Multi-agent construction method for high-voltage converter station fault handling

Abstract

The invention discloses a multi-agent construction method for fault handling of a high-voltage converter station, and relates to the technical field of electric power automation, and the method comprises the following steps: S1, mapping establishment of in-station equipment and agents; s2, constructing a unified identifier and event model; s3, organization and credibility calibration of evidence information; s4, performing consistency arbitration and conflict grading processing; s5, arranging and executing a fault disposal process; s6, running experience learning and rule self-evolution; and S7, auditing and tracing the disposal process. According to the method, a multi-agent construction mechanism is introduced into the converter station, devices such as a direct current pole, a valve tower, protection, measurement and control, operation and the like are respectively mapped into independent agents, and cross-domain cooperation is realized under a unified identification and event model; each agent not only has an autonomous information processing capability, but also can realize rapid fusion and consistency judgment of multi-source information through an evidence chain and an arbitration mechanism, thereby forming explainable and traceable disposal suggestions when a fault occurs.

Description

Technical Field

[0001] This invention relates to the field of power automation technology, specifically to a multi-agent construction method for fault handling in high-voltage converter stations. Background Technology

[0002] High-voltage direct current (HVDC) converter stations serve as crucial hubs for inter-regional power grid transmission and large-scale renewable energy consumption. Their operational safety and fault handling efficiency directly impact the stability of the entire power system. Currently, common converter station fault handling methods largely rely on a centralized control architecture, where the station control layer uniformly receives signals from various protection devices, monitoring equipment, and operation ticket systems before making centralized decisions. While structurally simple, this approach has significant shortcomings: First, at the moment a fault occurs, signals from DC poles, valve towers, converter transformers, AC filters, and other components flood in concurrently. Centralized processing easily leads to information congestion and computational delays, making it difficult to respond promptly to rapidly evolving conditions. Second, differences in message formats and naming conventions among different manufacturers' equipment, coupled with a lack of unified identifiers and event models, make it difficult to directly align and integrate multi-source data, often requiring manual intervention and increasing the risk of erroneous or missed actions. These shortcomings make it difficult for existing technologies to simultaneously achieve real-time performance, reliability, and interpretability when facing complex, multi-source, and high-frequency fault scenarios. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a multi-agent construction method for fault handling in high-voltage converter stations, thereby solving the problems mentioned in the background section.

[0004] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a multi-agent construction method for fault handling in high-voltage converter stations, comprising the following steps: S1. Establishment of mapping between on-site devices and agents; The main equipment and functional components in the converter station are clearly defined as real objects, and corresponding intelligent agents are established based on the granularity of the real objects in the converter station. S2, Construction of a unified identifier and event model; After completing the mapping from device to agent, establish a unified identification system and event model; S3. Organization and credibility assessment of evidentiary information; After completing the unified identifier and event model, the original records reported by each agent are extracted as evidence units, evidence chains are formed according to the constraint rules, and credibility files are established for each agent participating in the reporting, generating candidate failure hypotheses. S4. Consistency arbitration and conflict classification handling; Using the evidence chain and candidate failure hypotheses formed in step S3 as input, a four-dimensional scoring framework consisting of time alignment, physical consistency, multi-source mutual verification, and proxy credibility is established and summarized into a total score according to preset weights. Based on this, each candidate failure hypothesis is ranked and conflicts are classified. S5. Arrangement and execution of fault handling procedures; Using the fault assumptions sorted and graded in step S4 as the triggering condition, the process is executed according to the state flow of pre-check—action orchestration—secondary check—rollback / upgrade. S6. Learning from operational experience and self-evolution of rules; In the fault handling process, the credibility of the agent is continuously improved, and high-frequency evidence patterns are explored to form new rules; S7. Auditing and tracing of the disposal process; It records the entire process in its entirety, generates auditable and replayable data, and supports scenario replay and minimal disclosure export.

[0005] To further optimize this technical solution, in step S1, Real object partitioning: Establish a topology tree at the device level, expanding layer by layer from station level to pole level, valve tower level, and interval level, with each node at each level bound to one or more intelligent agent agents; The established intelligent agent agent: clearly defines its input signal source, output command type, scope of participation in fault handling, and information interaction mode.

[0006] To further optimize this technical solution, in step S2, a hierarchical unique identifier encoding rule is constructed for the unified identifier system: It is structured in six levels: station area, pole area, bay, equipment, terminal, and signal. Each level is encoded using a combination of numbers and characters, so that the status of any signal or equipment can be uniquely identified through a complete identification link. It is accurately represented in the form of "station area code + pole area number + bay identifier + equipment category + terminal number + signal sequence number".

