A method and system for generating training corpus of large language model based on power operation event reconstruction and knowledge enhancement

Abstract

The application discloses a kind of big language model training corpus generation method and system based on power operation event reconstruction and knowledge enhancement.The multiple-source power operation data is obtained, and the electrical objects and time correlation are established by time reference unification and equipment identification mapping;According to the candidate event trigger record determined by the change of device state, the change of operating parameter, the protection action or alarm trigger, search associated operation data and identify event boundary;Reconstruct the power operation event unit based on the state, event, operation, operating parameter and time sequence relationship within the event boundary;Conduct consistency check using electrical topology relationship, state transition rule, operation time sequence rule and operating parameter constraint, and construct event fact anchor point combined with data source effectiveness;According to the training task type, select event fact anchor point and corresponding knowledge constraint to generate training sample, and establish the association between training sample, event fact anchor point, event unit and original operation data, form traceable big language model training corpus, improve the fact consistency and traceability of training corpus.

Description

Technical Field

[0001] This invention belongs to the technical fields of artificial intelligence, power system data processing, and large language model training data construction. Specifically, it relates to a method and system for generating large language model training corpus based on power operation event reconstruction and knowledge enhancement. It is applicable to the construction of large language model training corpus in power dispatching, equipment operation monitoring, fault analysis, operation and other power business scenarios. Background Technology

[0002] With the development of large language model technology, professional large language models for the power industry are gradually being applied to business scenarios such as power grid operation analysis, equipment status assessment, fault assistance, dispatch knowledge Q&A, and operation decision support. The professionalism, authenticity, and relevance of the training corpus to the actual power operation process directly affect the power industry's large language model's ability to understand and reason about equipment status, operating events, operational behaviors, parameter changes, and time sequence relationships.

[0003] Power systems generate a large amount of heterogeneous operational data during actual operation, including equipment operating status data, telemetry data, remote signaling data, protection action records, alarm records, fault waveform data, dispatching operation logs, and operation records. This data originates from different business systems, monitoring devices, and operational processes, and varies in data structure, equipment identification, sampling period, and time accuracy. The same power operation event is typically recorded across multiple data sources; for example, protection actions, circuit breaker status changes, operating parameter changes, and dispatching operations may be stored in different systems, making it difficult to directly form training samples that can fully describe the actual operational process.

[0004] Existing methods for constructing training corpora for large language models typically rely on textual data such as regulations, technical documents, knowledge bases, and historical cases. Question-answer pairs, instruction data, or inference training data are generated through text segmentation, entity relation extraction, question-answer template matching, or large language model generation. Some methods also utilize power operation data, equipment monitoring data, or historical fault data to construct training datasets.

[0005] However, for multi-source asynchronous data from actual power operation processes, simply splicing data according to time windows or directly converting it into natural language text makes it difficult to accurately determine whether different data records belong to the same real operation event. For example, multiple device status changes, alarm messages, and operation records may exist simultaneously within adjacent time periods, and simply relying on time proximity can easily lead to incorrect aggregation of different operation events; furthermore, protection actions, switch changes, and operating parameter responses within the same operation event may exhibit time differences due to the sampling period and recording delay of different systems.

[0006] Furthermore, directly generating training questions and answers or inference samples based on discrete operation records may result in inconsistencies between the equipment status, event causal relationships, operation timing, or changes in operating parameters in the training samples and the actual operation process. Especially in specialized tasks such as fault analysis, status judgment, and operation inference, if the target output of the training samples lacks verifiable operational facts, it is easy to introduce content that does not conform to the actual operating state or professional rules into the training data.

[0007] Therefore, there is a need for a method that can identify data records belonging to the same operation process from multi-source asynchronous power operation data, reconstruct the state evolution process of real power operation events, and generate a large language model training corpus based on operation facts that have been verified by power professional knowledge, thereby improving the consistency and traceability between the training corpus and the actual power operation process. Summary of the Invention

[0008] I. Technical problems to be solved

[0009] This invention aims to solve the problems in the construction of training corpora for existing power large language models, such as the difficulty in accurately associating multi-source asynchronous running data with real running events, the difficulty in representing the complete equipment state evolution process by discrete running records, and the lack of real running fact constraints in the target output of training samples.

[0010] To this end, this invention unifies the time base and maps equipment identifiers for different power operation data sources, identifies event boundaries by combining electrical topology relationships, equipment status changes, operating parameter changes, and time relationships, and aggregates operation data belonging to the same operation process; further reconstructs the pre-event state, triggering event, operation behavior, post-event state, operating parameter response, and time sequence relationship, and determines verifiable event fact anchors through power professional knowledge constraints, selects corresponding event fact anchors to generate training samples according to different large language model training tasks, thereby establishing a traceable generation process from raw operation data to training corpus.

