A method and system for analyzing coordinated dc control and protection actions based on size model

Abstract

The application discloses a kind of based on size model cooperative direct current control protection action analysis method and system, it is related to direct current transmission control protection technical field.The method includes: mechanism small model is packaged as tool description set and constructs control protection prior vector knowledge base;Generate context context;And concurrently instantiate multiple agents to generate fault hypothesis set, and plan corresponding tool call sequence;Calculate the slot difference of tool parameter slot, when over threshold, correct;Mechanism small model is executed based on the parameter after correction, constructs physical conflict tensor and obtains conflict degree score;Compare the semantic consensus of fault hypothesis with calculation result, if not consistent, then determine pruning penalty weight in combination with negative constraint memory bank, calculate branch evaluation score to execute pruning, and iterate and regenerate until check passes.The application realizes the physical level decoupling of semantic logic and deterministic numerical calculation, eliminates the numerical illusion of large model, and improves the reliability of complex fault tracing.

Description

Technical Field

[0001] This invention relates to the field of DC transmission control and protection technology, and in particular to a method and system for analyzing DC control and protection actions based on large and small models. Background Technology

[0002] DC control and protection systems are core equipment in ultra-high voltage direct current (UHVDC) transmission projects, playing a crucial role in the stability, reliability, and transmission efficiency of power transmission. As power grids develop towards larger capacity, longer distances, and higher reliability, the protection logic of DC control and protection systems is becoming increasingly complex, and fault scenarios are exhibiting diverse and multifaceted characteristics, placing higher demands on the accurate analysis and rapid tracing of protection actions.

[0003] However, existing methods for analyzing the behavior of DC control and protection actions have the following bottlenecks: First, traditional expert systems based on fixed rules and manual backtracking are fragile. Traditional analysis methods mainly rely on manual backtracking by comparing with protection schematic diagrams, or online analysis using expert systems based on fixed rules and fuzzy logic. When faced with complex dynamic scenarios such as multiple fault superpositions, dynamic adjustment of settings, or atypical logic blocking (such as commutation failure predictive control logic), existing methods, due to the fragility of static rule bases, struggle to cover all dynamic scenarios and accurately pinpoint the underlying causes of protection maloperation or failure to operate.

[0004] Secondly, direct reasoning based on the Large Language Model (LLM) faces the challenges of "physical illusion" and logical distortion: In recent years, existing technologies have attempted to introduce the Large Language Model (LLM) for protection specification parsing and fault-aided analysis. However, the underlying autoregressive generation mechanism of LLM is essentially text generation based on probability statistics, lacking the inherent modeling and deterministic reasoning capabilities for protection logic gates (such as AND, OR, and NOT gates, flip-flops) and setting comparisons. When dealing with industrial-grade verification tasks involving strong physical constraints and zero fault tolerance, the model is highly susceptible to numerical calculation-related "physical illusions." For example, when analyzing differential protection actions, LLM may incorrectly determine that the current differential setting is valid or confuse the delay setting in sequential logic, leading to unreliable analysis conclusions that fail to meet the stringent requirements for safe grid operation.

[0005] Furthermore, while existing technologies include large-model application solutions for electrical equipment fault diagnosis, none have effectively addressed the aforementioned issues. For example, Chinese patent application CN202411519337.3 discloses a method, device, and equipment for electromagnetic transient fault analysis based on a large model, which can improve the accuracy of fault analysis and solve the problems of high computational load and insufficient analysis accuracy in traditional technologies. However, this method mainly targets electromagnetic transient fault analysis and does not cover the comprehensive question-and-answer and knowledge reasoning scenarios of DC control and protection systems. Chinese patent application CN202311403372.4 discloses an intelligent fault analysis method, system, and storage medium for UHVDC transmission systems based on a domestically developed artificial intelligence large-model framework. Through multi-source heterogeneous data fusion, CNN-Transformer hybrid network feature extraction, and knowledge graph-assisted localization, it achieves full automation of the fault diagnosis and report generation process. However, this method is not specifically optimized for the question-and-answer and knowledge reasoning scenarios of DC control and protection systems and lacks a deep integration of large and small models in terms of control and protection logic understanding, real-time interactive question-and-answer, and collaborative reasoning of complex protection strategies.

[0006] Therefore, there is an urgent need for a method and system for analyzing DC control and protection actions based on a large-scale model-based collaborative approach. This approach should deeply integrate the semantic planning capabilities of a large language model for processing complex relational information with the deterministic feature calculation and threshold determination capabilities of a small mechanism model under strong physical constraints. Under the premise of effectively isolating control flow and data flow, this approach should overcome the static rule fragility of traditional expert systems when facing complex DC scenarios such as multiple fault superposition and atypical logic blocking, and reduce the physical illusion of large language models when processing underlying numerical logic. This would enable accurate diagnosis and reliable tracing of complex protection actions in UHVDC systems. Summary of the Invention

[0007] The purpose of this invention is to address the fragility of traditional rules caused by the complexity of protection logic in ultra-high voltage direct current (UHVDC) transmission systems, and the numerical illusion and logical distortion caused by the lack of physical determinism in general large-scale models in industrial scenarios. This invention provides a method and system for analyzing DC control and protection actions based on a small-scale model-coordinated approach, comprising: encapsulating the mechanistic small-scale model into a tool description set and constructing a control and protection prior vector knowledge base; generating a contextual context; concurrently instantiating multiple agents to generate a set of fault hypotheses and planning the corresponding tool call sequence; calculating the slot divergence degree of tool parameter slots and correcting when it exceeds a threshold; the mechanistic small-scale model performing calculations based on the corrected parameters, constructing a physical conflict tensor and obtaining a conflict degree score; comparing the semantic consensus of the fault hypotheses with the calculation results, and if inconsistent, determining the pruning penalty weight by combining a negative constraint memory, calculating the branch evaluation score to perform pruning, and iteratively regenerating until verification is passed. This invention achieves physical-level decoupling between semantic logic and deterministic numerical calculation, significantly eliminating the numerical illusion of large-scale models and improving the reliability of complex fault tracing.

[0008] To achieve the above objectives, the present invention adopts the following technical solution.

