a process scheduling module, which controls the sequential execution and data transmission of the local rapid diagnosis module, the deep analysis and diagnosis module, the simulation verification and decision-making module, and the closed-loop optimization module, and sets timeout thresholds for key processing steps, and starts backup analysis paths when timeout occurs.

[0016] Furthermore, when the simulation verification and decision-making module generates the fault handling operation instruction, it specifically performs the following: retrieves successful handling cases from the historical fault case library based on the final diagnostic conclusion to obtain an operation template, and instantiates and populates the operation template by combining real-time equipment parameters and operating environment data obtained from external systems.

[0017] An AI-based power fault diagnosis and handling method, applied to the aforementioned system, includes the following steps: S1: Collect real-time fault data, extract features and quickly match them with the historical case database. If the match is successful, output the result; otherwise, proceed to S2. S2: Perform power flow analysis to generate physical characteristics, and integrate multi-source characteristics to obtain a list of candidate faults sorted by probability through an AI analysis model; S3: Perform sequential digital twin simulation verification on the candidate fault list, determine the first fault that passes the verification as the final diagnosis, and generate handling operation instructions accordingly; S4: Evaluate the accuracy of the diagnosis based on the feedback after the instruction is executed, and use the feedback information to drive the closed-loop optimization of the historical case library, the simulated fault library, and the AI ​​analysis model.

[0018] The technical effects and advantages of this invention are as follows: This approach enhances the accuracy of fault diagnosis and the reliability of decision-making. By deeply integrating physical features, electrical waveform features, protection action logic features, and power grid topology data extracted from power grid power flow calculations, and utilizing AI models for reasoning, the limitations of single information sources are overcome. Furthermore, digital twin simulation is introduced to rigorously verify the AI ​​reasoning results physically, transforming probabilistic outputs into empirically validated, high-confidence conclusions. This improves the accuracy of locating complex faults and provides a reliable basis for safe handling.

[0019] It achieves rapid emergency response and precise handling plans: A rapid matching mechanism based on a historical case database enables the system to respond to and invoke contingency plans within seconds for a large number of known typical faults. The handling plan generation process is not a simple copy; instead, it retrieves similar successful case templates and automatically binds them to real-time equipment parameters, protection settings, and environmental data to generate specific and unambiguous standardized operation tickets. This improves response speed while ensuring the safety and accuracy of operations.

[0020] A closed-loop mechanism for continuous self-optimization and knowledge accumulation has been constructed: the system transforms the feedback from each real-world action into learning resources through a closed-loop optimization module. This mechanism drives the dynamic updating and confidence management of the historical case library, the incremental learning of the AI ​​diagnostic model, and the parameter calibration of the simulation model. It can also adaptively adjust verification strategies for recurring specific misjudgment patterns. This allows the system's diagnostic and decision-making capabilities to continuously evolve autonomously through operational practice.

[0021] The system's overall robustness and service availability have been enhanced: through the built-in process scheduling and monitoring mechanism, when core computing modules (such as deep AI analysis and high-precision simulation) time out or malfunction, the system can automatically degrade to a rule-based fast backup analysis path, ensuring that valuable decision support can be provided under any circumstances, avoiding system-wide failure due to local module failure, and meeting the stringent requirements for high availability in power production control. Attached Figure Description

[0022] Figure 1 This is a flowchart of the overall system processing of the present invention.

[0023] Figure 2 This is a flowchart of the rapid diagnosis and in-depth analysis process of the present invention.

[0024] Figure 3 This is a flowchart of the verification and decision-making branches of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1 As attached Figures 1 to 3 The following is a detailed implementation of an AI-based power fault diagnosis and handling method and system: This embodiment describes a complete work cycle of the system from the occurrence of a fault to the completion of processing and knowledge updating. The system software is deployed on the security zone III server of the power dispatch data network, and establishes data interfaces with the dispatch automation system, fault recording system, protection information management system, production management system, and meteorological information system located in security zones I / II through a forward isolation device.

[0027] Step S100: Rapid matching diagnosis based on the historical case library.

[0028] When the scheduling automation system generates a record containing a precise timestamp When a fault alarm is triggered by the associated device identifier, the system initiates this process.

[0029] S101: Multi-source fault data acquisition and standardization.

