Switching operation ticket generation method based on large electric power model and electronic equipment

Abstract

The invention provides a switching operation ticket generation method based on an electric power large model and electronic equipment, and the method comprises the steps: receiving a dispatching instruction, carrying out the semantic analysis of the dispatching instruction through a large model pre-trained by an electric power corpus, and extracting a structured operation element; retrieving real-time state information of related equipment from an equipment knowledge base; retrieving matched sample ticket data from a sample ticket knowledge base; and inputting the operation elements, the real-time state information and the sample ticket data into the large model to generate a switching operation ticket. According to the scheme, the scheduling instruction is subjected to semantic analysis by adopting the large model pre-trained by the power corpus, and the equipment topological relation data generated by automatic analysis of the model from the electrical wiring diagram and the equipment state information obtained by real-time linkage with the anti-error system are combined; according to the method, deep fusion of the scheduling instruction and the topology and real-time state of the field equipment is realized, and the logic consistency and the safety compliance in the operation ticket generation process are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of operation ticket generation, in particular to a switching operation ticket generation method based on a power large model and an electronic device. BACKGROUND

[0002] In the field of power system automation, switching operation ticket is the core file to ensure the safe operation of power grid, and its accuracy and standardization directly affect the operation safety. With the development of artificial intelligence technology, natural language processing technology based on large model has been applied to the field of operation ticket generation, which converts dispatching instructions into structured operation sequences through semantic understanding. At the same time, knowledge graph and vector database technology are used to construct device information library and historical sample ticket library to provide support for device topology relationship and operation rules.

[0003] Although the related technology attempts to introduce a large language model to realize automatic generation of operation ticket, it is derived from a static knowledge base architecture, and the model only relies on general corpus training, lacking a deep coordination mechanism with the structured knowledge base in the power field, and cannot effectively associate dispatching instructions with field device topology relationship and historical operation experience, resulting in logical consistency and safety compliance defects in the generated operation ticket, which is difficult to meet the high reliability requirements of power system. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a switching operation ticket generation method based on a power large model and an electronic device to solve the above problems.

[0005] In a first aspect, the embodiments of the present application provide a switching operation ticket generation method based on a power large model, which comprises: receiving a dispatching instruction, and using a large model pre-trained by power corpus to perform semantic analysis on the dispatching instruction to extract structured operation elements; based on the operation elements, retrieving real-time state information of related devices from a device knowledge base; wherein the device knowledge base stores device topology relationship data generated by the large model analyzing an electrical wiring diagram, and the device knowledge base is linked with a misoperation prevention system to synchronize device state data; based on the operation elements and the real-time state information, retrieving matching sample ticket data from a sample ticket knowledge base; wherein the sample ticket knowledge base stores typical operation ticket data, and the sample ticket knowledge base is configured to receive execution feedback results for dynamic updating; inputting the operation elements, the real-time state information and the sample ticket data into the large model to generate a switching operation ticket.

[0006] In the implementation process of the scheme, the scheduling instruction is subjected to semantic analysis by using a large model pre-trained by power corpus, and the device topology relationship data automatically generated by the model from the electrical wiring diagram, and the device state information obtained in real time through linkage with the anti-misoperation system are combined, thereby realizing deep fusion of the scheduling instruction with the field device topology and real-time state, effectively improving the logical consistency and safety compliance in the operation order generation process, and reducing the risk of misoperation caused by incomplete device information or missing topology relationship. On the other hand, the feedback result of execution is received to update dynamically by configuring a sample order knowledge base, thereby building a continuous learning optimization mechanism based on actual execution effect, so that the above-mentioned switching operation order generation method can continuously iterate typical operation order data according to operation experience, thereby enhancing the adaptability to different power grid operation scenarios and scheduling instructions, and realizing dynamic evolution of the knowledge base and continuous improvement of the model performance.

[0007] In an implementation form of the first aspect, the inputting the operation element, the real-time state information and the sample order data into the large model to generate a switching operation order comprises: decomposing a switching operation order generation task into a plurality of subtasks; wherein each of the subtasks is used to generate an operation sequence of a single device; respectively invoking a subtask agent deployed on a plurality of computing nodes to execute the subtasks to generate corresponding sub-operation sequences; and integrating each of the sub-operation sequences into a complete switching operation order according to the device operation sequence.

[0008] In the implementation process of the scheme, the switching operation order generation task is decomposed into subtasks for a single device, and the subtask agents deployed on a plurality of computing nodes are invoked to execute in parallel, thereby effectively breaking through the Token length limit of a single large model in processing long sequence operation orders, improving the feasibility and response speed of large-scale switching operation order generation; the modular task decomposition mechanism enables each subtask agent to run independently and not to interfere with each other, thereby enhancing the scalability and fault tolerance of the above-mentioned switching operation order generation method, and meanwhile, the sub-operation sequences are integrated according to the device operation sequence, thereby ensuring the logical coherence and execution timing accuracy of the complete operation order finally generated.

[0009] In an implementation form of the first aspect, the subtask agent executes the subtask to generate the sub-operation sequence, comprising: retrieving a sample order knowledge segment matched with the corresponding device in the sample order knowledge base; wherein the sample order knowledge segment comprises pre-operation check items and post-operation confirmation items; and replacing the device information in the sample order knowledge segment with the actual device information in the operation element to generate the sub-operation sequence.

[0010] In the implementation process of the scheme, by searching the sample ticket knowledge base for a sample ticket knowledge fragment matched with the corresponding device, directly reusing the standardized operation logic containing pre-operation check items and post-operation confirmation items, the uncertainty caused by the complete reliance of the sub-task agent on model generation is avoided, and the normativeness and accuracy of the sub-operation sequence generation are improved. On the other hand, by using a device information replacement mechanism to replace the device information in the sample ticket knowledge fragment with the actual device information in the operation element, accurate adaptation from a typical template to a specific device instance is realized, the operation logic consistency is maintained, the generated content is ensured to be accurately matched with the target device, and the applicability and reliability of the operation ticket are enhanced.