[0009] In a first aspect, the present invention provides a method for analyzing DC control and protection actions based on a large-scale model, comprising: The mechanism small model is encapsulated into a tool description set containing parameter slots, and the pre-set power prior documents are segmented and vectorized to construct a control and protection prior vector knowledge base. Transient waveform data is decoded and stored in a time-series database for isolation, and a hybrid retrieval is performed in the control and protection prior vector knowledge base based on the input fault characteristics to generate contextual information. Based on the context, M agents are concurrently instantiated to generate a set of fault hypotheses, and combined with the tool description set, a corresponding set of tool call sequences is generated. Extract the parameter slots of each sequence in the tool call sequence set and calculate the slot divergence degree; when the slot divergence degree exceeds a preset threshold, correct the corresponding parameter slots based on the event sequence log; the mechanism mini-model performs calculations based on the corrected parameter slots and outputs the corresponding physical attributes, constructs a physical conflict tensor based on the physical attributes and calculation results of each agent, and calculates the conflict degree score; Extract the semantic consensus of the fault hypothesis set and verify its consistency with the calculation result; if the verification fails, determine the pruning penalty weight based on the currently invoked physical attribute and negative constraint memory, and calculate the branch evaluation score in combination with the conflict degree score; prune fault hypotheses whose branch evaluation scores are lower than the preset evaluation threshold, write the conflicting physical attributes into the negative constraint memory, and regenerate fault hypotheses until the verification passes or the preset maximum number of iterations is reached, and then output a diagnostic report.

[0010] Preferably, the mechanism model is encapsulated into a tool description set containing parameter slots, and the pre-set power prior documents are segmented and vectorized to construct a control and protection prior vector knowledge base, specifically including: The mechanism model is encapsulated into a computing component that conforms to a standard interface protocol; based on a structured data description pattern, the unique tool identifier, natural language function description, parameter slots, and corresponding data type restriction information of the computing component are defined; the definition results are serialized and injected into the context of a large language model to form the tool description set; The pre-set power prior document is segmented into multiple text fragments; each text fragment is converted into a corresponding knowledge vector using a text embedding model; a mapping index between text fragments and knowledge vectors is established to form the control and protection prior vector knowledge base.

[0011] Preferably, the transient waveform data is decoded and stored in a time-series database for isolation, and a hybrid retrieval is performed in the control and protection prior vector knowledge base based on the input fault characteristics to generate a contextual information, specifically including: Transient waveform data conforming to the general transient data exchange format standard is decoded into a time-series data matrix containing instantaneous sample values, and the time-series data matrix is ​​stored in the time-series database; Extract the fault features as query text, and convert the query text into a query vector; Calculate the dense vector similarity between the query vector and each knowledge vector in the control and protection prior vector knowledge base, and the sparse text similarity between the query text and each text fragment; Based on the set weight coefficients, the dense vector similarity and the sparse text similarity are weighted and summed to obtain a hybrid retrieval score; a preset number of text segments are selected in descending order of the hybrid retrieval score to generate the context.

[0012] Preferably, M agents are concurrently instantiated based on the context to generate a set of fault hypotheses, and combined with the tool description set to generate a corresponding set of tool call sequences, specifically including: Different analysis prompts are configured for the M agents, and fault reasoning is performed based on the context to generate corresponding fault hypotheses, forming the fault hypothesis set; For each fault hypothesis in the fault hypothesis set, the corresponding mechanism tool is selected according to the tool description set, and the calling order and parameter slots of the mechanism tool are determined to generate the corresponding tool calling sequence set.

[0013] Preferably, the parameter slots of each sequence in the tool call sequence set are extracted and the slot divergence degree is calculated; when the slot divergence degree exceeds a preset threshold, the corresponding parameter slot is corrected based on the event sequence log; the mechanism mini-model performs calculations based on the corrected parameter slots and outputs the corresponding physical attributes, constructs a physical conflict tensor based on the physical attributes and calculation results of each agent, and calculates the conflict degree score, specifically including: Extract the filling results of each agent for the same parameter slot, count the candidate values ​​and their corresponding occurrence frequencies, calculate the information entropy based on the occurrence frequencies, and normalize the information entropy to obtain the slot divergence degree. When the slot divergence exceeds a preset threshold, a timestamp matching retrieval is performed in the event sequence log based on the inverted index, and the matched real value is used to cover and correct the parameter slot. The mechanism mini-model performs calculations based on the corrected parameter slots to obtain the physical attributes and corresponding value sets of each agent; Based on the physical properties and the corresponding set of values, a physical conflict tensor is constructed. The elements of the physical conflict tensor are indicator functions used to characterize whether different agents obtain a specific value on the same physical property. For any physical attribute, the percentage of agents corresponding to each value is counted, and the conflict score is calculated based on the difference between the value and the sum of the squares of the percentages of each value.

[0014] Preferably, extracting the semantic consensus of the fault hypothesis set specifically includes: Extract isomorphic semantic conclusions from the fault hypotheses generated by each agent; The frequency of occurrence of each isomorphic semantic conclusion is counted, and the isomorphic semantic conclusion with the highest frequency is determined as the semantic consensus.

[0015] Preferably, the pruning penalty weight is determined based on the currently invoked physical attributes and the negative constraint memory, and a branch evaluation score is calculated in conjunction with the conflict score. Fault hypotheses with branch evaluation scores below a preset evaluation threshold are pruned, and the physical attributes that generate conflicts are written into the negative constraint memory. Specifically, this includes: The intersection operation is performed between the set of physical attributes corresponding to the current failure hypothesis and the set of historical default attributes stored in the negative constraint memory; the pruning penalty weight is calculated based on the ratio of the number of intersection elements to the total number of elements in the physical attribute set; the arithmetic mean of the conflict scores of each physical attribute invoked by the current failure hypothesis is calculated to obtain the average physical conflict score; the branch evaluation score is calculated by subtracting the product of the pruning penalty weight and the average physical conflict score from the basic verification score of the corresponding failure hypothesis. When the branch evaluation score is lower than the preset evaluation threshold and pruning is triggered, the current fault hypothesis that causes the conflict, the physical attribute identifier that causes the conflict, and the corresponding physical conflict cause are extracted; the current fault hypothesis, the physical attribute identifier, and the physical conflict cause are constructed into a structured negative constraint triplet and written into the negative constraint memory; when regenerating the fault hypothesis, if the newly generated fault hypothesis matches the negative constraint triplet already stored in the negative constraint memory, the pruning penalty weight of the corresponding fault hypothesis is increased.