[0007] To further optimize this technical solution, in step S2, the event model is used to abstractly describe different types of faults or anomalies; The event model includes the event's category label, triggering conditions, duration, scope of impact, severity level, and priority; The event model also introduces logical relationships between events, including causal relationships, corroborative relationships, and contradictory relationships.

[0008] To further optimize this technical solution, in step S3, the evidence unit includes event category, unique identifier of device and signal, occurrence time window, numerical summary, source fingerprint, and acquisition conditions; The constraint rules are a triple constraint based on time sequence, topological reachability within the station, and physical conservation.

[0009] To further optimize this technical solution, the constraint rules automatically generate a chain of evidence based on the triple constraints of chronological order, intra-station topological reachability, and physical conservation, establishing causal, corroborative, and contradictory relationships between pieces of evidence. This chain includes: First, filter by adjacent time windows and upper limit of propagation delay, then determine the influencing chain based on the device connection path, and finally eliminate pseudo-correlation by physical verification of power balance, DC loop connectivity, and valve arm thermo-electric consistency. For duplicate reports with the same identifier and within the same time window, duplicates are deduplicated using the source fingerprint and merged according to the completeness of the record; For waveform recording and event fragmentation, a process of fragment splicing and consistency verification is adopted. Adjacent fragments are spliced ​​according to the buffer before and after the trigger, and the waveform characteristics are verified to be consistent with the event description, thereby forming a loop-free chain of evidence.

[0010] To further optimize this technical solution, in step S4, the four-dimensional scoring framework includes: Time alignment is checked against the order of evidence occurrence and the upper limit of internal transmission delay. High scores are awarded when the time deviation is within the allowable window. Physical consistency is tested by examining power balance, inter-electrode current difference, valve arm conduction relationship and topological reachability. If the power imbalance is below the upper limit and the topology is reachable, it is considered consistent. Multiple sources of evidence are required to be from independent sources. The more sources and the more consistent the conclusions, the higher the score. Agent credibility is based on historical accuracy, clock synchronization quality, and data integrity, with agents that have been stable over a long period of time given higher weight.

[0011] To further optimize this technical solution, step S5, the state flow of pre-verification—action orchestration—secondary verification—rollback / upgrade, includes: Pre-verification includes topology reachability and isolation point location verification, protection and interlocking condition matching, power flow and reactive power margin verification, local / remote control authority and operation ticket authorization confirmation, and time synchronization and signal consistency checks; if any of these conditions are not met, the system will switch to safe waiting or derated operation. Action orchestration breaks down the process into verifiable atomic actions. Each action clearly defines the target device, expected measurement, and success criteria, and sets tiered timeouts and retry counts. Actions avoid mutual interference through conditional jumps and interlocking links. Secondary verification follows a closed loop of suggestion-confirmation-execution-receipt: the system can only issue the order after double confirmation on the running interface, and the receipt source is consistent across both channels; If new contradictory evidence or key criteria are not met during the rollback / upgrade process, subsequent actions will be immediately stopped and the process will enter the rollback branch, where reversible operations will be undone in the preset sequence or safety interlocks will be executed.

[0012] To further optimize this technical solution, in step S6, high-frequency evidence patterns are mined to form new rules, including: The system performs pattern mining on the high-frequency evidence subgraphs and stable chronological order in historical samples to obtain candidate causal constraints or disposal templates. Each candidate rule must undergo two levels of verification. First, the explanatory power and safety boundaries of past events are verified in simulation playback. Then, it is enabled in a small-scale grayscale to observe the trigger rate, accuracy and impact on operating indicators. Only after the threshold is met can it be enabled on the whole site. All changes to parameters and rules are recorded in a versioned manner, including the effective time, triggering background, comparison indicators, and rollback path.

[0013] To further optimize this technical solution, in step S7, the auditable and reproducible data includes the following structured content: Meta-information list, evidence catalog, evidence relationship diagram, arbitration details, conflict classification and resolution path, resolution process sequence, and learning and rule change records.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, they implement the steps of a multi-agent construction method for fault handling of a high-voltage converter station as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of a multi-agent construction method for fault handling of a high-voltage converter station as described in the first aspect of the present invention.