[0011] In an implementation form of the first aspect, the retrieving, from the sample ticket knowledge base, a matched sample ticket data based on the operation element and the real-time state information comprises: performing vectorization processing on the operation element to generate an instruction feature vector; calculating the similarity of the instruction feature vector with each typical operation ticket data in the sample ticket knowledge base, and obtaining a candidate sample ticket set based on the similarity filtering; filtering the typical operation ticket in the candidate sample ticket set based on the task attribute in the operation element to obtain a target sample ticket set; wherein the task attribute comprises a device type, a voltage level and a risk level; and performing weighted sorting on the typical operation ticket in the target sample ticket set based on the update time, the historical execution times, the execution success rate and the similarity with the instruction feature vector to obtain an optimal operation ticket.

[0012] In the implementation process of the scheme, by using a multi-level retrieval mechanism of vectorization similarity filtering-task attribute filtering-multi-index weighted sorting, and especially by introducing dynamic business indexes such as historical execution times and execution success rate to participate in weight calculation, effective balance between semantic matching accuracy and practical reliability of sample ticket retrieval is realized, and the accuracy of optimal operation ticket selection is improved. On the other hand, by taking business attributes such as device type, voltage level and risk level as filtering conditions, and combining dynamic indexes such as execution success rate and timeliness for weighted sorting, the sample ticket retrieval process deeply integrates business scenario constraints and operation experience feedback, and the business adaptability and safety compliance of the retrieval result to a specific scheduling instruction are enhanced.

[0013] In an implementation form of the first aspect, the method further comprises: obtaining an execution feedback result of the optimal operation ticket; if the execution feedback result is failure, pushing an operation ticket reselection prompt to a user, and obtaining an optimal operation ticket reselected by the user from the target sample ticket set to continue the execution of the switching operation ticket generation process; and if the execution feedback result is success, increasing the weights of the historical execution times and the execution success rate in the weighted sorting.

[0014] In the implementation process of the scheme, the execution feedback result of the optimal operation ticket is obtained, and a failure reselection mechanism is established, a man-machine collaborative continuous learning closed loop is constructed, so that the above switching operation ticket generation method can feed back the execution failure case as a negative sample to the retrieval link, correct the retrieval deviation through manual intervention, improve the adaptability and fault tolerance to complex boundary scenarios. On the other hand, by dynamically increasing the weight coefficient of the historical execution times and the execution success rate in the weighted sorting when the execution is successful, the online evaluation of the sample ticket quality and the adaptive optimization of the sorting strategy based on the actual operation effect are realized, so that the high-quality sample ticket is preferentially selected in the subsequent retrieval, thereby continuously improving the overall success rate and reliability of the switching operation ticket generation.

[0015] In an implementation form of the first aspect, the method further includes: obtaining an execution feedback result of the optimal operation ticket; if the execution feedback result is successful, storing the switching operation ticket and the corresponding dispatching instruction to the sample ticket knowledge base; if the execution feedback result is failed, recording error information of the current failure operation; updating parameters of the large model based on the new data in the sample ticket knowledge base and / or the error information.

[0016] In the implementation process of the scheme, by establishing a bidirectional closed loop mechanism of execution feedback and knowledge base accumulation, the operation ticket with successful execution is automatically deposited to the sample ticket knowledge base to continuously expand the sample size, and the error information of the failure case is recorded to form a negative sample library, the dynamic evolution of the knowledge base based on the actual operation effect is realized, and the self-improvement ability of the above switching operation ticket generation method to the power grid operation scene is improved. On the other hand, by continuously updating and incrementally training the large model based on the new sample ticket data and the error information, the model can quickly adapt to new scenarios such as power grid topology changes, equipment commissioning and dispatching rule evolution, avoiding the low efficiency problem of traditional offline full retraining, effectively enhancing the timeliness and generalization ability of the model, and reducing the long-term maintenance cost.

[0017] In an implementation form of the first aspect, the method further includes: constructing the device knowledge base, and / or, constructing the sample ticket knowledge base, wherein: the method of constructing the device knowledge base includes: obtaining an electrical wiring diagram; using the power large model to analyze and reason the electrical wiring diagram, extracting device information and topological relationship; generating structured data and storing it to the device knowledge base; interacting with the anti-misoperation system to obtain device state information and store it to the device knowledge base; the method of constructing the sample ticket knowledge base includes: obtaining the typical operation ticket data and storing it to the sample ticket knowledge base; based on the execution feedback result, dynamically updating the newly generated switching operation ticket to the sample ticket knowledge base; using the new data in the sample ticket knowledge base to iteratively train the large model.

[0018] In the implementation process of the above scheme, by using a large model to analyze and infer the electrical wiring diagram, the device information and topological relationship are automatically extracted, the automatic construction of the device knowledge base is realized, the manual maintenance cost and error rate are reduced, and through real-time state synchronization with the anti-misoperation system, the consistency of the device knowledge base and the actual power grid operation state is ensured. On the other hand, by dynamically updating the sample ticket knowledge base based on the execution feedback result and using the new data to iteratively train the large model, a continuous learning closed loop is constructed, so that the model can quickly adapt to new scenarios such as power grid topology changes, device commissioning and dispatching rule evolution, effectively improving the long-term adaptability and generation accuracy of the above switching operation ticket generation method.