[0030] The system automatically executes the data acquisition program based on the device identifier in the alarm. The acquisition includes: 1. Send a request to the fault recording system to obtain the time window. The system contains all waveform recordings (COMTRADE format) electrically associated with the device. Inputs are the timestamps and device identifiers from the alarm signals; processing involves parsing the waveform recordings; the output is the obtained three-phase voltage. and three-phase current The sampling sequence (a typical sampling rate) ).

[0031] 2. Send a request to the dispatch automation system to retrieve the sequential records of all relevant events within the same time window. Inputs are the same timestamp and device identifier; processing involves querying the database; outputs a list of event types such as protection activation, protection output, and circuit breaker opening / closing, along with their timestamps accurate to milliseconds.

[0032] Final output: Time-aligned and format-normalized fault recording data and event sequence .

[0033] S102: Fault feature extraction and fast matching calculation.

[0034] The historical case matching module processes the collected data as follows: 1. Waveform Feature Extraction: The well-known abrupt change initiation algorithm is used to extract waveform features from the current sequence. The starting time of fault detection Extract The first power frequency cycle after ( The fault phase current data is used. The full-cycle Fourier algorithm is applied to calculate: Fundamental ( RMS value of current Unit: Ampere (A).

[0035] 2nd, 3rd, and 5th harmonic content ,in For the first RMS value of subharmonic current .

[0036] Constructing waveform feature vectors: .

[0037] 2. Action Logic Feature Extraction: Analysis ,extract All protection and switching events within the time window. Encode event types with numeric IDs and sort them by occurrence time to form an event type sequence. Simultaneously record the time interval between adjacent events. Together they constitute the action logic feature vector. .

[0038] 3. Comprehensive similarity calculation: For the first case in the historical case database... One case, whose pre-stored features are and Calculate the degree of match between the real-time fault and this case: Waveform similarity ,in This represents the Euclidean distance between vectors.

[0039] Action logic similarity By calculating the event sequence and The normalized edit distance is then mapped.

[0040] Overall similarity score Among them, the weighting coefficient and satisfy This is used to adjust the contribution of different characteristics. Given that protection action logic typically has higher determinism in fault location, it is usually set... Within the range of 0.6 to 0.8, The corresponding value is in the range of 0.4 to 0.2. As an example, this embodiment uses... .

[0041] Output: A set of comprehensive similarity scores .

[0042] S103: Matching decision and process branch.

[0043] Preset a similarity threshold Its value ranges from [0.80, 0.90]. Calculate and find the maximum score. .

[0044] like (This embodiment takes) If the match is successful, then it is considered a successful quick match. The output will be the same as the given value. The corresponding historical cases contain pre-stored diagnostic conclusions and treatment procedures. The process is terminated.

[0045] like If the match fails, it is considered a quick match failure. Output the "match failed" flag and any associated information. , , , The complete data packet. The process proceeds to step S200.

[0046] Furthermore, as a preferred implementation, each case in the historical case database is associated with a dynamic confidence score. The initial value is 0.5, and the value range is [0, 1].

[0047] The rating is dynamically updated based on feedback after a case is cited: when a case is successfully matched and subsequent feedback confirms the diagnosis is correct. Among them, the increment Values ​​between 0.05 and 0.15 are acceptable (e.g., 0.1); if a diagnostic error is reported after a match, The attenuation factor Values ​​between 0.5 and 0.8 are acceptable (e.g., 0.7).

[0048] In the matching sorting of S102, a modified priority score can be used. Sort, where This is the gain factor, which can be configured in the range of 0.1 to 0.3 (e.g., 0.2).

[0049] Step S200: Deep diagnostics that integrates physical mechanisms and data-driven approaches.

[0050] The system initiates this phase when a quick match fails.

[0051] S201: Power Flow Comparison Analysis and Physical Feature Extraction.

[0052] The power flow calculation engine in the fusion diagnostics module is invoked: 1. Data Input: Obtain network-wide telemetry data at two time points from the dispatch automation system: the steady-state section before the fault (time point). (and the first available steady-state section after the fault. Data includes voltage amplitudes at all bus nodes.) Phase angle and the active power of all branches. reactive power .

[0053] 2. Power Flow Calculation and Comparison: Perform AC power flow calculations for both cross-sections. For each branch... Calculate the change in active power (Unit: megawatt, MW). For each node Calculate the voltage change (Unit: per unit, pu).