[0016] Compared with the prior art, the present invention provides a multi-agent construction method for fault handling in high-voltage converter stations, which has the following beneficial effects: This multi-agent construction method for fault handling in high-voltage converter stations introduces a multi-agent construction mechanism within the converter station, mapping DC poles, valve towers, protection, measurement and control, and operation equipment as independent agents, and achieving cross-domain collaboration under a unified identifier and event model. Each agent not only possesses autonomous information processing capabilities but also achieves rapid fusion and consistent judgment of multi-source information through evidence chains and arbitration mechanisms, thereby generating explainable and traceable handling suggestions when a fault occurs. This method can significantly reduce information bottlenecks in centralized architectures, improve the response speed and accuracy of fault handling, and possess self-learning and rule evolution capabilities during continuous operation, making the system increasingly stable and enhancing the overall safety and intelligence level of the converter station. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a multi-agent construction method for fault handling in high-voltage converter stations proposed in this invention. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0022] Example 1: Reference Figure 1 This is the first embodiment of the present invention, which provides a multi-agent construction method for fault handling in high-voltage converter stations, including the following steps: S1. Establishment of mapping between on-site devices and agents; This method clearly defines the main equipment and functional components of a converter station as real-world objects and establishes corresponding intelligent agent agents defined at the granularity of these real-world objects. A converter station typically includes multiple components such as DC poles, valve towers, converter transformers, AC filters, DC field equipment, protection devices, measurement and control systems, and interfaces for interaction with the dispatch center. Each of these components plays a different role in fault handling. Relying solely on traditional centralized control methods can easily lead to information congestion, decision-making delays, and cross-domain conflicts. This method maps different equipment and their functional units as independent agents, enabling each agent to have clearly defined boundaries and responsibilities while collaborating within a unified framework.

[0023] The real-world object segmentation involves establishing a topology tree at the equipment level, expanding layer by layer from the station level to the pole level, valve tower level, and bay level. Each node at each level is bound to one or more intelligent agent agents. For example, the DC pole corresponds to the pole control agent, the valve tower corresponds to the valve control agent, the converter transformer corresponds to the transformer health agent, the AC filter and reactive power compensation device corresponds to the reactive power regulation agent, the protection and measurement and control devices correspond to the protection agent and the measurement and control agent, and the operation tickets and human-machine interface are managed by the operation agent.

[0024] The established intelligent agent agent is characterized by clearly defining its input signal sources, output command types, scope of participation in fault handling, and information interaction modes. For example, the valve control agent is mainly responsible for collecting and uploading information such as valve tower temperature, trigger angle, and conduction status, and can receive derating or blocking commands from the station control agent; the protection agent is responsible for determining the relay protection action and transmitting the action signal to the arbitration stage, and also has the function of publishing the action results to the outside world.

[0025] In this embodiment, the following agent types and responsibilities can be adopted, as shown in Table 1. Table 1 Table 1 defines the agent family tree and responsibility boundaries at the granularity of the actual objects in the converter station: StationCtrl-Agent (disposal orchestration), Arbiter-Agent (arbitration), PoleCtrl / ValveCtrl-Agent (pole / valve control), DCYard-Agent (DC field), ACFilter&Q-Agent (AC reactive power / filtering), ConverterTR-Health-Agent (converter transformer health), Protection-Agent (protection IED), M&C-Agent (measurement and control), Record&SOE-Agent (waveform recording / SOE), OpTicket-Agent (operation ticket / HMI), and Dispatch-IF-Agent (interaction with EMS / dispatch).

[0026] During the design of the agents, their time synchronization methods, message transmission channels, and access control also need to be specified. For example, each agent relies on a unified clock signal across the entire station (such as a high-precision time synchronization system) to ensure the comparability of multi-source information in the time dimension; information transmission adopts a distributed message bus structure, with different agents only subscribing to and publishing messages within their authorized scope, avoiding interference from irrelevant data. Through this structured mapping, the entire converter station forms a network composed of multiple independent yet closely cooperating agent units. Each agent has clear responsibilities and well-defined boundaries, effectively supporting subsequent event identification, evidence organization, and arbitration decisions, thus providing a fundamental guarantee for fault handling. Compared with the centralized control mode in existing technologies, this mapping method has innovative advantages such as strong scalability, high fault tolerance, and clear division of responsibilities.

[0027] S2, Construction of a unified identifier and event model; After completing the mapping from device to agent, a unified identification system and event model are established.

[0028] Establish a hierarchical and unique identifier coding rule for the unified identifier system: The system is structured in six layers: station area, pole area, bay, equipment, terminal, and signal. Each layer is encoded using a combination of numbers and characters, ensuring that the status of any signal or device can be uniquely identified through a complete identification chain. This is precisely represented in the format of "station area code + pole area number + bay identifier + equipment category + terminal number + signal sequence number." This identification method avoids inconsistencies or duplications in names when connecting equipment from different manufacturers and lays the foundation for subsequent event correlation and evidence chain construction.

[0029] Event models are used to abstractly describe different types of faults or anomalies; The event model includes event category labels (such as "protection trip", "valve tower overheating", "DC grounding", "AC short circuit", "monitoring and control anomaly", etc.), triggering conditions, duration, impact range, severity level, and priority. For example, the triggering condition for the "valve tower overheating" event could be that the valve body temperature exceeds a set threshold and persists for a certain period of time; its impact range includes the corresponding valve module and its corresponding polar region; and the priority is automatically adjusted according to the rate of temperature increase. Through this model-based design, complex real-time signals are transformed into standardized event descriptions, enabling different agents to use the same logical framework when parsing events, thereby achieving cross-system interoperability.