[0019] In an implementation form of the first aspect, the method further includes: performing anti-misoperation checking on the switching operation ticket; wherein the anti-misoperation checking includes at least one of device name checking, operation range checking, pre-operation device state checking, post-operation device state checking, missing item checking, operation terminology checking, and misoperation checking; and / or performing anti-misoperation simulation rehearsal on the switching operation ticket; and / or pushing an artificial review prompt to a user and obtaining an artificial review result of the user.

[0020] In the implementation process of the above scheme, by configuring a multi-dimensional AI anti-misoperation checking mechanism covering device name consistency, operation range compliance, pre-operation and post-operation state logical integrity, terminology specification, and anti-misoperation logic conflict detection, comprehensive cross-validation of the generated operation ticket is realized, effectively intercepting complex logical errors that cannot be identified by single rule checking, improving the detection rate and accuracy of automatic checking. On the other hand, by introducing a dual review mechanism of anti-misoperation simulation rehearsal and artificial review prompt, a depth defense system of intelligent checking-virtual rehearsal-artificial confirmation is constructed, so that the operation ticket needs to be verified multiple times before entering the execution link, minimizing the safety risks caused by model illusion or incomplete rule coverage, and ensuring the safety and reliability of the operation ticket in real scenarios.

[0021] In an implementation form of the first aspect, the method further includes: obtaining historical data related to power operation; wherein the historical data includes historical operation tickets, operation procedures, safety procedures, and professional terminology library; preprocessing the historical data to construct a training data set containing operation logic, device state changes, and operation rules; performing domain-enhanced training on the large model based on the training data set to obtain the trained large model.

[0022] In the implementation process of the above scheme, by systematically collecting multi-source heterogeneous data covering historical operation tickets, operation procedures, safety procedures and professional term library, and constructing a structured training data set containing operation logic, device state change and operation rules, the knowledge system in the power field is modeled in all directions, providing a high-quality data basis for subsequent model training. On the other hand, by performing domain-enhanced training on the large model based on the training data set, the general large model deeply integrates the professional knowledge and business logic of power operation, improves the understanding accuracy of the model to the dispatching instruction and the standardization of the operation ticket generation, and effectively alleviates the illusion problem of the general model in the power vertical field and the risk that the generated content does not comply with the safety procedures.

[0023] In a second aspect, the embodiments of the present application provide an electronic device, comprising: a processor, a memory and a communication bus, wherein the processor and the memory complete mutual communication through the communication bus; the memory stores computer program instructions that can be executed by the processor, and when the computer program instructions are read and run by the processor, the method provided by the first aspect or any one of the possible implementation manners of the first aspect is executed. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0025] Figure 1 The flowchart of the method for generating switching operation tickets based on a power large model provided by the embodiments of the present application is shown in the figure. Figure 2 The flowchart of pre-training the large model in some application scenarios provided by the embodiments of the present application is shown in the figure. Figure 3 The initialization flowchart of the device knowledge base in some application scenarios provided by the embodiments of the present application is shown in the figure. Figure 4 The flowchart of converting the wiring diagram into SVG format by the vector conversion tool in some application scenarios provided by the embodiments of the present application is shown in the figure. Figure 5 The initialization flowchart of the sample ticket knowledge base in some application scenarios provided by the embodiments of the present application is shown in the figure. Figure 6 The flowchart of the intelligent ticket generation scheme in some application scenarios provided by the embodiments of the present application is shown in the figure. Figure 7A schematic diagram of an architecture of a switching operation order generation system based on a power large model is provided for an embodiment of the present application. Figure 8 A schematic diagram of a structure of an electronic device is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

[0027] Power system switching operation involves multi-device collaborative switching and strict timing control. The operation order needs to accurately describe the device state change sequence, operation conditions and safety check items, and is a key technical document to ensure the safe operation of the power grid. With the increasing complexity of the power grid structure and the increasing response time requirement of the dispatching instructions, the traditional manual preparation of the operation order mode faces the bottleneck of low efficiency, easy error and strong experience dependence, and it is urgent to realize the automatic generation of the operation order through intelligent means. The development of large model technology provides a new path for the semantic understanding and text generation of dispatching instructions, but the fusion mode of its professional knowledge system in the power field becomes the core problem restricting the application effect.

[0028] The prior art mainly adopts a general large model combined with rule processing to realize operation order generation. Its typical implementation modes include: fine-tuning a general pre-trained model, injecting domain knowledge by using limited power corpus; or storing device account and typical operation order text through an external knowledge base, and matching related cases by vector similarity retrieval; some schemes also introduce a knowledge graph to describe the correlation of devices, and combine a rule engine to check the operation logic. Most of the prior art injects external knowledge into the model reasoning process through retrieval enhancement generation paradigm, but the knowledge base construction and model training are independent of each other, and cannot form a deep collaborative joint optimization mechanism. In the prior art, although the general model has a semantic understanding basis, its parameter space does not deeply integrate multi-dimensional domain knowledge such as device topology constraints, operation logic rules and historical execution feedback, resulting in the separation of the generation process and the knowledge base retrieval. The specific performance is as follows: the model cannot actively associate the real-time state and topology constraints in the device knowledge base when analyzing the dispatching instruction, and the generated operation sequence is easy to violate the five-prevention logic; the retrieval results of the sample ticket knowledge base are only used as text splicing input, and cannot form gradient back propagation with the model reasoning process, resulting in low knowledge utilization rate; the knowledge base update depends on offline manual maintenance, and lacks online iteration ability based on execution result feedback. This defect causes insufficient accuracy and safety compliance of the operation order generation, and it is difficult to adapt to the scene expansion demand brought by the dynamic change of the power grid topology and the operation of new devices.