[0011] II. Technical Solution

[0012] To address the aforementioned technical problems, the first aspect of this invention provides a method for generating training corpus for a large language model based on power operation event reconstruction and knowledge enhancement, comprising the following steps.

[0013] Step 1: Acquisition of Multi-Source Power Operation Data

[0014] Obtain raw operational data from multiple power operation data sources.

[0015] The power operation data source may include at least two of the following: dispatch automation system, energy management system, power distribution management system, substation monitoring system, relay protection device, synchronous phasor measurement device, and power business management system.

[0016] The original operating data may include equipment operating status data, operating parameter data, event logs, and operation logs; further, it may include at least two of the following: telemetry data, remote signaling data, protection action logs, alarm logs, fault waveform characteristic data, scheduling operation logs, and operation logs.

[0017] Different data sources can have different data structures, device encoding methods, sampling periods, and time precision. This invention does not require all data sources to adopt a uniform data storage format beforehand.

[0018] Step 2: Time base unification and electrical object association

[0019] The time base is unified and the equipment identifier is mapped for raw operating data from different power operation data sources.

[0020] A unified time base is used to convert the time recorded from different data sources to a unified time base. For data with different sampling periods, the original timestamps can be retained, and a correspondence can be established on a unified timeline.

[0021] Device identification mapping is used to map device codes, device names, or business identifiers representing the same electrical equipment or electrical object in different systems to a unified electrical object identifier.

[0022] In one implementation, a device identifier mapping relationship can be pre-established: Mid: IDs -> IDu Here, IDs represents the original device identifier from different data sources, and IDu represents the corresponding unified electrical object identifier.

[0023] Based on the unified equipment identification, electrical topology, and data time relationships, the electrical object associations and time associations between different original operating data are determined.

[0024] The electrical object association relationship includes the association between data records corresponding to the same electrical object, and it can also include the association between data records corresponding to multiple electrical objects that have direct or indirect electrical topological connections.

[0025] Step 3: Identification of Power Operation Event Boundaries

[0026] Candidate event trigger records are determined from the original operating data based on at least one of the following: changes in equipment operating status, changes in operating parameters, protection actions, and alarm triggers.

[0027] The candidate event trigger record may include at least one of the following: a record of a preset change in device status, a record of a change in operating parameters that meets a preset change condition, a protection action record, and an alarm trigger record.

[0028] Using the time point and electrical object corresponding to the candidate event trigger record as the event search benchmark, the system searches for related running data forward and backward within a preset time range.

[0029] During the search process, other electrical objects associated with the electrical object corresponding to the candidate event trigger record are determined based on the electrical topology relationship, and the running data corresponding to the other electrical objects are obtained.

[0030] For the associated operational data obtained from the search, the correlation degree of the event is determined based on the time interval between it and the candidate event trigger record, the electrical topology distance, the correlation degree of equipment status change, and the correlation degree of operational parameter change.

[0031] In one implementation, the event correlation degree is expressed as: Ri = β1Rt,i + β2Ro,i + β3Rs,i + β4Rp,i Where Ri represents the degree of event correlation between the i-th associated operational data and the candidate event trigger record; Rt,i represents the degree of time correlation; Ro,i represents the degree of electrical topology correlation; Rs,i represents the degree of correlation of equipment status changes; Rp,i represents the degree of correlation of operational parameter changes; β1 to β4 are the corresponding correlation coefficients.

[0032] The degree of time correlation Rt,i is determined based on the time interval between the i-th correlated running data and the candidate event trigger record. In one implementation, it can be: Rt,i = exp(-Δti / τt) Where Δti represents the time interval between the i-th associated running data and the candidate event trigger record, and τt represents the preset time decay parameter; the smaller the time interval, the larger Rt,i.

[0033] The electrical topology correlation degree Ro,i is determined based on the electrical topology connection distance between the electrical object corresponding to the i-th associated running data and the electrical object corresponding to the candidate event triggering record. In one implementation, it can be: Ro,i = 1 / (1 + di) Where di represents the topological connection distance between two electrical objects, which can be determined according to the number of shortest connection edges between the two electrical objects in the electrical topology diagram; the smaller the topological connection distance, the larger Ro,i is.

[0034] The correlation degree Rs,i of the device state change is determined based on the matching between the device state change in the i-th associated operational data and the preset state transition relationship corresponding to the candidate event. In one embodiment, when the device state change satisfies the corresponding preset state transition relationship, Rs,i takes a first preset value; when the preset state transition relationship is partially satisfied, Rs,i takes a second preset value; when the preset state transition relationship is not satisfied, Rs,i takes a third preset value, wherein the first preset value is greater than the second preset value, and the second preset value is greater than the third preset value.