[0054] 3. Generation of physical feature vectors: Identify branches with significant power shifts: Set a power change threshold. This threshold can be set according to the voltage level: for For lines of 2000 and above, a typical value can be taken as follows: to ;for And the following routes, can be selected to It can also be set as the line's rated capacity. of to All satisfied branch road Included in the set .

[0055] Identify nodes with significant voltage drops: Set a voltage change threshold. All satisfied nodes Included in the set .

[0056] Output: Physical feature vector It is used to characterize the abnormal state of the power grid caused by a fault, and can be represented as follows: .

[0057] S202: Multi-source information fusion and AI model inference.

[0058] 1. Feature fusion: Combining waveform features (dimension) Action logic characteristics (dimension) ), vectorized physical features and flattened topological adjacency matrix The data are concatenated and merged to form a unified input feature vector. Its total dimension is .

[0059] in, The operation will transform the physical feature vector set in and The changes and their associated variables are expanded and concatenated into a one-dimensional numerical vector; The operation will change the adjacency matrix Expand it into a one-dimensional vector by row.

[0060] 2. AI Model Inference: Input a pre-trained root cause inference model (e.g., a graph neural network). The model outputs the predicted probabilities of all predefined fault combinations. ( ,and ).

[0061] 3. Generate a candidate list: Set a minimum probability threshold This is a configurable parameter that can be adjusted within the range of 0.01 to 0.05. In this embodiment, we take... Filter out all The combination, and according to Sort in descending order.

[0062] Output: List of candidate root causes of failure .in This describes the specific fault assumptions.

[0063] Furthermore, as an optional implementation, the training data for the fault root cause reasoning model includes historical real-world fault cases and simulated cases generated based on power system simulation software.

[0064] The simulation case library covers different voltage levels (such as...) , , , ), fault type (single-phase grounding, phase-to-phase short circuit, etc.), transition resistance ( The simulation cases are designed to enhance the model's generalization ability by considering the system's operating conditions (e.g., -300 ohms) and electromagnetic transient simulation methods. The simulation examples are generated using power system electromagnetic transient simulation software (e.g., PSCAD / EMTDC or ATP-EMTP).

[0065] Different fault locations (along the line) can be automatically set via scripts. to Parameters such as fault type, transition resistance, fault initial phase angle, and system startup mode are used to run simulations in batches and collect electrical quantities and event data, thereby building a large-scale standardized simulation case library.

[0066] Step S300: Sequential simulation verification and diagnostic decision based on digital twin.

[0067] S301: Sequential digital twin simulation.

[0068] Simulation verification and disposal generation module according to Work in list order. For the current candidate hypothesis... : 1. Call from the simulated fault library and The corresponding digital twin simulation model.

[0069] 2. The pre-fault steady-state power flow obtained in S201 is used as the initial state of the simulation network.

[0070] 3. Set up the simulation model with... Describe a consistent fault and perform an electromagnetic transient simulation (duration). ).

[0071] Output: Virtual waveform data generated by simulation and virtual event sequences .

[0072] S302: Consistency Quantitative Assessment.

[0073] Compare the simulation output with the measured data collected by S101: 1. Waveform similarity calculation: Alignment and The fault initiation point. The first cycle after the fault is captured ( The normalized cross-correlation coefficient between the current waveforms of the two devices is calculated as the waveform similarity. .

[0074] 2. Time series fit calculation: Extraction and The same critical moment of protection and Calculate the time difference Timing consistency pass Calculation. Wherein, the time tolerance constant... This reflects the tolerance for protection action time deviation, based on the relay protection action speed and communication delay. Desirable to This embodiment takes .

[0075] 3. Overall Consistency Score: Calculated Weighting coefficients To adjust the two contributions, this embodiment takes... .

[0076] Output: Current candidate hypothesis Consistency score (Value range [0, 1]).

[0077] S303: Validate the decision.

[0078] Preset verification pass threshold The value range is [0.85, 0.95].

[0079] like (This embodiment takes) If so, then determine the hypothesis. Validated. Output As a final diagnostic conclusion The verification process has terminated, proceeding to step S400.

[0080] like If the result is not found, the verification fails. Return to the next candidate in the S301 verification list. If all candidates in the list fail verification, output a "Verification Failed" signal.

[0081] Step S400: Generate personalized treatment plan.