[0029] In view of this, the embodiment of the present application provides a switching operation order generation method based on a power large model. The method performs semantic analysis on dispatching instructions by using a large model pre-trained on power corpus, and combines device topology relationship data automatically generated by the model from an electrical wiring diagram, and device state information obtained in real time through linkage with a misoperation prevention system, thereby realizing deep integration of dispatching instructions with field device topology and real-time state, effectively improving logical consistency and safety compliance in the operation order generation process, and reducing the risk of misoperation caused by incomplete device information or missing topology relationship. On the other hand, by configuring a sample order knowledge base to receive execution feedback results for dynamic updating, a continuous learning optimization mechanism based on actual execution effect is constructed, so that the above-mentioned switching operation order generation method can continuously iterate typical operation order data according to operation experience, enhance the adaptability to different power grid operation scenarios and dispatching instructions, and realize dynamic evolution of the knowledge base and continuous improvement of the model performance.

[0030] Please refer to Figure 1 The embodiment of the present application provides a flowchart of the switching operation order generation method based on a power large model. The switching operation order generation method based on a power large model provided by the embodiment of the present application can be applied to an electronic device, which can include a server, a PC, a tablet computer or a smart phone, etc. physical device, or can also be a virtual machine or a container, etc. virtual device, and the electronic device can be a single device, or a combination of multiple devices or a cluster of a large number of devices. The above-mentioned switching operation order generation method based on a power large model can include: Step S110: receiving a dispatching instruction, and performing semantic analysis on the dispatching instruction by using a large model pre-trained on power corpus, and extracting structured operation elements.

[0031] The above-mentioned dispatching instruction is an electrical device state change command expressed in natural language form issued by a power system dispatching institution to a plant station operation on-duty personnel, and its content usually includes structured elements such as ordered plant station, voltage level, operation device, operation type, device initial state and target state. The dispatching instruction can be delivered in text or voice form, and follows the standard terminology system and expression specification in power dispatching regulations, and is used to clearly specify the operation requirement of converting a certain electrical device or device group from the current running state (such as running, hot standby, cold standby) to another target state (such as maintenance, test).

[0032] The operation elements refer to a set of structured data fields extracted by semantic analysis of the dispatching instructions to represent the core attributes of the switching operation task. This set is organized in key-value pairs and can include key dimensions such as the ordered station, voltage level, main device identifier, operation device number, operation type (such as closing, opening, switching), device type, initial state, and target state. Each element follows the standardized coding specifications of power dispatching regulations and has clear business semantics and value constraints, thus forming the minimum information unit that drives subsequent device knowledge base retrieval, sample ticket knowledge base matching, and large model generation.

[0033] The large model (also known as the power large model) is a generative pre-training language model based on the Transformer architecture. Its parameter space has been domain-adapted pre-trained on vertical domain corpus such as power system dispatching regulations, operation sample tickets, and device account books, with deep semantic understanding and contextual reasoning capabilities for power industry terminology. After receiving the dispatching instruction text in natural language form, the large model can encode and dependency analyze key entities such as station name, device number, operation verb, and state parameters in the instruction through self-attention mechanism, and then generate a sequence of structured operation elements including ordered station, voltage level, operation device, operation type, initial state, and target state through the decoder, achieving semantic mapping and information extraction from unstructured text to standardized business data format. The large model can use DeepSeek model and other mature models in related technologies. These mature large models are generally built based on the Transformer encoder-decoder architecture, with a core composed of multiple layers of stacked self-attention mechanism and position-aware feedforward neural network. The large model includes input embedding layer, multiple layers of Transformer blocks, and output projection layer. Each layer of Transformer block can integrate multi-head self-attention sublayer and position-wise feedforward network sublayer, supplemented by layer normalization and residual connection to achieve stable gradient propagation. The multi-head attention mechanism calculates multiple attention heads in parallel to capture the dependency relationships in different semantic subspaces of the dispatching instruction text, realizing cross-distance association modeling of key entities such as station, device, and state.

[0034] The semantic analysis process of the large model takes the user input scheduling instruction text as a query question, drives the large model to perform conditional generative reasoning, specifically: after the scheduling instruction is converted into a Token sequence by a tokenizer, it is input into a model encoder to generate a context-aware hidden state representation, and a decoder predicts a sequence of structured operation element key-value pairs Token by Token in a self-recurrent manner. This process dynamically aligns the natural language description in the instruction text with the device state transition rules, operation terminology mapping relationships and safety procedure constraints learned during the pre-training phase in the power domain through the attention mechanism, thereby converting unstructured free text into formalized data structures that conform to the semantics of the scheduling procedure, achieving accurate semantic mapping and information extraction from questions to answers.

[0035] Optionally, the above-mentioned power large model-based switching operation order generation method can further include: obtaining historical data related to power operation; wherein the historical data includes historical operation orders, operation procedures, safety procedures and a professional term library; preprocessing the historical data to construct a training data set containing operation logic, device state changes and operation rules; performing domain-enhanced training on the large model based on the training data set to obtain a trained large model.