[0030] The event model also introduces logical relationships between events, including causal, corroborative, and contradictory relationships. For example, protection tripping events are often accompanied by abnormal power imbalance measurements, and these two are connected in the event model through a causal relationship; waveform anomalies in the waveform recording file can serve as corroborating evidence of protection actions; and some measurement and control anomalies may contradict the judgment of valve tower overheating, which needs to be resolved in subsequent arbitration steps. By pre-defining these logical relationships, the event model not only has the functions of classification and description, but also provides structured support for the construction of evidence chains and consistent arbitration. Overall, the establishment of unified identification and event models enables the converter station fault handling process to be carried out within a unified semantic and standardized framework, improving the accuracy and innovation of the system.

[0031] S3. Organization and credibility assessment of evidentiary information; After completing the unified identification and event model, the original records reported by each agent are extracted as evidence units, evidence chains are formed according to the constraint rules, and credibility files are established for each agent participating in the reporting, generating candidate failure hypotheses.

[0032] The evidence unit includes the event category, the unique identifier of the device and signal, the occurrence time window (start and end time), the numerical summary (such as the over-limit amplitude, duration, power imbalance ratio), the source fingerprint (file check code or record number), and the acquisition conditions (sampling interval, trigger threshold). The constraint rules are a triple constraint based on time sequence, topological reachability within the station, and physical conservation.

[0033] The constraint rules, based on a triple constraint of chronological order, intra-station topological reachability, and physical conservation, automatically generate a chain of evidence by establishing causal, corroborative, and contradictory relationships among the evidence, including: First, filter by adjacent time windows and upper limit of propagation delay, then determine the influencing chain based on the device connection path, and finally eliminate pseudo-correlation by physical verification of power balance, DC loop connectivity, and valve arm thermo-electric consistency. For duplicate reports with the same identifier and within the same time window, duplicates are deduplicated using the source fingerprint and merged according to the completeness of the record; For waveform recording and event fragmentation, a process of fragment splicing and consistency verification is adopted. Adjacent fragments are spliced ​​according to the buffer before and after the trigger, and the waveform characteristics are verified to be consistent with the event description, thereby forming a loop-free chain of evidence.

[0034] Subsequently, credibility profiles were established for each participating agent, and a grading system (high, medium, and low) was calculated based on four indicators: historical consistency rate, time synchronization quality, data integrity, and source independence. Source independence was used to suppress homogeneous duplication amplification; only mutual verification from heterogeneous channels improved overall confidence. Evidence with contradictory relationships but failing to meet time or physical constraints was marked as "pending arbitration," and conflicting edges were retained on the chain for subsequent processing. For evidence clusters with sufficient causal and corroborating relationships, candidate failure hypotheses and their supporting lists were generated as direct input for the consistency arbitration process. Through the joint chain building based on "time, topology, and physics" and the credibility labeling based on "independence weighting," traceable organization and full disclosure of multi-source evidence were achieved, providing stable and verifiable basic data for subsequent ranking, judgment, and handling.

[0035] S4. Consistency arbitration and conflict classification handling; Using the evidence chain and candidate failure hypotheses formed in step S3 as input, a four-dimensional scoring framework consisting of time alignment, physical consistency, multi-source mutual verification, and proxy credibility is established and summarized into a total score according to preset weights. Based on this, each candidate failure hypothesis is ranked and conflicts are classified.

[0036] The four-dimensional scoring framework includes: Time alignment is checked against the order of evidence occurrence and the upper limit of the internal transmission delay. A high score is obtained when the time deviation is within the allowable window (usually 5 to 20 milliseconds).

[0037] Physical consistency is determined by examining power balance, inter-electrode current difference, valve arm conduction relationship, and topological reachability. If the power imbalance is below the upper limit (usually 5% to 10%) and the topology is reachable, it is considered consistent.

[0038] Multiple sources of evidence are required to be from independent sources. The more sources and the more consistent the conclusions, the higher the score.

[0039] Agent credibility is based on historical accuracy, clock synchronization quality, and data integrity, with agents that have been stable over a long period of time given higher weight.