[0035] The correlation degree Rp,i of the operating parameter changes is determined based on the direction and magnitude of the changes in the operating parameter corresponding to the i-th associated operating data before and after the event, as well as the matching relationship with the corresponding device state change. In one embodiment, when both the direction and magnitude of the operating parameter changes meet the preset parameter change conditions under the corresponding device state change, Rp,i takes the first parameter correlation value; when only the preset parameter change conditions are partially met, Rp,i takes the second parameter correlation value; when the preset parameter change conditions are not met, Rp,i takes the third parameter correlation value, wherein the first parameter correlation value is greater than the second parameter correlation value, and the second parameter correlation value is greater than the third parameter correlation value.

[0036] In one optional implementation, β1, β2, β3, and β4 are normalized, and the normalized β1, β2, β3, and β4 are all non-negative numbers and satisfy the following: β1 + β2 + β3 + β4 = 1 When the correlation between events meets the preset correlation conditions, the corresponding operational data will be grouped into the same candidate power operation event.

[0037] The event boundary of the current power operation event can be determined when at least one of the following conditions is met: the time interval between adjacent associated operation data exceeds a preset time threshold; the associated electrical object exceeds the preset electrical topology association range; or the equipment status reaches a preset event termination state.

[0038] The associated operational data belonging to the same power operation process within the event boundary are aggregated to form an event-related data set.

[0039] Step 4: Reconfiguration of Power Operation Event Unit Based on the changes in device status, event records, operation records, changes in operating parameters, and their time sequence in the event-related data set, the corresponding power operation process is reconstructed.

[0040] For the k-th power operation event, its power operation event unit can be represented as: Uk = {Ok, Sk^pre, Ek, Ak, Sk^post, Pk, Tk, Rk} Wherein, Uk represents the k-th power operation event unit; Ok represents the corresponding electrical object; Sk^pre represents the state before the event occurs; Ek represents the triggering event; Ak represents the operation behavior; Sk^post represents the state after the event occurs; Pk represents the operation parameter response; Tk represents the temporal relationship between event elements; and Rk represents the association relationship between related electrical objects and event elements.

[0041] The pre-event state is determined based on the valid device state of the corresponding electrical object before the triggering event occurs within the event boundary, and the post-event state is determined based on the valid device state of the corresponding electrical object after the triggering event and associated operations.

[0042] The operating parameter response is determined based on the changes in corresponding operating parameters within a preset time range before and after the event, and may include changes in current, voltage, active power, reactive power, frequency, and other parameters characterizing the operating status of power equipment.

[0043] Establish the correlation between triggering events and equipment state transitions, between triggering events and operating parameter responses, and between operational behaviors and post-event states, so that power operation event units can represent the state evolution process in actual operation, rather than just representing mutually independent data records.

[0044] Step 5: Knowledge Constraint Verification and Event Fact Anchor Point Construction Obtain the electrical topology relationship, equipment state transition rules, operation timing rules and operating parameter constraints corresponding to the power operation event unit.

[0045] Determine whether the pre-defined electrical connection relationship is satisfied between the associated electrical objects in the event unit based on the electrical topology relationship.

[0046] Based on the device state transition rules, determine whether the preset state transition conditions are met between the pre-event state, the triggered event, and / or the operation behavior and the post-event state.

[0047] For example, circuit breaker operation events can be verified according to the following relationship: S^pre = Closed, A = Open -> S^post = Open Where S^pre represents the device state before the operation, A represents the operation behavior, and S^post represents the device state after the operation. When the circuit breaker is in the closed position before the operation and a tripping operation is performed, the circuit breaker state after the operation should be in the open position.

[0048] The timing of triggering events, equipment status changes, operational behaviors, and parameter responses is determined according to the operation timing rules to determine whether the time sequence meets the preset timing conditions of the corresponding power operation process.

[0049] Based on the constraints of operating parameters, determine whether the changes in operating parameters before and after the event are within the range allowed by the corresponding equipment type and operating status.

[0050] For event elements that satisfy the corresponding knowledge constraints and the relationships between event elements, their source data identifier, source validity identifier, event unit identifier, element type, element value, and corresponding constraint identifier are recorded to form a fact record. The source validity identifier is used to characterize whether the corresponding original operational data originates from a pre-registered power operation data source and whether it has a traceable data source.

[0051] In one implementation, the original operational data is matched with pre-established valid data source information based on the data source identifier, data interface identifier, and / or acquisition device identifier corresponding to the original operational data. When a match is successful, the corresponding original operational data is determined to be from a valid source, and a corresponding source validity identifier is set. When the element value in the fact record can correspond to the original operational data from a valid source and satisfies the corresponding knowledge constraints, the fact record is determined as the event fact anchor point.