[0036] The above-mentioned historical operation order refers to the switching operation order electronic file that has been actually executed in the power system dispatching operation, and its data structure includes fields such as operation task description, operation step sequence, device double number, pre-and post-operation state marker, execution timestamp, operator signature and execution result feedback marker. Mainly derived from the operation order management module historical database of the power grid dispatching control system, the five-prevention operation record log of the substation monitoring system and the work order archive library in the power production management information system, etc. The above-mentioned operation procedure refers to a collection of technical documents that standardize the operation process and steps of electrical equipment, and its content covers chapters such as typical operation task decomposition, operation condition judgment logic, device interlocking relationship constraints and abnormal condition handling plans. Main sources include power industry standards, enterprise technical standards, implementation details of dispatching procedures prepared by regional dispatching agencies, and on-site operation special procedure documents of each voltage level substation, etc. The above-mentioned safety procedure is essentially a mandatory system document system that guarantees the safety of electrical operation personnel and equipment, and its provisions clearly define operation permission conditions, safety distance requirements, electric verification and grounding specifications, personal protective equipment standards and accident reporting processes, etc. Main sources include the power safety work procedure and its supporting interpretation documents, industry anti-accident measures compilation, safety technology notices issued by the safety supervision departments of power enterprises, and experience summary reports formed based on historical misoperation event analysis. The above-mentioned professional term library refers to a collection of standardized vocabulary and abbreviations in the field of power dispatching operation, and its entries cover semantic categories such as device categories, state definitions, operation verbs, protection device functions and fault types. Main sources include current relevant national standards, industry standards, internal standards of power enterprises, etc.

[0037] The preprocessing of the power operation related historical data can include data cleaning, data slicing and semantic segmentation, and data completion, etc., wherein: (1) data cleaning processing: integrity check, duplicate record removal and outlier removal can be performed, and noise data and missing fields caused by system input errors, communication interference or format incompatibility are removed to ensure the purity of the input corpus. (2) Data slicing and semantic segmentation: long text operation tickets can be disassembled into atomic operation step units according to operation task boundaries, device interval attribution and logical paragraph identifiers, and based on dependency syntax analysis and domain lexicon recognition, semantic components such as operation instructions, condition judgments and safety precautions are labeled in detail to realize the mapping of unstructured text to standardized sequences. (3) Data completion: for ambiguous samples with implicit topological relationship, fuzzy state transition or non-standard term usage, experts in the field of power dispatch intervene to perform manual sorting and ambiguity resolution, by supplementing missing context constraints, correcting device number alias mapping and labeling abnormal working condition handling logic.

[0038] After preprocessing the power operation related historical data, a vertical domain dataset for pre-training of power large models can be constructed based on the preprocessed multi-source heterogeneous data through structured packaging and task adaptation conversion. The dataset adopts a multi-task hybrid architecture: operation logic is encoded through operation reasoning procedure sequences that explicitly label operation premise conditions, device state transition logic and safety check rules; device state changes store dynamic information such as device state enumeration values, associated device topology indexes and state transition timestamps in key-value structures; operation rules can be expressed through operation step dependency relationships and execution order logic encoded in the form of directed acyclic graphs or state machines, as well as voltage level mapping tables, risk level labels and other power domain specific constraint fields. Data instances follow a unified specification, each sample contains input instruction text, structured operation elements, target operation ticket sequence and complete annotations of intermediate reasoning process, forming multi-dimensional training samples that support supervised fine-tuning, thought chain learning and instruction compliance ability training, meeting the vertical domain adaptation needs of large models in the power operation scenario.

[0039] The pre-training of the above large model can be performed in a paradigm combining full-parameter fine-tuning and incremental learning to perform domain-enhanced training of the large model. The structured power operation dataset is input into the large model base in a supervised learning mode. The training process optimizes the performance of the model on the three target tasks of dispatching instruction understanding, device state inference, and safety rule compliance by minimizing the cross-entropy loss between the predicted operation sequence and the real operation ticket. The training dataset is organized in the form of a thought chain, explicitly encoding the logical flow of operation premise-execution step-posterior confirmation, so that the model learns the mapping function from natural language intent to formal operation sequence. After training, the vertical domain power large model is obtained, whose parameter space has deeply captured the device state transition probability distribution, operation terminology semantic association, and five-prevention rule constraint mode, and has the ability to generate compliant operation tickets under zero-sample or few-sample conditions. When using the DeepSeek architecture, the pre-training can use the Mixture-of-Experts (MoE) sparse activation mechanism to trigger only the expert subnetworks related to the power operation logic, and realize computational efficiency optimization through dynamic routing of the gating network. The model integrates Multi-Head Latent Attention (MLA) to reduce the key-value cache overhead of long operation sequences, and introduces the Multi-Token Prediction (MTP) strategy to simultaneously predict multiple subsequent operation items, thereby improving the training signal density. The training process uses the FP8 mixed precision calculation framework to accelerate gradient update and reduce memory occupancy while maintaining numerical stability, and finally generates a high-capacity low-power model that realizes the synergistic improvement of semantic understanding depth and generation efficiency on the power operation ticket generation task.

[0040] The above scheme systematically collects multi-source heterogeneous data covering historical operation tickets, operation procedures, safety procedures, and professional terminology libraries, and constructs a structured training dataset containing operation logic, device state changes, and operation rules, thereby realizing all-around modeling of the power domain knowledge system and providing a high-quality data foundation for subsequent model training. On the other hand, by performing domain-enhanced training of the large model based on the training dataset, the general large model deeply integrates the professional knowledge and business logic of power operations, thereby improving the understanding accuracy of the model for dispatching instructions and the standardization of operation ticket generation, effectively alleviating the hallucination problem of general models in the power vertical domain and the risk of generated content not complying with safety procedures.

[0041] Step S120: Based on the operation elements, real-time state information of the related devices is retrieved from the device knowledge base; wherein the device knowledge base stores device topology relationship data generated by the large model analyzing the electrical wiring diagram, and the device knowledge base is linked with the anti-misoperation system to synchronize the device state data.

[0042] Step S130: retrieving matched sample ticket data from a sample ticket knowledge base based on the operation element and real-time state information; wherein the sample ticket knowledge base stores typical operation ticket data, and the sample ticket knowledge base is configured to receive execution feedback results for dynamic updating.