[0040] The system ranks the hypotheses accordingly and classifies conflicts: if there is a fundamental contradiction (such as a reversal of time sequence or violation of physical conservation), it is judged as a fatal conflict and must be transferred to manual review and trigger delayed locking; if it is an important conflict (such as missing key mutual evidence or highly homogeneous sources), it is downgraded and a prompt is issued; general conflicts only reduce the corresponding score and are retained in the audit record. When the difference between the scores of the first and second hypotheses is less than the threshold (usually 15% to 30%) or there is an unresolved fatal conflict, the system enters the manual confirmation branch; otherwise, it generates handling suggestions and execution priorities, and simultaneously outputs an interpretable evidence package, which includes the evidence number involved in the scoring, time window, key verification quantities and conflict sources, ensuring that the conclusions are verifiable and traceable, and providing consistent input for subsequent process arrangement and online learning.

[0041] S5. Arrangement and execution of fault handling procedures; Using the fault assumptions sorted and graded by the output of step S4 as the triggering condition, the process is executed according to the state flow of pre-verification—action orchestration—secondary verification—rollback / upgrade.

[0042] The state flow, including pre-verification, action choreography, secondary verification, and rollback / upgrade, includes: Pre-verification includes topology reachability and isolation point location verification, protection and interlocking condition matching, power flow and reactive power margin verification, local / remote control authority and operation ticket authorization confirmation, and time synchronization and signal consistency checks; if any of these conditions are not met, the system will switch to safe waiting or derated operation. Action orchestration breaks down the process into verifiable atomic actions. Each action clearly defines the target device, the expected measurement, and the success criteria (such as the power imbalance decreasing to within the set upper limit, the valve body temperature falling back within a limited time, and the circuit breaker position signal being consistent with the command). It also sets graded timeouts and retry counts. Actions are prevented from interfering with each other through conditional jumps and interlocking links. Secondary verification follows a closed loop of suggestion-confirmation-execution-receipt: the system can only issue the order after double confirmation on the running interface, and the receipt source is consistent across both channels; If new contradictory evidence or key criteria are not met during the rollback / upgrade process, subsequent actions will be immediately stopped and the process will enter the rollback branch, where reversible operations will be undone in the preset sequence or safety interlocks will be executed.

[0043] For typical scenarios, single-pole grounding prioritizes derating and polarity switching, while strengthening grounding current monitoring; valve arm overheating is first addressed with current limiting and delayed blocking, followed by evaluation of cooling recovery; AC side short circuits trigger reactive power and filter unit reconfiguration. The entire process generates an interpretable evidence package, recording evidence citations, action sequence, and criterion achievement status, providing data input for auditing and online learning.

[0044] S6. Learning from operational experience and self-evolution of rules; In the fault handling process, the credibility of the agent is continuously improved, and high-frequency evidence patterns are explored to form new rules.

[0045] Using the arbitration score, candidate hypothesis ranking, and conflict level generated in step S4, and the handling receipts and manual review opinions from step S5 as input, the system first performs unified time benchmark alignment and sample cleaning: records lacking key measurements or with abnormal time stamps are removed, and "evidence cluster - arbitration conclusion - handling result" are grouped into one-to-one corresponding sample pairs, labeled as "correct, misjudged, delayed effective, invalid". Then, structured features are extracted from the samples, including at least: the number and source differences of evidence, the magnitude of time and physical deviations, conflict type and level, the rate of compliance with handling actions, and the consistency of receipts, to establish a performance profile for operation (false alarm rate, missed alarm rate, average handling time, manual takeover rate, etc.). Based on this, the system performs small-step updates according to rolling time windows: for agents with long-term false alarms or time stamp jitter, their credibility is downgraded according to the error ratio and data missing rate; for agents with stable performance, their credibility is slowly increased according to robustness, ensuring smooth updates without overshoot. The adjustment of arbitration weights (the coefficients assigned to the four scoring factors) follows the principle of "residual-guided and safety-constrained": when a certain type of verification quantity consistently biases towards the same conclusion in a large number of samples, a small adjustment is made in that direction, with an upper limit on the step size and a boundary interval set; if key indicators deteriorate in the short term, the update is automatically frozen and rolled back to the previous version of parameters. To avoid bias caused by site differences, the update strategy introduces site characteristic factors and extreme operating condition exemption conditions to ensure that the convergence direction is not mistakenly triggered during special seasons or maintenance phases.

[0046] Regarding rule self-evolution, the system performs pattern mining on high-frequency evidence subgraphs and stable chronological sequences in historical samples to obtain candidate causal constraints or disposal templates. Each candidate rule undergoes two levels of verification: first, its explanatory power and safety boundaries for past events are verified in simulation playback; then, it is activated on a small scale to observe the trigger rate, accuracy, and impact on operational metrics. Only after meeting the threshold can it be fully activated across the entire system. All parameter and rule changes are recorded in a versioned manner, including the effective time, triggering background, comparison metrics, and fallback path, forming an auditable evolution archive. Through this mechanism, the system transforms the closed loop of "arbitration residual - disposal result - manual review" into a driving force for continuous learning, enabling the agent credibility and arbitration weight to adaptively converge with actual operation. The rule base is gradually enriched while ensuring safety, thereby improving the accuracy, stability, and interpretability of fault location and disposal.