[0016] Secondly, the present invention provides a DC control and protection action analysis system based on a large-scale model, comprising: The tool encapsulation and knowledge construction module is used to encapsulate the small mechanism model into a tool description set containing parameter slots, and to segment and vectorize the pre-set power prior documents to build a control and protection prior vector knowledge base. The data isolation and context generation module is used to decode transient waveform data and store it in a time series database for isolation, and to perform a mixed retrieval in the control and protection prior vector knowledge base based on the input fault characteristics to generate contextual information. The hypothesis generation and sequence planning module is used to concurrently instantiate M agents based on the context, generate a set of fault hypotheses, and combine them with the tool description set to generate a corresponding set of tool call sequences. The slot correction and mechanism execution module is used to extract the parameter slots of each sequence in the tool call sequence set and calculate the slot divergence degree; when the slot divergence degree exceeds a preset threshold, the corresponding parameter slot is corrected based on the event sequence log; the mechanism mini-model performs calculations based on the corrected parameter slots and outputs the corresponding physical attributes, constructs a physical conflict tensor based on the physical attributes and calculation results of each agent, and calculates the conflict degree score; The consistency verification and adaptive pruning module is used to extract the semantic consensus of the fault hypothesis set and perform consistency verification with the calculation results. If the verification fails, the pruning penalty weight is determined based on the currently invoked physical attributes and negative constraint memory, and the branch evaluation score is calculated in combination with the conflict degree score. Fault hypotheses with branch evaluation scores lower than the preset evaluation threshold are pruned, the conflicting physical attributes are written into the negative constraint memory, and fault hypotheses are regenerated until the verification passes or the preset maximum number of iterations is reached, at which point a diagnostic report is output.

[0017] Thirdly, the present invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method.

[0019] The beneficial effects of this invention are compared with those of the prior art: 1. This invention decouples the large language model from the underlying DC control and protection mechanism small model. The large language model only generates structured call sequences based on context and does not directly process massive transient waveform time-series data; numerical calculations (such as differential current feature extraction and setpoint threshold comparison) are all performed by the deterministic mechanism small model. This dual-track architecture avoids the probabilistic distortion that may occur in the numerical calculation stage of the large language model from the source, controls the physical illusions generated based on probability to a low level, and ensures the rigor and objectivity of fault calculation results.

[0020] 2. To address the semantic divergence problem that may occur during end-to-end generation of large language models, this invention introduces a slot divergence evaluation mechanism based on normalized information entropy. When parameter extraction is in a high-uncertainty state (i.e., the divergence is greater than a preset threshold), the system automatically transforms the task, using an inverted index to perform timestamp matching in the Sequence of Events (SOE) log and overwrite the high-entropy parameters. This mechanism effectively corrects the deviation in parameter mapping, enabling the system to ensure the accuracy of key parameter transmission under complex dynamic tasks and enhancing the robustness of the overall analysis chain.

[0021] 3. This invention constructs a multidimensional physical conflict tensor by comparing the semantic consensus of fault hypotheses with the objective calculation results of the mechanisms. In the concurrent exploration search space, branches that do not satisfy physical facts are penalized and pruned. Simultaneously, a negative constraint memory is introduced to record historical default attributes, preventing subsequent reasoning from repeatedly calculating on invalid paths. This forced backtracking and memory mechanism effectively converges the search space, significantly improving the accuracy and source tracing efficiency of comprehensive diagnosis of DC control and protection actions when dealing with multiple concurrent faults (such as valve short circuits superimposed with commutation failures).

[0022] 4. This invention adopts a pure reasoning paradigm, with large and small models interacting through standardized data description patterns. Universal large language models from different bases can be seamlessly integrated. When expanding new protection logic analysis capabilities, the system does not require costly parameter fine-tuning of the large language model; only the corresponding tool description set and mechanism components need to be updated to achieve capability integration. This effectively reduces the iteration and maintenance costs of this analysis scheme in large-scale industrial applications in substations. Attached Figure Description

[0023] Figure 1 This is a flowchart of a collaborative DC control and protection action analysis method based on a large-scale model. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0025] Example 1: like Figure 1 As shown, this invention provides a method for analyzing DC control and protection actions based on a large-scale model, comprising: In terms of system architecture deployment, it mainly includes a large language model at the station control layer, small mechanism models at the interval layer, and a bridging interface connecting the two. Specifically: the system deploys a large language model at the station control layer as the cognitive center of the system; deploys multiple highly lightweight small mechanism models at the interval layer as computational executors; and establishes a bridging interface protocol between the station control layer and the interval layer.

[0026] Among them, the large language model acts as the commander of the entire system. Its core responsibility is to read complex DC control and protection principle manuals, parse unstructured fault condition logs uploaded by operators, and explore different hypothesis branches by concurrently instantiating multiple agents to form multiple logical verification and exploration paths similar to a mind tree.

[0027] The proposed mechanism model does not include hard-coded links for specific sites, but is composed of professional power system protection mechanism algorithms (such as full wave transformation, sequence component calculation, differential current ratio calculation, etc.).

[0028] The station control layer and the interval layer communicate in full accordance with open application programming interface standards (such as OpenAPI 3.0). The mechanism small model exists in the form of microservices, waiting for the large language model to make accurate calls through structured data formats.

[0029] Based on the above collaborative architecture, the method formalizes the analysis and reasoning process of DC control and protection actions into a dynamic state machine model consisting of semantic planning, physical execution, and logical verification, specifically including the following evolution steps: Step 1: Encapsulate the small mechanism model into a tool description set containing parameter slots, and segment and vectorize the pre-set power prior documents to build a control and protection prior vector knowledge base.

[0030] 1.1 Standardized Registration of Mechanism Tools: During initialization, small physical mechanism models (such as control and protection algorithms like waveform analysis and sequence component calculation) are encapsulated into computational components conforming to standard interface protocols. These are then abstracted into semantic tools understandable by a Large Language Model (LLM) through a registration mechanism. Specifically, based on structured data description patterns (e.g., JSON Schema), the computational component is defined with a unique tool identifier, natural language function description, parameter slots, and corresponding data type restrictions. The parameter slots constrain the LLM's generation of structured tool call parameters and serve as information units for subsequent consistency analysis of multi-agent call results. For example, for a valve group overpressure calculation operator, the system defines its function using JSON Schema to calculate the voltage difference across the valve group within a specific time window and return a Boolean value. Its parameter slots are limited to index parameters such as start and end times, valve group number, and voltage threshold.

[0031] Subsequently, the above definition results are serialized and injected into the LLM context as part of the prompt words, or encapsulated through a standard function call interface to form the tool description set, so as to guide the LLM to output structured call instructions.

[0032] 1.2 Power Prior Document Vectorization Preprocessing: Obtain prior power documents, including static procedure documents and historical experience archives; segment the pre-set prior power documents into multiple text fragments. Text embedding models (such as pre-trained BERT or text2vec models) are used to convert each text fragment into a corresponding knowledge vector. This refers to high-dimensional dense vector representation. Ultimately, text fragments are constructed. With knowledge vectors A global mapping index is used to form the control and protection prior vector knowledge base. This mapping index is used for subsequent online retrieval to quickly locate and recall the original expert prior text fragments through vector similarity calculation.

[0033] Step 2: Decode the transient waveform data and store it in the time series database for isolation. Then, perform a mixed retrieval in the control and protection prior vector knowledge base according to the input fault characteristics to generate a contextual context.