[0043] The device knowledge base is a structured data storage system for the power switching operation scene. The core content includes device node information, inter-device topology connection relationship and electrical property data automatically extracted from electrical wiring diagrams by image analysis and inference technology of a large model, and dynamic synchronization of device operating states (such as closing, opening, maintenance, etc.) and position events through real-time interface linkage with the anti-misoperation system, to form a device digital mirror integrating static topology structure and dynamic state information. The device knowledge base can use a vector storage and indexing mechanism to support fast retrieval based on device name, number or topology path, and provide basic data support and real-time state constraints at the device level for operation ticket generation. The sample ticket knowledge base is a typical operation ticket storage system based on continuous evolution of historical execution data. The initial data set is derived from typical switching operation tickets in the power grid dispatching control system that have been audited and archived. Each sample ticket contains complete operation step sequences, device state transition logic and safety check rules. The sample ticket knowledge base is configured with a dynamic update interface that can receive execution feedback results. The successful execution of the operation ticket is automatically stored in the database and the index weight is increased. At the same time, the failure case is marked as a negative sample, supporting sample ticket quality self-learning evaluation based on execution frequency and success rate, forming a sample ticket resource pool with timeliness and reliability ranking, which is used to provide high-confidence reference templates for sample ticket generation of similar dispatching instructions.

[0044] The anti-misoperation system, also known as the anti-misoperation system, is a mandatory safety check device deployed at the substation level. Its core function is to prevent operating personnel from violating the five-prevention rules during switching operation through hardware locking and software logic dual mechanisms, i.e., preventing misoperation of opening and closing circuit breakers, preventing opening and closing disconnectors with load, preventing connecting ground wire with live line, preventing closing circuit breakers with ground wire, and preventing entering live interval. The anti-misoperation system can communicate with field intelligent electronic devices in real time through relevant protocols or private protocols, collect real position remote signaling, electrical interlocking state and logic node information of circuit breakers, disconnectors and grounding switches, and has a five-prevention logic judgment module based on a production rule engine. Virtual pre-performance and conflict detection are performed before any operation sequence. When the operation ticket instruction sequence triggers a safety constraint violation, the anti-misoperation system can output a locking signal through hardware relay contacts or software control instructions to block the field operation mechanism electric circuit, and provide the checking results and device state data through a standardized API interface, providing real-time and reliable device state data source and anti-misoperation rule checking services for the upper intelligent ticket generation system.

[0045] Optionally, the above-mentioned power big model-based switching operation ticket generation method further comprises: constructing a device knowledge base, and / or constructing a sample ticket knowledge base, wherein: The way of constructing the device knowledge base comprises: obtaining an electrical wiring diagram; using the power big model to analyze and reason the electrical wiring diagram, extracting device information and topological relationship; generating structured data and storing it to the device knowledge base; interacting with the anti-misoperation system to obtain device state information and store it to the device knowledge base; The way of constructing the sample ticket knowledge base comprises: obtaining typical operation ticket data and storing it to the sample ticket knowledge base; based on the execution feedback result, dynamically updating the newly generated switching operation ticket to the sample ticket knowledge base; using the new data in the sample ticket knowledge base to iteratively train the big model.

[0046] The construction method of the above-mentioned device knowledge base can include: uploading the substation original wiring diagram file (usually in CAD or raster format) to the Dify workflow management platform, triggering the power big model-driven image analysis pipeline: the power big model calls the vector conversion tool to translate the wiring diagram into SVG (Scalable Vector Graphics, Scalable Vector Graphics) format, extracts the device name, number, type (circuit breaker, disconnector, grounding switch, etc.) and electrical connection topological relationship (node connection order, bus-bar-bay attribution) from the SVG file through path element recognition and graph semantic reasoning technology, and encodes the extraction results into a JSON structured description file conforming to the specification. Then, the JSON data is layered written into the device knowledge base, which includes the primary device model library, the secondary device model library and the topological relationship library, forming a three-dimensional knowledge base with graph structure expression capability. To ensure the dynamic nature of knowledge, the electronic device executing the above-mentioned switching operation ticket generation method can establish a real-time data channel with the anti-misoperation system through the RESTful API interface, pull or subscribe device location remote signaling, state measurement value and logic node data regularly, and deploy an event-driven monitoring agent, which automatically triggers API calls to incrementally update the corresponding state fields in the device knowledge base when receiving the anti-misoperation system-pushed position change event or fault alarm, thereby ensuring millisecond-level synchronization consistency between device information and actual operation conditions.

[0047] The above-mentioned Dify workflow management platform is a low-code orchestration and integrated development platform for big model applications, and its functional core is to provide a visual workflow designer and an automated process scheduling engine, supporting users to connect heterogeneous task nodes such as file uploading, model calling, data conversion, knowledge base storage into an end-to-end processing pipeline in a drag-and-drop manner.

[0048] The construction of the above-mentioned sample ticket knowledge base adopts a two-stage strategy of initial import and dynamic update. In the initialization stage, the ETL (Extraction Transformation Loading) tool extracts the audited and archived typical switching operation tickets from the power grid dispatching control system or production management system, covering different voltage levels, equipment types and operation task types. Each record contains a standardized operation step sequence, device state transition path, operation timestamp and associated device ID, which is stored in a vector database after vectorization coding to form a basic template library. The dynamic update mechanism is deployed in the intelligent ticket generation process closed loop, automatically adding the operation tickets that have passed the verification and the optimized sequence corrected by manual audit to the sample ticket knowledge base through incremental writing, and updating the metadata such as execution frequency and success rate of the sample ticket, realizing the quality self-evaluation and priority reordering of the sample ticket based on the running effect. The sample ticket knowledge base and the equipment knowledge base are associated and mapped through the equipment topology index and the operation object code. Each operation step in the sample ticket is bound to the unique device identifier and real-time state constraint in the equipment knowledge base, forming a dynamic knowledge system of equipment-topology-sample ticket.