[0052] An event fact anchor can be represented as: Fj = {IDraw, Vsource, IDevent, Typej, Valuej, IDconstraint} Where Fj represents the j-th event fact anchor point; IDraw represents the source original operation data identifier; Vsource represents the source validity identifier; IDevent represents the power operation event unit identifier; Typej represents the fact element type; Valuej represents the corresponding fact value; and IDconstraint represents the knowledge constraint identifier used to verify the fact.

[0053] The event fact anchors may include state fact anchors, event fact anchors, operation fact anchors, parameter fact anchors, and time sequence fact anchors.

[0054] Event elements or relationships that do not meet the corresponding knowledge constraints will not be used as event fact anchors for generating the target output content of training samples.

[0055] Step 6: Training Task Mapping and Training Sample Generation The type of training task is determined based on the application requirements of the training corpus for the large language model to be built.

[0056] The training task types may include at least one of state judgment, event causal analysis, temporal reasoning, runtime parameter interpretation, and operational reasoning.

[0057] Pre-establish task mapping relationships between training task types and target event element types, as well as target association relationship types: Taskq -> {Typeq, Relq} Where Taskq represents the training task of type q, Typeq represents the target event element type required for the training task, and Relq represents the target association type required for the training task.

[0058] Based on the training task type of the training sample to be generated, query the task mapping relationship, and select the event fact anchor point that matches the target event element type and target association type from the corresponding power operation event unit.

[0059] Training sample input content is generated based on some event elements and their relationships in the selected event fact anchors. The target output content of the training sample is determined based on the target event elements used to answer the input content, the target relationships, and the corresponding knowledge constraints. For the same power operation event unit, different training tasks correspond to different combinations of event fact anchors and relationships. The target output content of the training sample is jointly determined by the fact values ​​recorded by the selected event fact anchors, the relationships between event fact anchors, and the corresponding knowledge constraints, rather than being generated independently of the event fact anchors.

[0060] In one implementation, for a state judgment training task, a trigger event fact anchor, a pre-event state fact anchor, and a post-event state fact anchor can be selected. Training sample inputs are generated based on the trigger event, and the target output is determined by the state transition relationship before and after the event.

[0061] For event causal analysis training tasks, you can select trigger event fact anchors, equipment state change fact anchors, and operating parameter response fact anchors, and form causal analysis training samples based on the correlation between each fact anchor.

[0062] For time-series reasoning training tasks, multiple event fact anchors, operation fact anchors, and time sequence fact anchors can be selected to form training samples of event execution order or state evolution order based on the corresponding time relationships.

[0063] For operational reasoning training tasks, the target operational fact anchor point can be determined based on the pre-event state, the triggering event, and the corresponding operational rules, and corresponding training samples can be formed.

[0064] The consistency of the device identifier, device status, event type, operation behavior, operating parameter value and time sequence relationship in the target output of the training sample is compared with the corresponding event fact anchor point.

[0065] After the consistency comparison is passed, the corresponding training sample is determined as the training sample of the target large language model.

[0066] The generation method of training samples is not limited to a specific generation model. It can be generated by rule templates, procedural generation, language model generation, or a combination of the above methods. When language model generation is used, the corresponding event fact anchors and knowledge constraints are used as generation constraints. The consistency of the device status, events, operation behaviors, operating parameters and time sequence relationships involved in the language model generation results with the corresponding event fact anchors is compared. After the consistency comparison is passed, the target output content is formed, so that the language model can be used for the language organization of training samples without changing the target facts determined by the event fact anchors.

[0067] Step 7: Construction of the training corpus Assign sample labels to the generated large language model training samples.

[0068] Establish a data hierarchy among sample identifiers, training task types, event fact anchor point identifiers, power operation event unit identifiers, and original operation data identifiers.

[0069] The data hierarchy relationship can be represented as follows: Sample_ID -> Fact_ID -> Event_ID -> Raw_Data_ID Among them, Sample_ID represents the training sample identifier, Fact_ID represents the event fact anchor identifier, Event_ID represents the power operation event unit identifier, and Raw_Data_ID represents the source raw operation data identifier.

[0070] Based on the aforementioned data hierarchy, when the target output of the training sample needs to be verified, the corresponding event fact anchor point, power operation event unit, and source original operation data can be located by reverse positioning of the training sample.

[0071] The above methods are used to create a traceable large language model training corpus.

[0072] III. System Technical Solution The second aspect of the present invention provides a large language model training corpus generation system based on power operation event reconstruction and knowledge enhancement, including a data acquisition module, a data association module, an event recognition module, an event reconstruction module, a fact anchoring module, a sample generation module, and a genealogy management module.