[0034] 2.1 Physical-level isolation extraction of transient data: When a fault occurs, a transient waveform data packet strictly adhering to the Common Transient Data Exchange Format (COMTRADE) standard is first obtained from the DC control and protection site. This data packet contains at least a configuration configuration file, a binary or text data file, and an event header file. At this time, a data diversion and isolation mechanism is executed: through a built-in preprocessing script, the original waveform file is decoded into a high-resolution time-series data matrix. This data matrix contains instantaneous sampled values ​​of DC line voltage at millisecond or even microsecond resolution, instantaneous sampled values ​​of current on each side of the converter transformer, and switch status.

[0035] To avoid context overflow or hallucinations caused by massive high-frequency sampling data, the massive time-series data matrix is ​​physically isolated and stored in a time-series database. The large language model only triggers the mechanism mini-model to perform calculations by generating symbolic calling instructions containing device identifiers and time window parameters. The mechanism mini-model reads the corresponding time-series data from the time-series database according to the calling instructions and returns the calculation results, thereby decoupling cognitive decision-making from massive data calculations.

[0036] 2.2 Hybrid Fault Feature Retrieval and Contextual Injection: The fault features (such as alarm time, faulty equipment, and basic action protection type) are extracted from the event header file or station control layer alarm log. Then, these fault features are extracted as query text Q, and the text embedding model is used to transform them into a query vector q with completely consistent dimensions.

[0037] Next, a hybrid retrieval (considering both semantic similarity and literal matching) is performed in the control and protection prior vector knowledge base using Retrieval Enhancement Generation (RAG) technology. Specifically, the query vector q and each knowledge vector in the control and protection prior vector knowledge base are compared. The dense vector similarity between them, and the query text Q and each text fragment The system calculates the sparse text similarity between the dense vectors and the sparse text similarity based on a set weight coefficient, and then performs a weighted sum of these two similarities to obtain a hybrid retrieval score. The system calculates the hybrid retrieval score using the following formula:

[0038] in For query vector, Let i be the knowledge vector of the i-th segment in the knowledge base. , These are the weight coefficients for the dense vector similarity and the sparse text similarity, respectively, and they satisfy the following conditions: In this DC control and protection embodiment, it is preferred to set =0.7, =0.3, unstructured fault logs are often highly colloquial, prioritizing the use of high-dimensional semantic representations from large models to delineate the scope of procedures, while supplementing with smaller... Weights serve as hard lexical constraints to ensure that key physical ledger identifiers are not missed during recall. Represents the similarity of dense vectors, used to measure the similarity of query vectors. With knowledge base vectors The semantic relevance between them is calculated using cosine similarity; it can be expressed as:

[0039] in, and They represent vectors respectively with vector The L2 norm.

[0040] This represents the sparse text similarity based on the BM25 algorithm, used to measure query text similarity. With knowledge base text The statistical correlation of word frequencies between them can be expressed as:

[0041] Where t represents the segmented word items of the query text Q after word segmentation; Let be the inverse document frequency of term t; For the term t in the text fragment Word frequency in the text; For text fragments Word length; The average word count length of all text segments in the control-preservation prior vector knowledge base; and In this embodiment, the preferred settings for adjusting the hyperparameters of the BM25 algorithm are as follows: =1.5, =0.75.

[0042] After sorting the data in descending order of scores, the system selects a preset number of prior knowledge items, such as the top N, to generate the contextual injection LLM, providing expert common sense boundaries. N is an adjustable parameter, with the threshold N set to a range of [2, 5]. In this embodiment, N = 3. This ensures sufficient contextual information while avoiding the introduction of excessive irrelevant interference.

[0043] Step 3: Concurrently instantiate M agents based on the context, generate a set of fault hypotheses, and combine them with the tool description set to generate a corresponding set of tool call sequences.

[0044] Multiple agent branches are launched simultaneously at the initial inference node to generate concurrent hypotheses. Specifically, this includes: 3.1 Concurrency Hypothesis Instantiation and Viewpoint Bias Injection: For the context C input in step 2, M parallel agent branches are first instantiated. Preferably, the number of branches M ranges from [3,6], and in this embodiment, M=4 is preferred. The system configures different analysis prompts for each parallel agent, such as "analysis from the perspective of protection action strategy", "analysis from the perspective of setting value adjustment", "analysis from the perspective of waveform anomaly", and "analysis from the perspective of system operation status". Based on the context, fault reasoning is performed respectively to generate corresponding fault hypotheses. For example, hypotheses such as line grounding, commutation failure leading to a surge in differential current, or current transformer disconnection are proposed, thereby constructing a set of divergent fault hypotheses. .

[0045] 3.2 Confirmatory tool invoking sequential autoregressive programming: At the initial node, the large language model addresses each fault hypothesis in the set of fault hypotheses. Based on the tool description set, the corresponding mechanistic tool is selected, and the invocation order and parameter slots of the mechanistic tool are determined. Specifically, the large language model performs a heuristic bundle search in a preset tool invocation sequence space T, and autoregressively plans out the parameters used to verify the hypothesis. Tool call strategy sequence The plan generates the aforementioned call sequence. The joint conditional probability distribution can be expressed as:

[0046] in, This represents the tool selection action at step k in the i-th path, which includes the selection of specific mechanism tools and the filling of corresponding parameter slots; K is the call sequence. The total number of planning steps; θ is the pre-training parameter inherent in the large model; C represents the context output of step 2; This refers to the tool description set generated during registration in step 1.1; This represents the sequence of historical tool calls generated before the generation of the k-th action.

[0047] To overcome the uncontrollable formatting issue when large language models generate data freely, the aforementioned autoregressive generation process follows the inference and action cue word paradigm and enforces output format constraints through JSON Schema. In practice, the large model generates each action step... When this happens, key-value pairs containing "thinking logic", "selected tool identifier", and "structured parameter slots" must be output in sequence.

[0048] For example, if we assume If the condition is "commutation failure", then the corresponding high-probability tool call sequence is... The priority is to call the "Converter Valve DC Voltage Recording Analysis Operator," and the "Action Input" field in the generated JSON call instruction will be filled with the corresponding "Start and End Time" and "Threshold" parameter slots; if we assume... If the error is "current transformer disconnection", then its calling sequence is... The priority is to invoke the "zero-sequence current calculation operator". Through a parallel sampling mechanism, the system obtains a diverse set of hypotheses H, which is transformed into a set of multiple tool call sequences to be executed. This provides a wealth of physical verification targets for the subsequent execution of deterministic mechanisms.