[0049] The above-mentioned scheme automatically extracts equipment information and topology relationship by using large models to analyze and infer electrical wiring diagrams, realizes the automatic construction of equipment knowledge base, reduces the cost and error rate of manual maintenance, and ensures the consistency of equipment knowledge base and actual power grid operation state through real-time state synchronization with the anti-misoperation system. On the other hand, by dynamically updating the sample ticket knowledge base based on the execution feedback results and iteratively training the large model with the new data, a continuous learning closed loop is constructed, so that the model can quickly adapt to new scenarios such as power grid topology changes, equipment commissioning and dispatching rule evolution, effectively improving the long-term adaptability and generation accuracy of the above-mentioned switching operation ticket generation method.

[0050] The implementation of retrieving real-time state information of the relevant device from the device knowledge base based on the operation element in the above step S120 can include: receiving a power dispatch instruction in the form of natural language input by a user, which generally specifies the substation, main device and target state transition requirement in a text description. Sequence labeling and dependency syntax analysis are performed on the dispatch instruction using a trained large model to extract structured operation elements including ordered substation identification, voltage level, main device type, operation device number, operation type verb, current device state and target state. Based on the element set, a vectorized query request or a structured query statement is constructed, and a retrieval is initiated through a query interface of the device knowledge base. After the device knowledge base responds to the query request, complete data records of the target device are returned, including device standardized name, logical node number, topological association path generated by the large model from the electrical wiring diagram (such as interval device, bus ownership relationship, etc.), and current position telemetered value and state measurement data obtained in real time through the API of the anti-misoperation system, thereby providing complete device-level information support for subsequent sample ticket matching and operation sequence generation.

[0051] Optionally, the above retrieval of matched sample ticket data from the sample ticket knowledge base based on the operation element and real-time state information includes: performing vectorization processing on the operation element to generate an instruction feature vector; calculating the similarity of the instruction feature vector with each typical operation ticket data in the sample ticket knowledge base, and obtaining a candidate sample ticket set based on the similarity; filtering the typical operation tickets in the candidate sample ticket set based on the task attributes in the operation element to obtain a target sample ticket set; wherein the task attributes include device type, voltage level and risk level; and performing weighted sorting on the typical operation tickets in the target sample ticket set based on the update time, historical execution times, execution success rate and similarity with the instruction feature vector to obtain an optimal operation ticket.

[0052] The vectorization processing on the operation element can be implemented through a dedicated embedding model. After the operation element is input into the embedding model, the embedding model can capture the semantic association and context dependency relationship between terms through a multi-layer self-attention mechanism, map the discrete text to a high-dimensional continuous vector space, and generate a dense vector representation containing the operation intent, device object and state transition semantics. The vector serves as a distributed semantic signature of the instruction, retains the structured features and logical relationships of the dispatch instruction in the power business scenario, and provides a numerical basis for subsequent similarity calculation with the typical operation ticket vector in the sample ticket knowledge base.

[0053] The similarity calculation between the instruction feature vector and the typical operation ticket data in the sample ticket knowledge base can adopt various measurement forms. The cosine similarity measures the direction consistency in the high-dimensional semantic space by calculating the cosine value of the included angle between two vectors. The Euclidean distance measures the straight-line distance between vectors, reflecting the absolute difference between the instruction features and the sample ticket features in each dimension. The dot product similarity comprehensively considers the direction consistency and the modulus scale through the vector inner product operation. The Manhattan distance is sensitive to the local difference between sparse vectors by calculating the sum of the absolute differences in each dimension. In addition, the above-mentioned various similarity forms can be used alone or weightedly fused to form a multi-scale and multi-angle similarity evaluation system.

[0054] Two strategies of threshold screening and / or sorting screening can be adopted to obtain the candidate sample ticket set. The threshold screening rule can be set as a static similarity threshold, that is, only the sample tickets with a similarity score higher than a preset empirical value (such as 0.75) are retained to ensure the basic semantic relevance of the candidate set. Alternatively, a dynamic threshold mechanism is adopted to adaptively calculate a quantile threshold (such as taking the upper 90% quantile point of the similarity distribution) according to the semantic complexity of the current query instruction and the statistical characteristics of the sample ticket library distribution to cope with the difference in retrieval difficulty of different operation tasks. The sorting screening strategy selects the K sample tickets with the highest similarity to constitute the candidate set through the fixed number strategy. The value of K can be dynamically adjusted according to the device complexity of the operation task, for example, K=5 for a single device task and K=10 for a multi-device association task to balance the retrieval accuracy and the calculation overhead. The mixed rule can be implemented as a two-stage screening: first, preliminarily filter the sample tickets with a loose threshold, and then perform Top-K secondary selection on the remaining samples; or an iterative strategy is adopted, the candidate set is obtained by Top-K initially, and if the highest similarity is lower than the confidence threshold, the backoff mechanism is triggered to expand the retrieval range.

[0055] The candidate sample set is filtered based on the task attributes in the operation elements, aiming to optimize the business adaptability of the semantic retrieval result through explicit business constraints. The filtering process mainly includes: first, extracting key attributes such as device type code, voltage level identifier and risk level label from structured operation elements, and constructing a multi-dimensional constraint query vector. Then, traverse the candidate sample set, read the pre-defined attribute fields in the typical operation ticket data annotation, including the device type enumeration, voltage level range and risk level classification applicable to the sample ticket. The filtering rule adopts a combination strategy of strict matching and hierarchical compatibility: the device type must be completely matched or meet the inheritance relationship (such as "circuit breaker" operation compatible with "smart circuit breaker" sample ticket); the voltage level performs accurate equality or upward compatibility (such as 35kV operation can refer to 110kV sample ticket, but not vice versa); the risk level implements downward compatible filtering, only retaining the sample ticket with risk level equal to or higher than the current operation task, ensuring that the selected template has sufficient safety redundancy. For the sample tickets that do not pass the attribute consistency check, they are excluded from the candidate set, and finally the target sample set that meets the three business constraints of device, voltage and risk is formed, thereby adding business logic correctness guarantee on the basis of semantic similarity.