[0073] The data acquisition module is used to acquire raw operating data from multiple power operation data sources.

[0074] The data association module is used to unify the time base and map the equipment identifiers of the raw operating data from different power operation data sources, and to determine the electrical object association and time association relationships between different raw operating data.

[0075] The event recognition module is used to identify power operation event boundaries based on the electrical object association and time association, combined with equipment status changes, operating parameter changes, and time intervals between adjacent original operating data, and to form an event association data set.

[0076] The event reconstruction module is used to determine the pre-event state, triggering event, operation behavior, post-event state, operating parameter response and timing relationship based on the event-related data set, and reconstructs it into a power operation event unit.

[0077] The fact anchoring module is used to perform consistency verification on power operation event units based on electrical topology relationships, equipment state transition rules, operation timing rules, and operating parameter constraints, and to determine the event fact anchor point based on the event elements that have passed the consistency verification and their correlation relationships.

[0078] The sample generation module is used to determine the target event element type and target association type according to the training task type of the large language model, select the corresponding event fact anchor point, and generate large language model training samples based on the selected event fact anchor point and its corresponding knowledge constraints.

[0079] The genealogy management module is used to establish the association between large language model training samples and training task types, power operation event units, event fact anchors, and raw operation data.

[0080] The above modules can be deployed on servers, cloud computing platforms, power data middleware, artificial intelligence training platforms, or other computing devices with data processing capabilities. Each module can be implemented in software, hardware, or a combination of software and hardware.

[0081] IV. Beneficial Effects Compared with the prior art, the present invention has at least the following beneficial effects.

[0082] First, it can improve the event-level correlation capability of multi-source asynchronous power operation data. This invention does not directly splice different data sources according to temporal proximity, but rather combines temporal relationships, electrical object associations, electrical topology relationships, equipment status changes, and operating parameter changes to identify event boundaries, which helps reduce the situation where data from different operating events are incorrectly aggregated.

[0083] Second, it can form event units that characterize the evolution of real power operation states. By reconstructing the pre-event state, triggering event, operational behavior, post-event state, operating parameter response, and timing relationship, discrete records originally scattered across multiple systems are transformed into power operation event units with state evolution relationships, providing a structured operational factual basis for training corpus generation.

[0084] Third, it can reduce the risk of discrepancies between the target output of the training samples and actual power operation facts. By verifying the consistency of event units through electrical topology relationships, equipment state transition rules, operation timing rules, and operating parameter constraints, and determining the event fact anchor points from the verified event elements and their relationships, the target output of the training samples is constrained by real operating facts and power industry rules.

[0085] Fourth, it can reuse the same real-world operational events for different large-scale language model training tasks. By establishing a mapping relationship between training task types and event element types and correlation types, training samples for different training tasks such as state judgment, event causal analysis, temporal reasoning, operational parameter interpretation, and operational reasoning can be generated from the same power operation event unit, thereby improving the ability to transform and utilize training corpora from real-world operational data.

[0086] Fifth, it can improve the verifiability and traceability of training corpora. By establishing a data hierarchy relationship between training samples, event fact anchors, power operation event units, and original operation data, the corresponding real operation records can be located from the training samples, providing a data foundation for the verification, maintenance, and subsequent use of training corpora. Attached Figure Description

[0088] Figure 1 This is a flowchart illustrating the overall process of a method for generating training corpus for a large language model based on power operation event reconstruction and knowledge enhancement, as proposed in this invention.

[0089] Figure 2 This is a schematic diagram illustrating the unification of time base for multi-source power operation data and the association of electrical objects in this invention.

[0090] Figure 3 This is a flowchart of the power operation event boundary identification process of the present invention.

[0091] Figure 4 This is a schematic diagram of the power operation event unit reconstruction of the present invention.

[0092] Figure 5 This is a flowchart for the knowledge constraint verification and event fact anchor point construction of this invention.

[0093] Figure 6 This is a flowchart illustrating the training sample generation process based on training task mapping in this invention.

[0094] Figure 7 This is a schematic diagram of the genealogical relationships of the training corpus data of this invention.

[0095] Figure 8 This is a schematic diagram of the structure of the large language model training corpus generation system of the present invention. Detailed Implementation

[0097] The present invention will be further described below with reference to specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0098] Example 1: Event Reconstruction of Multi-Source Operational Data The implementation of this invention is illustrated by taking the operation process of a power transmission line where a protection action occurs and causes a change in the state of the circuit breaker as an example.