[0049] Step 4: Extract the parameter slots of each sequence in the tool call sequence set and calculate the slot divergence degree; when the slot divergence degree exceeds a preset threshold, correct the corresponding parameter slots based on the event sequence log; the mechanism mini-model performs calculations based on the corrected parameter slots and outputs the corresponding physical attributes, constructs a physical conflict tensor based on the physical attributes and calculation results of each agent, and calculates the conflict degree score.

[0050] 4.1 Semantic-to-parameter intelligent mapping: The ReAct paradigm is introduced to drive each agent branch to call the corresponding mechanism calculation tools. The system automatically parses unstructured logs and performs intelligent slot extraction. For example, when the large language model parses an overpressure alarm message for an ultra-high pressure valve group at a specific time, it automatically extracts the timestamp and maps it to the time slot based on semantic understanding, identifies the corresponding valve group and maps it to the equipment slot, and retrieves the corresponding set value from the knowledge base generated by retrieval enhancement and maps it to the threshold slot, forming a structured request.

[0051] Furthermore, a slot divergence evaluation mechanism is introduced in the semantic-to-parameter intelligent mapping stage. When the large language model outputs a text sequence that fills a specific slot, the normalized information entropy X of the sequence is extracted simultaneously. The extraction method is as follows: assuming that M parallel agents all output prediction results for the same parameter slot, the extracted M deduplicated candidate values ​​are statistically analyzed. L Let candidate values ​​be... Frequency of occurrence Forming a frequency distribution Then calculate the information entropy of the distribution:

[0052] in, = / M is a candidate value The probability of occurrence satisfies Finally, the information entropy is normalized to... When L>1; if L=1, then the divergence degree X=0 is directly determined, meaning the value range of X is [0,1], and the normalized information entropy X is used as the slot divergence degree of that slot. When X is close to 1, it indicates that the outputs of each agent are highly divergent; when X is close to 0, it indicates that the outputs of each agent are highly consistent. This parameter is used for subsequent consensus detection and pruning decisions.

[0053] A threshold r is preset. When the slot divergence X ≤ r, it indicates that the outputs of each agent are highly consistent, the parameters extracted by the large language model have high confidence, the system directly retains the extraction result of the parameter slot, and triggers the execution stage of the downstream mechanism small model.

[0054] Conversely, if X>r for a certain key slot parameter, it indicates that the large model's extraction of this parameter is in a state of high uncertainty. In this case, the small model execution stage is not entered. Instead, the underlying slot adaptive alignment subroutine is triggered: the system transforms the high-entropy slot into a local retrieval task, that is, it abandons the generation result of the large language model and instead uses the inverted index to perform a secondary hard match retrieval on the strongly structured timestamp with the device tag in the Sequence of Events (SOE) log. After a successful match, the high-entropy prediction value of the large language model is directly covered and corrected with the true value obtained from the match, and then sent down to the interval layer for execution.

[0055] 4.2 Physical Mechanism Black-Box Execution: After receiving a structured parameter vector that has been directly verified or corrected, the mechanistic mini-model internally performs data reading, feature calculation, and threshold determination operations. For example, when an agent based on a differential protection hypothesis calls a differential current calculation tool, the mechanistic mini-model calculates based on the waveform data and returns the objective fact that the current differential current has not actually reached the set value. Since all calculations are completed within a defined symbol system, the large language model no longer directly performs probabilistic comparisons of specific numerical values, thus effectively avoiding numerical calculation illusions and reducing the error rate of logical reasoning.

[0056] The mechanistic mini-model no longer simply returns a single boolean value, but instead outputs a multi-dimensional physical conflict tensor to the backend. Let the set of output results of M parallel agents be . , where each output It contains a set of key physical property-value pairs. Define the conflict tensor. Let be a three-dimensional tensor with dimensions (A, V, I), corresponding to the number of physical attributes A, the range size V, and the number of agent branches I, respectively, and its coordinate index i satisfies 1 ≤ i ≤ M; where the elements Let be an indicator function, used to indicate the i-th agent in the _ ... Whether the i-th agent acquires a specific value v on a physical attribute, that is, when the i-th agent acquires a specific value v on a physical attribute. The value is 1 when the calculated result equals v, and 0 otherwise. For example, when the i-th agent calls the differential current tool to calculate the fact that "the differential current is not exceeded", the value state corresponding to this physical attribute is recorded as a specific discrete state identifier. Thus Set to 1.

[0057] For the a-th physical attribute, the conflict score The calculation formula is:

[0058] The conflict score It belongs to the [0,1) interval. The core function of this physical conflict tensor and its conflict degree score is to map the multi-agent mechanism execution results to the physical attribute dimension to construct a conflict tensor, and to use this conflict tensor to quantify the degree of consistency between different fault hypotheses and the objective facts of mechanism execution. The physical attributes are defined, for example, as action timing limit exceedance, electrical quantity magnitude margin (such as the difference between differential current and tripping threshold), and logic state XOR value, providing a quantitative basis for subsequent consensus detection and pruning decisions.

[0059] Step 5: Extract the semantic consensus of the fault hypothesis set and verify its consistency with the calculation result; if the verification fails, determine the pruning penalty weight based on the currently invoked physical attribute and negative constraint memory, and calculate the branch evaluation score in combination with the conflict degree score; prune fault hypotheses whose branch evaluation scores are lower than the preset evaluation threshold, write the conflicting physical attributes into the negative constraint memory, and regenerate fault hypotheses until the verification passes or the preset maximum number of iterations is reached, and then output a diagnostic report.

[0060] 5.1 Consensus Extraction and Two-Way Comparison: The system summarizes the preliminary conclusions of all parallel branches. Let the diagnostic conclusions output by the M agents be sets. Extract isomorphic semantic conclusions from the fault hypotheses generated by each agent, i.e., natural language texts where the core fault type and action result representation are the same. Count the frequency of occurrence of each isomorphic semantic conclusion, and determine the isomorphic semantic conclusion with the highest frequency as the semantic consensus. This can be modeled as finding the most frequent isomorphic semantic conclusions in a set:

[0061] In the formula, Let I be the set of candidate conclusions after deduplication in set D; The parentheses are indicator functions; they take the value 1 if the identity within the parentheses is true, and 0 otherwise. This indicates that two natural language diagnostic conclusions are semantically isomorphic in a pre-built power industry semantic ontology library. That is, when the core fault type and protection action result are completely consistent after the two natural language conclusions are extracted by pre-trained power entity relations, they are determined to be isomorphic.

[0062] Subsequently, the semantic consensus (Assuming the differential protection action is triggered by a commutation failure) The system performs a consistency check with the calculation results returned by the mechanism mini-model (the differential setting did not actually exceed the limit). If the two are inconsistent, the system determines that the physical consistency status is unsuccessful.