[0082] S401: Historical successful case retrieval and template extraction.

[0083] by Using the core search term, we searched the historical case database for all similar cases that were successfully handled. Through natural language processing, we analyzed the handling report text, extracted common operational steps, and created a structured operation template containing variable placeholders (such as {{device ID}}).

[0084] Output: Standardized operation template.

[0085] S402: Real-time parameter binding and operation ticket instantiation.

[0086] The system automatically obtains real-time parameters through an interface: 1. Obtain the precise dispatch numbers of faulty equipment and related switches and disconnectors from the production management system.

[0087] 2. Obtain the current setting parameters of the relevant protection devices from the protection information management system.

[0088] 3. Obtain real-time weather data for the faulty area from the meteorological information system.

[0089] Replace all placeholders in the template with the obtained parameter values.

[0090] Output: An instantiated "Standardized Operation Ticket for Fault Handling".

[0091] Step S500: On-site handling and result feedback.

[0092] After performing on-site operations based on the operation ticket, maintenance personnel submit structured feedback through the system terminal.

[0093] Output: Feedback information It must contain the following fields: (1) Actual Fault (actual cause of failure confirmed on-site); (2) Outcome Evaluation.

[0094] Step S600: Closed-loop self-optimization.

[0095] S601: Diagnostic accuracy assessment and case structuring.

[0096] Comparison and If they match, mark Flag = "Correct"; otherwise, mark Flag = "Incorrect". Integrate all data from this fault.

[0097] Output: A new structured case study with labeled flags .

[0098] S602: Multiple knowledge bases are updated synchronously.

[0099] Will Store in the historical fault case database. The data is used to create or calibrate corresponding digital twin simulation models in the simulated fault library.

[0100] Output: The updated knowledge base.

[0101] S603: Model optimization and adaptive parameter adjustment.

[0102] Optimize based on Flag tags: 1. Incremental model learning (when Flag="Correct"): [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] This data is combined with other recent (e.g., within the past 30 days) correct diagnostic cases to form an incremental training dataset. .

[0103] Specifically, cases that are similar in feature space to those in the historical case library can be selected. Most similar One (e.g.) Historical correct cases, and Together constitute The fault root cause reasoning model is fine-tuned using an elastic weight consolidation algorithm. The input is: parameters of the old model. Importance matrix Incremental dataset The output is the updated model parameters. .

[0104] 2. Verify adaptive threshold adjustment (when Flag = "Error", and it is a specific pattern of repeated error): If a "high resistance grounding" type fault is detected, the recent consecutive If the "bus fault" assumption is incorrectly accepted, the verification threshold of all "bus high resistance grounding" scenarios in the simulated fault database will be automatically adjusted. Increase by an increment (For example ),Right now .

[0105] S604: Process monitoring and downgrade handling mechanism.

[0106] The process scheduling module sets the maximum allowable processing time for steps S200 and S300 respectively. and (This can be adjusted according to server performance) Adjustments are made within 30 seconds.

[0107] If S200 times out, the backup rule engine will be activated, directly based on... A diagnostic conclusion is generated by applying preset simplified rules. The simplified rules include, for example, "if the power anomaly transfer set..." The data contains only one line, and the node of the busbar connected to this line (called the head busbar) is located in the voltage anomaly set. If the result is "indicated, then the line is faulty." The generated conclusion will be labeled "low confidence - timeout," and the process will jump to S400.

[0108] If S300 times out, then adopt directly. The hypothesis with the highest probability is taken as The system will display a message indicating that the simulation has not been fully verified, and the process will then redirect to S400.

[0109] Through the cyclical execution of the above steps S100 to S600, the system realizes intelligent diagnosis, handling and continuous self-optimization of power faults.