[0099] The system acquires operational data from relay protection devices, substation monitoring systems, dispatch automation systems, and dispatch operation logs. Specifically, relay protection devices record distance protection action information, substation monitoring systems record circuit breaker remote signaling changes, dispatch automation systems record line current and active power changes, and dispatch operation logs record subsequent operational procedures.

[0100] The above operational data is standardized in terms of time reference, and the line names, protection device identifiers and circuit breaker identifiers of different systems are mapped to a unified electrical object and its topology associated devices through device identifier mapping.

[0101] When a distance protection action is detected or the circuit breaker status changes from closed to open, the corresponding record will be identified as a candidate event trigger record.

[0102] Based on candidate event trigger records, search for related data forward and backward, and determine related operational data according to time intervals, electrical topology relationships, equipment status changes, and line current changes.

[0103] In one specific implementation, events can be ranked according to their relevance: Ri = β1Rt,i + β2Ro,i + β3Rs,i + β4Rp,i Calculate the degree of association between the running data to be associated and the candidate event triggering record, and classify the running data whose event association degree reaches the preset association threshold into the same candidate event.

[0104] Protection action records, circuit breaker status change records, current change records, and dispatch operation records belonging to the same operation process are aggregated to form an event-related data set.

[0105] Reconstruct the state before the event occurs, the triggering event, the state transition, the running parameter response, the operation behavior, and the state after the event occurs based on the event-related data set.

[0106] For example: Pre-event status: The line is operating normally and the circuit breaker is in the closed position; Triggering event: Distance protection action; State transition: The circuit breaker changes from the closed position to the open position; Operating parameter response: Line current decreased significantly; Operational action: Execute the corresponding operational procedures; Post-incident status: The line enters the corresponding outage state or the operational status after handling.

[0107] The above-mentioned event elements, their temporal relationships, state transition relationships, and electrical object association relationships are combined to form a power operation event unit.

[0108] Example 2: Construction of Event Fact Anchors The knowledge constraint verification is performed on the power operation event unit formed in Example 1.

[0109] Verify the electrical relationships between protective operating devices, circuit breakers, and corresponding lines based on the electrical topology.

[0110] Verify the state changes of the circuit breaker before and after the corresponding event and operation according to the equipment state transition rules.

[0111] Verify the temporal sequence of protection actions, circuit breaker changes, operating parameter responses, and subsequent operations according to the operation sequence rules.

[0112] The system checks whether changes in parameters such as line current and active power, and changes in the corresponding equipment status, meet the preset operating conditions based on the constraints of the operating parameters.

[0113] For the event elements and relationships that have passed the above verification, establish state fact anchors, event fact anchors, parameter fact anchors, operation fact anchors, and time sequence fact anchors respectively, and record the corresponding source data identifiers, event unit identifiers, element types, element values, and knowledge constraint identifiers.

[0114] Event elements or relationships that cannot be verified by corresponding knowledge constraints will not be used as event fact anchors for the target output content of training samples.

[0115] Example 3: Generation of Multi-Task Training Samples Based on the event fact anchor points obtained in Example 2, multiple training task samples can be generated for the same power operation event.

[0116] For the state judgment task, the input content is constructed based on the protection action fact anchor point and the circuit breaker state fact anchor point, and the state change of the circuit breaker from closed to open is taken as the target output fact.

[0117] For the task of event causal analysis, the association is established based on the event fact anchor points corresponding to protection actions, circuit breaker status changes and line current changes, and the target output is formed based on the association relationship verified by knowledge constraints.

[0118] For time-series reasoning tasks, training samples are generated based on the time fact anchors corresponding to protection actions, circuit breaker changes, parameter changes, and operational operations, so that the event sequence in the target output is consistent with the time relationship corresponding to the real operational data.

[0119] For the task of interpreting operating parameters, the input content is generated based on the fact anchors of triggering events, the fact anchors of state changes, and the fact anchors of operating parameters, and the target output is determined based on the corresponding changes in operating parameters and their correlation with the device state.

[0120] For operational reasoning tasks, operational fact anchors are selected based on the pre-event state, the triggering event, and the corresponding operational rules to form operational reasoning training samples.

[0121] Each training sample retains the association identifier between itself and the corresponding event fact anchor point, power operation event unit and original operation data, so that it is possible to trace back its corresponding real operation fact based on the training sample.

[0122] Example 4: Deployment of the Training Corpus Generation System In one embodiment, the system of the present invention is deployed on a power artificial intelligence training platform or a power data middleware platform.

[0123] The data acquisition module acquires data generated by different power business systems and operation monitoring devices through data interfaces; the data association module completes time base unification, equipment identifier mapping, and topology association; the event identification module completes candidate event trigger record identification and event boundary determination; the event reconstruction module forms power operation event units; the fact anchoring module performs consistency verification according to power professional rules and forms event fact anchors; the sample generation module generates corresponding training samples according to the training task type; and the genealogy management module saves the association relationship between training samples and original operation data.