[0063] 5.2 State handling and closed-loop: (1) Consistency passed: If the semantic consensus is completely consistent with the physical facts, the system determines the logical loop, stops the iteration and outputs a structured fault diagnosis report.

[0064] (2) Adaptive backtracking: If the verification fails, the system will convert the failed path and objective facts into "negative constraint memory" and feed it back to the starting node.

[0065] First, the system writes the currently conflicting physical properties into a pre-defined negative constraint memory. Specifically, when the execution result of a certain branch violates physical constraints, the key information of that branch is represented as a triple. Write to negative constraint memory ,in For the current fault hypothesis, To generate a physical attribute identifier for default, The cause of the conflict is as follows: When regenerating the fault hypothesis in subsequent reasoning, if the newly generated fault hypothesis matches an existing negative constraint triple in the negative constraint memory, the system increases the pruning penalty weight of the corresponding fault hypothesis, for example, by dynamically increasing the base penalty coefficient. Or cumulative penalty decay factor This allows the repeated invalid hypotheses to be quickly removed in subsequent evaluations, thus preventing the system from repeatedly consuming computing power on paths that have been disproven.

[0066] Meanwhile, the system determines the pruning penalty weights based on the physical properties corresponding to the current fault hypothesis and the negative constraint memory. The specific calculation process is as follows: The set of physical properties corresponding to the current fault hypothesis... With the negative constraint memory library The intersection set is obtained by performing an intersection operation on the historical default attribute set stored in the database. (Characterizing the degree of overlap between the current hypothesis and negative memories); subsequently, based on the number of intersection elements... Total number of elements in the physical property set The ratio is used to calculate the pruning penalty weight. The calculation formula is as follows:

[0067] in, The base penalty coefficient is λ=0.5 in this embodiment, and γ is the cumulative penalty decay factor (γ=1.0 in this embodiment). For the current branch The set of physical attribute identifiers to be invoked. For set With memory bank The intersection of the set of physical attribute identifiers contained in all historical default triples (i.e., the degree of overlap between the current branch and the negative memory).

[0068] Next, the system calculates the branch evaluation score based on the conflict score obtained in step 4.2. Specifically, a heuristic evaluation function is constructed as follows, which assigns the basic verification score of the corresponding fault hypothesis to the evaluation function. Subtract the pruning penalty weight Compared with average physical conflict score The product of these factors is used to calculate the branch evaluation, which is divided into:

[0069] in, For large language models in generating this fault hypothesis The basic verification score assigned at that time is set to 1.0 in this embodiment; The physical set representing the current hypothesis call Conflict scores of various physical attributes The arithmetic mean of the values. This formula means that the final score of the hypothesis is significantly suppressed by the product of the penalty weight and the degree of objective physical conflict.

[0070] When the branch evaluation score Below the preset evaluation threshold When the fault hypothesis branch is removed, its information is written into the negative constraint memory.

[0071] Subsequently, the large language model regenerates the fault hypothesis based on the updated negative constraint memory and replans the corresponding tool invocation path; the mechanism small model performs physical verification again, and the consistency verification module re-executes the semantic consensus and mechanism fact comparison. This process is repeated until the semantic inference result and the physical verification result reach consistency, or the preset maximum number of iterations is reached. (In this embodiment, it is preferred to set) After that, the system terminates the iteration and outputs the final fault diagnosis report.

[0072] In some embodiments, the conflict results are further classified according to the degree of physical conflict feedback from the mechanism small model, and a differentiated backtracking strategy is adopted.

[0073] First, a conflict level index is constructed based on the deviation corresponding to the conflict attribute. When the conflict level index is lower than the first threshold, it is judged as a minor conflict. The system maintains the current fault hypothesis unchanged, only adjusts the sampling constraint parameters in the fault hypothesis generation stage, and regenerates the tool call sequence in the neighborhood of the current fault time window. The underlying physical basis of this mechanism is that the clock synchronization jitter at the microsecond to millisecond level between the merging unit (MU) and the waveform recording device in the DC power system field is very likely to occur, or the inherent mechanical opening and closing transient delay of the circuit breaker. At this time, the fault hypothesis and physical formula inferred by the large language model based on the log are logically completely valid. The operator verification fails only because the phase of the underlying sampled waveform has a slight drift. By shifting the time window in the neighborhood, the real transient peak can be accurately captured, thereby avoiding misjudging the clock synchronization error as a mechanism logic error.

[0074] When the conflict level index is higher than the first threshold but lower than the second threshold, it is judged as a moderate conflict. The system retains the current fault hypothesis and re-extracts, corrects, and verifies the mechanism for the parameter slots that caused the conflict. The underlying physical basis of this mechanism is that there is a cognitive gap between the "static program setting value" and the "dynamic actual value" in the power field. For example, the "commutation failure" hypothesis derived by the large language model is completely consistent with the current electrical quantity characteristics. However, because the field dispatcher has temporarily issued a setting area switching command, the actual voltage threshold value recorded in the log is inconsistent with the standard setting ledger in the RAG program library. At this time, the actual electrical physical condition has occurred. The system only needs to trigger the slot adaptive correction program and perform inverted hard matching in the Sequence of Events (SOE) log to make the operator call of the small mechanism model return to the physical reality.

[0075] When the conflict level index exceeds the second threshold, it is judged as a severe conflict. The system directly prunes the current fault hypothesis branch and regenerates the fault hypothesis based on the negative constraint memory. The underlying physical basis of this mechanism is that the conflict tensor calculated by the small mechanism model has seriously violated physical hard constraints such as Kirchhoff's Current Law (KCL), the law of conservation of energy, or DC topological connectivity. For example, the large language model infers "pole-one line grounding" based on log false alarms, but after the mechanism operator performs full-wave Fourier transform, it is found that the current at this end and the other end of the line is strictly conserved within the allowable range of measurement error. There is absolutely no possibility of energy leakage in the objective physical entity. This kind of conflict is a serious semantic illusion generated by the cognitive center and must be forcibly blocked and the branch information must be entered into the negative constraint memory.

[0076] Through the aforementioned hierarchical backtracking mechanism, adaptive closed-loop control of the reasoning search space based on physical conflict results is achieved.