[0047] S7. Auditing and tracing of the disposal process; It records the entire process in its entirety, generates auditable and replayable data, and supports scenario replay and minimal disclosure export.

[0048] The auditable and retrievable data shall include at least the following structured content: First, a metadata list, recording station area, level, interval, equipment and agent number, version number, unified clock verification result and generation time; Second, an evidence catalog, listing each piece of evidence number, source system, time window, summary value (including file checksum), integrity verification, deduplication and merging instructions; Third, an evidence relationship diagram, marking causal, corroborating, and contradictory edges and their generation basis; Fourth, arbitration details, providing time alignment, physical consistency, etc. The system includes the following data sets: 1) Multi-source mutual verification and proxy credibility (four sub-scores, weights, thresholds, and ranking results); 2) Conflict classification and handling paths, indicating the trigger points and resolution results for fatal / important / general conflicts; 3) Handling process sequence, listed item by item according to "pre-verification—atomic action—conditional jump—interlock verification—success criterion—abnormal interruption—rollback / upgrade," with consistency records of manual confirmation and dual-channel receipts; 4) Learning and rule change records, specifying parameters before and after adjustment, trigger background, gray-scale activation results, and rollback paths. All the above data is written to a non-modifiable medium and stored on both the system and the superior platform, with daily / weekly summary value reconciliation to ensure immutability and verifiability. To support post-event review and training, the system can automatically generate "replay scenarios" from the audit package. The generation logic is as follows: adjacent buffers are extracted according to the event time window, the topology and initial running point are restored, evidence sequences are injected in chronological order, and the sampling period and trigger thresholds are reused to ensure consistency between the replay behavior and the actual situation. To protect sensitive information, a "minimum explorable export" approach is provided. Sensitive fields such as personnel identity, geographical location, and device serial number are anonymized, retaining only essential evidence fingerprints, timestamps, and criterion achievement status. Access to this export requires authorization and operational traceability. The system also establishes a multi-dimensional index (by time, device, agent, event category, and handling process) and a health self-check mechanism. When a broken evidence chain, timestamp drift, or inconsistent summary value is detected, an alarm is automatically triggered and the relevant entries are frozen. Through the above design, a closed loop is achieved in the handling process: "full-link recording—interpretable verification—controllable disclosure—repeatable experimentation," providing a directly usable technical foundation for continuous optimization and accountability.

[0049] Example 2: In practical applications, this method can be deployed in converter stations of UHVDC transmission projects to enable real-time fault detection, assessment, and handling during operation. A typical scenario, a "unipolar DC grounding fault," will be used for illustration.

[0050] When a ground fault occurs at the DC electrode, the abnormal signal is detected almost simultaneously by the valve control agent, DC field agent, and protection agent. The valve control agent reports a sudden drop in electrode voltage and an imbalance in valve tower current, the DC field agent detects a sharp increase in ground current, and the protection agent triggers ground protection action. This information is transformed into standardized events under a unified identification system and automatically incorporated into the event model. Subsequently, multi-source data are used to form a chain of evidence: a causal relationship is established between the protection action and the current surge, with valve tower current imbalance serving as corroborating evidence. If some measurement and control data do not show a consistent trend, they are marked as contradictory evidence awaiting arbitration.

[0051] The arbitration phase involves a comprehensive evaluation based on chronological order, physical consistency, cross-verification from multiple sources, and the credibility of the proxy. For example, if a sudden increase in ground current coincides with the protection system's activation time, the power imbalance exceeds a set threshold, and signals from different sources corroborate each other, the total arbitration score is significantly higher than other hypotheses, thus determining "single-pole grounding" as the optimal fault conclusion. If the score difference between this conclusion and the second candidate exceeds a set threshold, no manual confirmation is required, and the process proceeds directly to the handling phase.

[0052] The handling procedure is programmed and the corresponding handling script is retrieved. At this point, the station control agent issues an instruction to the valve control agent: execute derated operation, close some valve arm trigger angles, switch polarity operating mode, and simultaneously activate enhanced monitoring of DC grounding current. The operation ticket agent generates a list of pending confirmations on the interface. The instruction is only officially issued after the operator approves it through a double confirmation mechanism. If new evidence emerges during execution (such as the grounding current not recovering significantly), a delayed interlocking procedure is automatically triggered to ensure safety is prioritized.

[0053] After the incident is resolved, all events, evidence chains, arbitration results, enforcement instructions, and manual confirmation information will be recorded in the audit package, forming a complete traceability file. A replayable scenario file will also be automatically generated for subsequent training or algorithm optimization. Based on the effectiveness of this resolution, the system will adjust the credibility and arbitration weight of the relevant agents and mark new patterns of this type of failure in the rule base to enhance subsequent identification and resolution capabilities.