[0124] The generated training corpus can be used for pre-training, supervised fine-tuning, instruction fine-tuning, or other model training processes of the power industry's large language model, or it can be stored in the training corpus for subsequent model training.

[0125] This invention is not limited to transmission line protection operation scenarios, but can also be applied to transformer operation abnormalities, bus faults, power distribution equipment faults, new energy output abnormalities, changes in energy storage operation status, and other business scenarios where the state evolution process can be characterized by multi-source power operation data.

[0126] Those skilled in the art can adjust the event triggering conditions, event association conditions, knowledge constraint set, training task type and training sample template according to specific business scenarios, without changing the basic technical concept of the present invention to construct training corpus through multi-source running data association, running event reconstruction, event fact anchoring and fact-constrained training sample generation.

[0127] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, various modifications, equivalent substitutions, or improvements can be made to the present invention without departing from its spirit and principles. All modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating training corpus for a large language model based on power operation event reconstruction and knowledge enhancement, characterized in that, Includes the following steps: S1, acquire raw operating data from multiple power operation data sources, the raw operating data including at least equipment operating status data, operating parameter data, event records and operation records; S2 unifies the time base and maps the equipment identifiers for the raw operating data from different power operation data sources. Based on the equipment identifiers, electrical topology relationships, and data time relationships, it determines the electrical object association relationships and time association relationships between different raw operating data. S3, based on the electrical object association relationship and time association relationship, the original operating data of the equipment state undergoing preset changes, the operating parameters changing to meet preset change conditions, the protection action occurring and / or the alarm being triggered are determined as candidate event trigger records; Based on the candidate event trigger record, related operation data is searched forward and backward according to the electrical object associated with the candidate event trigger record, the time interval between adjacent original operation data, and the electrical topology association range. When the time interval between adjacent original operation data exceeds a preset time threshold, the corresponding electrical object exceeds the electrical topology association range, or the equipment status meets a preset event termination condition, the event boundary is determined. The related operation data within the event boundary is aggregated to form an event related data set. S4. Based on the changes in equipment status, event records, operation records, changes in operating parameters and their time sequence in the event-related data set, determine the equipment status before the event, the triggering event, the operation behavior, the equipment status after the event, the response and time sequence of operating parameters, and associate the above elements according to the state transition relationship and the time sequence relationship based on the electrical object association relationship, and reconstruct to form a power operation event unit that represents the corresponding power operation process. S5. Obtain the electrical topology relationship, equipment state transition rules, operation timing rules, and operating parameter constraints corresponding to the power operation event unit. Perform consistency verification on the equipment state transition, timing relationship between events and operations, changes in operating parameters, and topology relationship between associated electrical objects in the power operation event unit. For event elements and their associated relationships that meet the corresponding constraints, record their source data identifier, source validity identifier, event unit identifier, element type, and corresponding constraint identifier. Determine the event elements or the associated relationships between event elements that have valid source identifiers and pass the corresponding consistency verification as event fact anchors. S6, determine the training task type of the large language model, determine the corresponding target event element type and target correlation type according to the training task type, and select event fact anchor points that match the target event element type and target correlation type from the power operation event unit; The input content of the training sample is generated based on the selected event fact anchor, and the target output content of the training sample is determined based on the event elements, correlations and corresponding knowledge constraints represented by the selected event fact anchor. The device status, events, operations, running parameters and time sequence relationships in the target output content are kept consistent with the corresponding event fact anchor, forming a large language model training sample corresponding to the training task type. S7. Establish the association between the training samples of the large language model and their corresponding training task types, power operation event units, event fact anchors and original operation data to form a traceable large language model training corpus.

2. The method according to claim 1, characterized in that, The multiple power operation data sources include at least two of the following: dispatch automation system, energy management system, distribution management system, substation monitoring system, relay protection device, synchronous phasor measurement device, and power business management system; the original operation data includes at least two of the following: telemetry data, remote signaling data, protection action record, alarm record, fault waveform characteristic data, dispatch operation log, and operation record.

3. The method according to claim 1, characterized in that, Step S3 includes: Candidate event trigger records are determined based on changes in equipment operating status and / or operating parameters. Using the time point and electrical object corresponding to the candidate event trigger record as the event search benchmark, the associated electrical object is determined along the electrical topology connection relationship of the electrical object within a preset time range, and the associated operation data corresponding to the associated electrical object is obtained; The degree of event correlation is determined based on the time interval between the associated operational data and the candidate event triggering record, the electrical topology distance, the degree of correlation between equipment status changes and the degree of correlation between operational parameter changes; Related operational data whose event correlation meets the preset correlation conditions are grouped into the same candidate power operation event, and the event boundary is determined when the time interval between adjacent related operational data exceeds a preset time threshold, the related electrical object exceeds the preset electrical topology range, or the equipment status meets at least one of the preset event termination conditions.