[0110] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change. Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based power fault diagnosis and processing system, characterized in that, include: The local rapid diagnosis module is used to quickly match the real-time collected fault waveform data and event sequence records with the historical fault case library to calculate the comprehensive similarity. If the highest comprehensive similarity exceeds the first preset threshold, the corresponding pre-stored diagnosis results and handling plans in the historical cases will be output. The deep analysis and diagnosis module is activated when the highest comprehensive similarity does not exceed the first preset threshold. This module is configured to: perform power flow calculations before and after the fault to generate physical features characterizing power transfer and voltage anomalies, and input the physical features, features extracted from the fault waveform data and event sequence records, and power grid topology data into an AI analysis model to obtain a list of candidate fault root causes sorted by probability of occurrence. The simulation verification and decision-making module is used to simulate each fault hypothesis in the candidate fault root cause list by calling the corresponding digital twin model in the simulation fault library in probability order, and calculate the consistency score between the simulation data and the real-time data. The first fault hypothesis whose consistency score exceeds the second preset threshold is determined as the final diagnosis conclusion, and fault handling operation instructions are generated based on the conclusion and the real-time parameters. The closed-loop optimization module is used to receive on-site feedback after the execution of the fault handling operation command, evaluate the accuracy of the diagnosis based on the actual fault cause in the feedback, and synchronously update and optimize the historical fault case library, the simulated fault library and the AI ​​analysis model based on the evaluation results.

2. The AI-based power fault diagnosis and processing system according to claim 1, characterized in that, When the local rapid diagnosis module calculates the comprehensive similarity, it specifically calculates the first similarity of the current fault and historical cases in terms of waveform features and the second similarity in terms of action logic features, and then performs a weighted sum of the first similarity and the second similarity, wherein the weight coefficient of the second similarity is higher than the weight coefficient of the first similarity.

3. The AI-based power fault diagnosis and processing system according to claim 1, characterized in that, The physical features generated by the deep analysis and diagnosis module specifically include a set of lines whose active power changes exceed a set threshold as identified by power flow calculation, and a set of buses whose voltage changes exceed a set threshold.

4. The AI-based power fault diagnosis and processing system according to claim 1, characterized in that, When the simulation verification and decision-making module calculates the consistency score, it specifically performs a weighted fusion calculation on the waveform similarity between the simulated waveform and the measured waveform, as well as the timing consistency between the simulated protection action time and the measured protection action time.

5. The AI-based power fault diagnosis and processing system according to claim 1, characterized in that, The closed-loop optimization module optimizes the AI ​​analysis model through incremental learning, specifically by constructing new cases from the correct fault diagnosis data of the current process, merging them with historical correct cases into an incremental training set, and then fine-tuning the parameters of the AI ​​analysis model accordingly.

6. The AI-based power fault diagnosis and processing system according to claim 1, characterized in that, The closed-loop optimization module is also used to perform adaptive adjustment: when the same misjudgment occurs a preset number of times for a specific type of fault, the second preset threshold of the corresponding fault scenario in the simulated fault library is automatically increased.

7. The AI-based power fault diagnosis and processing system according to claim 1, characterized in that, Each historical case in the historical fault case library is associated with a dynamic confidence score; the confidence score is dynamically updated based on the diagnostic correctness feedback after the case is matched and referenced; when performing matching, the local rapid diagnosis module prioritizes recommending historical cases with higher confidence scores.

8. The AI-based power fault diagnosis and processing system according to claim 1, characterized in that, The system also includes a process scheduling module, which controls the sequential execution and data transmission of the local rapid diagnosis module, the deep analysis and diagnosis module, the simulation verification and decision-making module, and the closed-loop optimization module, and sets timeout thresholds for key processing steps, and starts backup analysis paths when timeout occurs.

9. The AI-based power fault diagnosis and processing system according to claim 1, characterized in that, When the simulation verification and decision-making module generates the fault handling operation instruction, it specifically performs the following: retrieves successful handling cases from the historical fault case library based on the final diagnosis conclusion to obtain an operation template, and instantiates and populates the operation template by combining real-time equipment parameters and operating environment data obtained from external systems.

10. An AI-based method for power fault diagnosis and handling, characterized in that, Applied to the system as described in any one of claims 1-9, the method comprises the following steps: S1: Collect real-time fault data, extract features and quickly match them with the historical case database. If the match is successful, output the result; otherwise, proceed to S2. S2: Perform power flow analysis to generate physical characteristics, and integrate multi-source characteristics to obtain a list of candidate faults sorted by probability through an AI analysis model; S3: Perform sequential digital twin simulation verification on the candidate fault list, determine the first fault that passes the verification as the final diagnosis, and generate handling operation instructions accordingly; S4: Evaluate the accuracy of the diagnosis based on the feedback after the instruction is executed, and use the feedback information to drive the closed-loop optimization of the historical case library, the simulated fault library, and the AI ​​analysis model.

Images