[0077] Example 2: This invention provides a DC control and protection action analysis system based on a large-scale model collaborative model, comprising: To further illustrate the practical application effects of this invention, the following uses a DC valve group overvoltage protection verification scenario as an example to elaborate in detail the parameter slot mapping and collaborative execution workflow between the large language model and the small mechanism model in this system, specifically including: The tool encapsulation and knowledge construction module is used to encapsulate the small mechanism model into a tool description set containing parameter slots, and to segment and vectorize the pre-set power prior documents to build a control and protection prior vector knowledge base. The data isolation and context generation module is used to decode transient waveform data and store it in a time series database for isolation, and to perform a mixed retrieval in the control and protection prior vector knowledge base based on the input fault characteristics to generate contextual information. The hypothesis generation and sequence planning module is used to concurrently instantiate M agents based on the context, generate a set of fault hypotheses, and combine them with the tool description set to generate a corresponding set of tool call sequences. The slot correction and mechanism execution module is used to extract the parameter slots of each sequence in the tool call sequence set and calculate the slot divergence degree; when the slot divergence degree exceeds a preset threshold, the corresponding parameter slot is corrected based on the event sequence log; the mechanism mini-model performs calculations based on the corrected parameter slots and outputs the corresponding physical attributes, constructs a physical conflict tensor based on the physical attributes and calculation results of each agent, and calculates the conflict degree score; The consistency verification and adaptive pruning module is used to extract the semantic consensus of the fault hypothesis set and perform consistency verification with the calculation results. If the verification fails, the pruning penalty weight is determined based on the currently invoked physical attributes and negative constraint memory, and the branch evaluation score is calculated in combination with the conflict degree score. Fault hypotheses with branch evaluation scores lower than the preset evaluation threshold are pruned, the conflicting physical attributes are written into the negative constraint memory, and fault hypotheses are regenerated until the verification passes or the preset maximum number of iterations is reached, at which point a diagnostic report is output.

[0078] 1. System initialization and mechanism tool abstract registration The system defines and registers valve group overpressure calculation operators using a standardized data description schema (such as JSON Schema). The definition includes: tool name, natural language function description, such as: explaining that the tool is used to read voltage data on both sides of the DC converter valve group within a specific time window, calculate the voltage difference and compare it with the set value, return the objective fact of whether the overpressure trip criterion has been met, and the required start and end times, valve group number, voltage threshold and other parameter slot formats.

[0079] 2. Semantic parsing and slot alignment of large language models When the large language model parses an overpressure alarm message from an ultra-high pressure valve group at a specific time, the system, acting as a semantic planning hub, automatically extracts the timestamp based on semantic understanding and maps it to a time slot, identifies the corresponding valve group and maps it to an equipment slot, and retrieves the corresponding setpoint value from the control and protection prior vector knowledge base and maps it to a threshold slot. Simultaneously, the system calculates the slot divergence of the extracted parameters. If the slot divergence exceeds a preset threshold, a local retrieval task is automatically triggered, utilizing Sequence of Events (SOE) logs for real value matching and overwrite correction, thereby generating a high-confidence structured tool invocation request.

[0080] 3. Mechanistic Small Model Deterministic Physical Execution After receiving the structured parameter vector, the small mechanistic model performs data reading, feature calculation, and threshold determination operations within its own confines. Since such feature extraction and comparison are completed within a defined physical symbol system, the large language model no longer directly participates in the probabilistic comparison of specific numerical values, thus effectively avoiding the probabilistic fluctuations in numerical operations inherent in pure large language models and keeping the illusion of numerical computation at an extremely low level.

[0081] This invention compares the overall performance of a large language model as the semantic planning hub in this embodiment. In the aforementioned collaborative analysis architecture, the robustness and overall performance of the system highly depend on the large language model's adherence to interface (API) specifications and its tool invocation capabilities. Preferably, this embodiment deploys a general-purpose large language model with high tool invocation accuracy at the site control layer. The specific technical effects of this configuration are as follows: (1) High accuracy and compliance: The model can accurately convert natural language instructions into structured JSON parameters that conform to the schema specification. Combined with the adaptive pruning mechanism of multidimensional physical conflict tensor, it achieves the best accuracy in fault diagnosis and action verification. (2) Low illusion rate and physical isolation: By decoupling the physical level of the mechanism small model, the numerical generation task is transformed into deterministic mechanism execution, reducing the physical illusion in the computation process; at the same time, combined with the built-in vector knowledge base retrieval, the factual reliability of the initial semantic routing stage is guaranteed. (3) High robustness and low migration cost: With the standardized tool routing and parameter mapping mechanism, this system does not need to perform expensive full fine-tuning on large models (i.e., training-free fine-tuning), which reduces the industrial deployment overhead and the migration and adaptation cost of new protection logic.

[0082] In summary, the collaborative analysis architecture of large and small models constructed in this invention cleverly combines the precise filling technology of instruction slots with a deterministic mechanism verification mechanism. While retaining the ability of large language models to broadly and generalize their understanding of unstructured text, it perfectly inherits the physical rigor of power system mechanism calculations, achieving an intelligent leap in the field of DC control and protection action analysis without the need for fine-tuning.

[0083] Example 3: A terminal, comprising a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method.

[0084] Example 4: A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method.

[0085] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0086] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0087] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0088] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for analyzing DC control and protection actions based on a large-scale model collaborative approach, characterized in that, include: The mechanism small model is encapsulated into a tool description set containing parameter slots, and the pre-set power prior documents are segmented and vectorized to construct a control and protection prior vector knowledge base. Transient waveform data is decoded and stored in a time-series database for isolation, and a hybrid retrieval is performed in the control and protection prior vector knowledge base based on the input fault characteristics to generate contextual information. Based on the context, M agents are concurrently instantiated to generate a set of fault hypotheses, and combined with the tool description set, a corresponding set of tool call sequences is generated. Extract the parameter slots of each sequence in the tool call sequence set and calculate the slot divergence degree; when the slot divergence degree exceeds a preset threshold, correct the corresponding parameter slot based on the event sequence log; The mechanism mini-model performs calculations based on the corrected parameter slots and outputs the corresponding physical properties. It constructs a physical conflict tensor based on the physical properties and calculation results of each agent and calculates the conflict degree score. Extract the semantic consensus of the fault hypothesis set and verify its consistency with the calculation results; If the verification fails, the pruning penalty weight is determined based on the currently invoked physical attributes and negative constraint memory, and the branch evaluation score is calculated in combination with the conflict score; fault hypotheses with branch evaluation scores lower than the preset evaluation threshold are pruned, the conflicting physical attributes are written into the negative constraint memory, and fault hypotheses are regenerated until the verification passes or the preset maximum number of iterations is reached, and then a diagnostic report is output.

2. The method for analyzing DC control and protection actions based on a large-scale model according to claim 1, characterized in that, The mechanism model is encapsulated into a tool description set containing parameter slots, and pre-built power prior documents are segmented and vectorized to construct a control and protection prior vector knowledge base, specifically including: The mechanism model is encapsulated into a computing component that conforms to a standard interface protocol; based on a structured data description pattern, the unique tool identifier, natural language function description, parameter slots, and corresponding data type restriction information of the computing component are defined; the definition results are serialized and injected into the context of a large language model to form the tool description set; The pre-set power prior document is segmented into multiple text fragments; each text fragment is converted into a corresponding knowledge vector using a text embedding model; a mapping index between text fragments and knowledge vectors is established to form the control and protection prior vector knowledge base.