4. The method according to claim 1, characterized in that, The power operation event unit includes an electrical object identifier, a pre-event state, a triggering event, an operational action, a post-event state, an operational parameter response, and a timing relationship. It also establishes the association between the triggering event and the equipment state transition, the triggering event and the operational parameter response, and the operational action and the post-event state. The pre-event state and the post-event state are determined based on the equipment operation state of the corresponding electrical object before and after the event boundary, respectively. The operational parameter response is determined based on the changes in operational parameters within a preset time range before and after the event.

5. The method according to claim 1, characterized in that, For event elements or relationships between event elements that originate from at least one original operational data source and have passed the consistency verification of corresponding knowledge constraints, a fact record is established, including a source data identifier, a source validity identifier, an event unit identifier, an element type, an element value, and a constraint identifier. Based on the source validity identifier, it is determined whether the corresponding original operational data is valid. When the element value in the fact record is consistent with the corresponding original operational data, the source of the corresponding original operational data is valid, and the corresponding knowledge constraints are satisfied, the fact record is determined as the event fact anchor point, and a reverse index relationship is established between the event fact anchor point and the corresponding original operational data.

6. The method according to claim 1, characterized in that, A task mapping relationship is pre-established between training task types and target event element types and target association relationship types; the task mapping relationship is queried according to the training task type of the training sample to be generated to determine the corresponding target event element type and target association relationship type, and a matching event fact anchor point is selected from the power operation event unit; wherein, the training task type includes at least one of state judgment, event causal analysis, time sequence reasoning, operation parameter interpretation and operation reasoning.

7. The method according to claim 6, characterized in that, Based on the sample template corresponding to the training task type, some event elements and their relationships in the selected event fact anchors are organized into training sample input content, and the target event elements, target relationships and corresponding knowledge constraints used to answer the input content are organized into training sample target output content. The device identifier, device status, event type, operation behavior, running parameter value and time relationship in the target output content are compared with the corresponding event fact anchor point. After the consistency comparison is passed, the target large language model training sample is formed.

8. The method according to claim 1, characterized in that, Assign sample identifiers to the training samples of the large language model, and establish a data hierarchy relationship between sample identifiers, training task types, event fact anchor point identifiers, power operation event unit identifiers, and original operation data identifiers; when the target output content in the training samples needs to be verified, the corresponding event fact anchor point and original operation data are located in reverse according to the data hierarchy relationship.

9. The method according to claim 1, characterized in that, The consistency verification includes: determining whether the associated electrical objects in the event unit meet the preset electrical connection relationship based on the electrical topology relationship; determining whether the state before the event, the operation behavior, and the state after the event meet the preset state transition conditions based on the equipment state transition rules; determining whether the time sequence between the triggering event, the operation behavior, and the state change meets the preset timing conditions based on the operation timing rules; and determining whether the changes in the operating parameters before and after the event are within the parameter range allowed by the corresponding equipment and operating state based on the operating parameter constraints. For event elements or associations that do not meet the corresponding constraints, they are not determined as event fact anchors for generating the target output content.

10. A system for generating training corpus for a large language model based on power operation event reconstruction and knowledge enhancement, characterized in that, include: The data acquisition module is used to acquire raw operating data from multiple power operation data sources; The data association module is used to unify the time base and map the equipment identifiers of the raw operating data from different power operation data sources, and to determine the electrical object association and time association relationships between the raw operating data. The event recognition module is used to identify power operation event boundaries based on the electrical object association and time association, combined with equipment status changes, operating parameter changes, and time intervals between adjacent original operating data, and to form an event association data set; The event reconstruction module is used to determine the pre-event state, triggering event, operation behavior, post-event state, operating parameter response and timing relationship based on the event-related data set, and reconstruct the power operation event unit. The fact anchoring module is used to perform consistency verification on the power operation event unit based on electrical topology relationships, equipment state transition rules, operation timing rules and operating parameter constraints, and to determine the event fact anchor point based on the event elements that pass the consistency verification and their correlation relationships. The sample generation module is used to select corresponding event fact anchors according to the training task type of the large language model, and generate large language model training samples corresponding to the training task type based on the selected event fact anchors and their corresponding knowledge constraints. The genealogy management module is used to establish the association between the training samples of the large language model and the training task type, power operation event unit, event fact anchor point and original operation data.

Images