[0054] Example 3: In converter station operation, valve tower overheating is a common operational risk, and the timeliness of its handling directly affects the safety of the DC system. Suppose that the temperature of a valve arm continuously rises above a set threshold. First, the valve control agent detects the abnormal real-time temperature data and converts it into a "valve tower overheating" event, uploading it to the event model. Simultaneously, the cooling system monitoring agent detects insufficient coolant flow, and the protection agent records the corresponding temperature protection alarm signal. After this evidence is collected by a unified identification system, a causal chain is generated by the evidence organization module: insufficient coolant flow is the precursor to the valve temperature rise, the real-time temperature data from the valve control agent is the core evidence, and the protection alarm is corroborating evidence. If the data from the monitoring and control agent is delayed or inconsistent at this time, it is marked as contradictory evidence.

[0055] The candidate fault hypothesis was scored. Time alignment showed that the temperature rise and cooling flow rate decrease matched within a second range; physical consistency verification confirmed that the temperature rise was related to the power transmission state and that there was no violation of energy conservation; multi-source mutual verification was supported by evidence from valve control, cooling, and protection; valve control proxy and protection proxy scored highly in terms of credibility. Therefore, this hypothesis scored significantly higher than other candidates such as "simple measurement and control anomaly", and the arbitration conclusion was determined to be "valve tower overheating".

[0056] In the handling process, the station control agent retrieves the corresponding script and first issues a "flow-limiting operation" command to the valve control agent, reducing the valve arm conduction angle to alleviate heat generation. Simultaneously, it notifies the cooling system agent to implement cooling and pressurization measures. If the temperature does not recover within the set time, a "delayed interlock" is triggered, temporarily detaching the faulty valve arm and reconfiguring the power distribution. The operation ticket agent generates an execution list and requires operator confirmation, monitoring temperature changes in real time during execution. If normal operation is restored, the system maintains derated operation with enhanced monitoring; if the temperature continues to rise, a forced interlock is executed. Finally, the entire handling chain is recorded and an audit package is generated for subsequent review and rule optimization.

[0057] Example 4: When a short circuit occurs in the AC system, the electrical impact on the converter station is severe, requiring rapid identification and stabilization measures. Assuming a single-phase short circuit occurs in the AC side filter bay, the AC filter agent detects a sharp rise in the filter branch current and triggers protection action. Simultaneously, the protection agent reports an "AC short circuit" event. The DC field agent records a power drop and electrode current fluctuation, and the converter transformer agent detects voltage imbalance. These events, integrated through a unified identification system, establish a causal relationship: the short circuit action causes a power drop, the power drop is accompanied by voltage imbalance, the filter branch current is the core evidence, and the protection trip provides supporting evidence.

[0058] In the arbitration scoring, time alignment showed that the multiple events occurred simultaneously within milliseconds; physical consistency verification confirmed that the DC power drop and voltage anomaly caused by the AC short circuit conformed to electrical laws; multi-source mutual verification was corroborated by multiple agents such as AC filters, protection, and converter transformers; in terms of credibility scoring, both the AC-side agent and the protection agent received high ratings. In summary, the "AC short circuit" hypothesis scored significantly higher than other hypotheses, resulting in a clear arbitration conclusion.

[0059] The handling process involves the station control agent retrieving the corresponding script. First, "rapidly blocking the fault filter branch" is executed to prevent continuous current surges; simultaneously, reactive power compensation reconfiguration is initiated to maintain reactive power voltage support within the station and prevent further bus voltage drops; if the short circuit range expands, the AC circuit breaker is activated for isolation. The operation ticket agent generates a confirmation list, which is executed after manual confirmation. After execution, if DC power gradually recovers and voltage stabilizes, operation is maintained and continuously monitored; if recovery fails, derating operation or standby switching is triggered. Throughout the process, the system monitors in real time and enters a fallback branch if the anomaly is not resolved to ensure operational safety. Finally, all handling information is stored in the audit package, providing a basis for subsequent training and parameter optimization.

[0060] Example 5: This embodiment also provides a computer device applicable to a multi-agent construction method for fault handling in high-voltage converter stations, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the multi-agent construction method for fault handling in high-voltage converter stations as proposed in the above embodiment.

[0061] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a multi-agent construction method for high-voltage converter station fault handling as proposed in the above embodiments.