3. The method for analyzing DC control and protection actions based on a large-scale model according to claim 2, characterized in that, Transient waveform data is decoded and stored in a time-series database for isolation. Based on the input fault characteristics, a hybrid retrieval is performed in the control and protection prior vector knowledge base to generate a contextual information, specifically including: Transient waveform data conforming to the general transient data exchange format standard is decoded into a time-series data matrix containing instantaneous sample values, and the time-series data matrix is ​​stored in the time-series database; Extract the fault features as query text, and convert the query text into a query vector; Calculate the dense vector similarity between the query vector and each knowledge vector in the control and protection prior vector knowledge base, and the sparse text similarity between the query text and each text fragment; Based on the set weight coefficients, the dense vector similarity and the sparse text similarity are weighted and summed to obtain a hybrid retrieval score; a preset number of text segments are selected in descending order of the hybrid retrieval score to generate the context.

4. The method for analyzing DC control and protection actions based on a large-scale model according to claim 3, characterized in that, Based on the aforementioned context, M agents are concurrently instantiated to generate a set of fault hypotheses. Combined with the tool description set, a corresponding set of tool call sequences is generated, specifically including: Different analysis prompts are configured for the M agents, and fault reasoning is performed based on the context to generate corresponding fault hypotheses, forming the fault hypothesis set; For each fault hypothesis in the fault hypothesis set, the corresponding mechanism tool is selected according to the tool description set, and the calling order and parameter slots of the mechanism tool are determined to generate the corresponding tool calling sequence set.

5. The method for analyzing DC control and protection actions based on a large-scale model according to claim 4, characterized in that, Extract the parameter slots of each sequence in the tool call sequence set and calculate the slot divergence degree; when the slot divergence degree exceeds a preset threshold, correct the corresponding parameter slot based on the event sequence log; The mechanistic mini-model performs calculations based on the corrected parameter slots and outputs the corresponding physical properties. A physical conflict tensor is constructed based on the physical properties and calculation results for each agent, and a conflict degree score is calculated, specifically including: Extract the filling results of each agent for the same parameter slot, count the candidate values ​​and their corresponding occurrence frequencies, calculate the information entropy based on the occurrence frequencies, and normalize the information entropy to obtain the slot divergence degree. When the slot divergence exceeds a preset threshold, a timestamp matching retrieval is performed in the event sequence log based on the inverted index, and the matched real value is used to cover and correct the parameter slot. The mechanism mini-model performs calculations based on the corrected parameter slots to obtain the physical attributes and corresponding value sets of each agent; Based on the physical properties and the corresponding set of values, a physical conflict tensor is constructed. The elements of the physical conflict tensor are indicator functions used to characterize whether different agents obtain a specific value on the same physical property. For any physical attribute, the percentage of agents corresponding to each value is counted, and the conflict score is calculated based on the difference between the value and the sum of the squares of the percentages of each value.

6. The method for analyzing DC control and protection actions based on a large-scale model according to claim 5, characterized in that, Extracting the semantic consensus of the set of fault hypotheses specifically includes: Extract isomorphic semantic conclusions from the fault hypotheses generated by each agent; The frequency of occurrence of each isomorphic semantic conclusion is counted, and the isomorphic semantic conclusion with the highest frequency is determined as the semantic consensus.

7. The method for analyzing DC control and protection actions based on a large-scale model according to claim 6, characterized in that, Based on the currently invoked physical attributes and the negative constraint memory, the pruning penalty weight is determined, and the branch evaluation score is calculated in conjunction with the conflict score. Fault hypotheses with branch evaluation scores below a preset evaluation threshold are pruned, and the physical attributes that cause conflict are written into the negative constraint memory. Specifically, this includes: The intersection operation is performed between the set of physical attributes corresponding to the current failure hypothesis and the set of historical default attributes stored in the negative constraint memory; the pruning penalty weight is calculated based on the ratio of the number of intersection elements to the total number of elements in the physical attribute set; the arithmetic mean of the conflict scores of each physical attribute invoked by the current failure hypothesis is calculated to obtain the average physical conflict score; the branch evaluation score is calculated by subtracting the product of the pruning penalty weight and the average physical conflict score from the basic verification score of the corresponding failure hypothesis. When the branch evaluation score is lower than the preset evaluation threshold and pruning is triggered, the current fault hypothesis that causes the conflict, the physical attribute identifier that causes the conflict, and the corresponding physical conflict cause are extracted; the current fault hypothesis, the physical attribute identifier, and the physical conflict cause are constructed into a structured negative constraint triplet and written into the negative constraint memory; when regenerating the fault hypothesis, if the newly generated fault hypothesis matches the negative constraint triplet already stored in the negative constraint memory, the pruning penalty weight of the corresponding fault hypothesis is increased.

8. A DC control and protection action analysis system based on a large-scale model, characterized in that, include: The tool encapsulation and knowledge construction module is used to encapsulate the small mechanism model into a tool description set containing parameter slots, and to segment and vectorize the pre-set power prior documents to build a control and protection prior vector knowledge base. The data isolation and context generation module is used to decode transient waveform data and store it in a time series database for isolation, and to perform a mixed retrieval in the control and protection prior vector knowledge base based on the input fault characteristics to generate contextual information. The hypothesis generation and sequence planning module is used to concurrently instantiate M agents based on the context, generate a set of fault hypotheses, and combine them with the tool description set to generate a corresponding set of tool call sequences. The slot correction and mechanism execution module is used to extract the parameter slots of each sequence in the tool call sequence set and calculate the slot divergence degree; when the slot divergence degree exceeds a preset threshold, the corresponding parameter slot is corrected based on the event sequence log. The mechanism mini-model performs calculations based on the corrected parameter slots and outputs the corresponding physical properties. Based on the physical properties and calculation results of each agent, a physical conflict tensor is constructed, and the conflict degree score is calculated. The consistency verification and adaptive pruning module is used to extract the semantic consensus of the fault hypothesis set and perform consistency verification with the calculation results. If the verification fails, the pruning penalty weight is determined based on the currently invoked physical attributes and negative constraint memory, and the branch evaluation score is calculated in combination with the conflict score; fault hypotheses with branch evaluation scores lower than the preset evaluation threshold are pruned, the conflicting physical attributes are written into the negative constraint memory, and fault hypotheses are regenerated until the verification passes or the preset maximum number of iterations is reached, and then a diagnostic report is output.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method according to any one of claims 1-7.

Images