[0062] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0063] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0064] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0065] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0066] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0067] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-agent construction method for fault handling of a high-voltage converter station, characterized in that, The method comprises the following steps: S1, mapping establishment of in-station equipment and agents; S2, uniform identification and event model construction; S3, evidence information organization and credibility calibration; S4, consistency arbitration and conflict grading processing; S5, fault disposal process arrangement and execution; S6, learning of operation experience and self-evolution of rules; S7, audit and traceability of disposal process. In the step S1, The division of real objects: a topological tree of equipment levels is established, and the tree is expanded layer by layer from the station level to the pole level, the valve tower level and the interval level. Each level node is bound to one or more agent proxies. The established agent proxy: the input signal source, the output instruction type, the fault disposal link participation range and the information interaction mode are determined. In the step S2, the hierarchical unique identification coding rules of the uniform identification system are constructed: The six levels of station area, pole area, interval, equipment, terminal and signal are expanded, and each level is coded by combining numbers and characters, so that any signal or equipment state can be uniquely determined through a complete identification link. The form of "station area code + pole area number + interval identification + equipment category + terminal number + signal serial number" is used for accurate expression. The event model is used for abstract description of different types of faults or abnormalities. The event model includes event category label, trigger condition, duration, influence range, severity level and priority.

2. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 1, characterized in that, The event model also introduces the logical relationship between events, including cause-effect relationship, corroboration relationship and contradiction relationship. In the step S3, the evidence unit includes event category, unique identification of equipment and signal, occurrence time window, numerical summary, source fingerprint and collection condition. The constraint rules are three constraints of time sequence, in-station topological accessibility and physical conservation.

3. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 1, characterized in that, According to the three constraints of time sequence, in-station topological accessibility and physical conservation, the cause-effect, corroboration and contradiction relationships between evidences are automatically generated to form evidence chains, including: ​ 4. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 1, characterized in that, ​ ​ ​ 5. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 1, characterized in that, ​ ​ 6. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 5, characterized in that, ​ First, filter according to time window adjacency and upper limit of propagation delay, then judge the influence chain according to the device connection path, and finally eliminate false associations by physical checks such as power balance, DC loop connectivity, and valve arm thermal-electric consistency; For the same identifier and the same time window, the repeated reports are de-duplicated by source fingerprints and merged according to the record completeness; For the recording and event fragmentation, the process of segment splicing-consistency check is adopted, the adjacent segments are spliced according to the pre-trigger buffer, and the waveform characteristics and event description are checked for consistency, thereby forming a loop-free evidence chain.

7. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 1, characterized in that, In step S4, the four-dimensional scoring framework includes: Time alignment, according to the evidence occurrence and station propagation delay upper limit check, time deviation within the allowed window gets high score; Physical consistency, check power balance, inter-electrode current difference, valve arm conduction relationship and topological reachable path, power imbalance below the upper limit and topological reachability are consistent; Multi-source mutual evidence, the evidence comes from independent channels, the more the sources and the more consistent the conclusions, the higher the score; Agent credibility, based on historical accuracy, clock synchronization quality and data integrity, long-term stable agent is given higher weight.

8. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 1, characterized in that, In step S5, the state flow of pre-checking-action arrangement-secondary checking-backup / upgrade includes: Pre-checking, topological reachability and isolation point position check, protection and blocking condition matching, power flow and reactive power margin test, local / remote control right and operation ticket authorization confirmation, time synchronization and signal consistency check; any one does not meet the requirements, it will be transferred to safe waiting or reduced capacity operation; Action arrangement, the treatment is decomposed into verifiable atomic actions, each action has clear target device, expected observation and success criterion, and sets hierarchical timeout and retry number; the actions are connected through conditional jump and interlock link to avoid mutual interference; Secondary checking, follow the closed loop of suggestion-confirmation-execution-feedback: the operation interface double confirms and then the command can be issued, and the feedback sources are consistent through two channels; Backup / upgrade, if new contradictory evidence appears or key criterion is not met, stop subsequent actions and enter the backup branch, reverse the reversible operations or execute the safety lock in the preset order.

9. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 1, characterized in that, In step S6, high-frequency evidence patterns are mined to form new rules, including: The system mines the high-frequency evidence subgraph and stable sequence in historical samples to obtain candidate causal constraints or treatment templates; Each candidate rule needs to pass two levels of checking, first verify the explanation power for past events and safety boundary in simulation playback, then enable in a small range, observe the trigger rate, accuracy and influence on operation indicators, and enable in the whole station after meeting the threshold; All parameter and rule changes are recorded in a versioned manner, including effective time, trigger background, comparison indicators and rollback path.

10. The multi-agent construction method for fault disposal of a high-voltage converter station according to claim 1, characterized in that, In step S7, the auditable and replayable materials include the following structured contents: Meta-information list, evidence directory, evidence relationship diagram, arbitration details, conflict classification and treatment path, treatment process timing, and learning and rule change records.

Images