[0056] The typical operation ticket in the target sample set can be weighted and sorted by constructing a comprehensive scoring function through a multi-dimensional business indicator fusion mechanism. Before constructing the comprehensive scoring function, the four core indicators of update time, historical execution times, execution success rate and instruction similarity can be normalized respectively to eliminate dimensional differences: the update time indicator is mapped through a time decay function, giving higher timeliness weight to recently updated sample tickets; the historical execution times are converted through a non-linear saturation degree, ensuring the reliability of high-frequency sample tickets while avoiding overfitting to a single pattern; the execution success rate directly uses the statistical probability value and adds a confidence correction factor, smoothed by sample size as the denominator; the instruction similarity keeps the original cosine similarity or the probability score converted from the Euclidean distance. The update time, historical execution times, execution success rate and instruction similarity are weighted and fused to form a comprehensive score, where the weight coefficients of each dimension can be dynamically adjusted according to the risk level of the operation task, for example: increasing the success rate weight for high-risk tasks, and strengthening the timeliness and similarity weight for new device operations. Finally, the head sample ticket is selected as the optimal operation ticket in descending order of comprehensive score, which is used to drive the subsequent generation process, achieving a balance between timeliness, reliability and semantic relevance of the retrieval result.

[0057] The above scheme realizes effective balance between semantic matching accuracy and practical reliability of sample ticket retrieval by adopting a multi-level retrieval mechanism of vectorized similarity screening-task attribute filtering-multi-index weighted sorting, especially introducing dynamic business indicators such as historical execution times and execution success rate to participate in weight calculation, and improves the accuracy of selecting the optimal operation ticket. On the other hand, by taking business attributes such as device type, voltage level and risk level as filtering conditions, and combining dynamic indicators such as execution success rate and timeliness for weighted sorting, the sample ticket retrieval process deeply integrates business scenario constraints and operation experience feedback, and enhances the business adaptability and safety compliance of the retrieval result to the specific dispatching instruction.

[0058] Optionally, the above power model-based switching operation ticket generation method can further include: obtaining an execution feedback result of the optimal operation ticket; if the execution feedback result is failure, pushing an operation ticket reselection prompt to the user, and obtaining the optimal operation ticket reselected by the user from the target sample ticket set to continue the switching operation ticket generation process; if the execution feedback result is success, increasing the weight of the historical execution times and the execution success rate in the weighted sorting.

[0059] When it is determined that the execution result of the optimal operation ticket is failure, an artificial reselection mechanism can be automatically activated. The mechanism visualizes all candidate operation tickets in the target sample ticket set and their metadata such as matching degree and historical execution confidence through an interactive interface, and manually specifies a substitute sample ticket as a new reference template by the user based on field experience. After capturing the artificial reselection instruction, the initial conditions of the switching operation ticket generation process are automatically reset, the reselected sample ticket data is injected into the inference context of the large model, and the operation sequence is regenerated and error prevention verification is triggered, without repeating the pre-implementation of the dispatching instruction analysis and device information retrieval. This reselection mechanism effectively makes up for the semantic misjudgment defects of pure algorithm retrieval in boundary scenarios by embedding the judgment ability of dispatching experts into the exception handling branch of the automated process, and simultaneously records the reselection operation log and the final execution result as incremental training data for subsequent model iteration optimization, forming a closed-loop feedback loop from error detection to manual intervention and then to model improvement.

[0060] When the execution feedback result indicates that the switching operation order passes the anti-misoperation check and simulation pre-rehearsal and is highly matched with the logic structure of the typical sample order, a positive weight adjustment mechanism can be triggered to increase the weight coefficients of the historical execution times and the execution success rate of the typical order in the weighted ranking function, so that the sample order has a higher priority and confidence in the subsequent retrieval of similar dispatching instructions, thereby strengthening the path dependence of successful experience. On the contrary, if the execution feedback is a failure or there is a key step deviation, a negative feedback or elimination mechanism is started to reduce the execution success rate weight of the sample order, or mark it as a disabled state and remove it from the candidate set, to avoid repeating the reference of the error mode. This dynamic weight adjustment strategy relies on the real-time feedback and statistical update of the execution result, forming a closed-loop control loop of retrieval-generation-execution-feedback-weight optimization, so that the quality evaluation and ranking strategy of the sample order knowledge base can learn online and adapt to the changes of the power grid operation mode, continuously improving the accuracy and business adaptability of the optimal sample order selection.

[0061] The above scheme obtains the execution feedback result of the optimal operation order and establishes a failure reselection mechanism, constructs a continuous learning closed loop of human-computer cooperation, so that the above switching operation order generation method can feed the execution failure cases as negative samples to the retrieval link, correct the retrieval deviation through manual intervention, and improve the adaptability and fault tolerance to complex boundary scenarios. On the other hand, by dynamically increasing the weight coefficients of the historical execution times and the execution success rate in the weighted ranking when the execution is successful, the sample order quality online evaluation and ranking strategy adaptive optimization based on the actual operation effect are realized, so that the high-quality sample order is preferentially selected in the subsequent retrieval, thereby continuously improving the overall success rate and reliability of the switching operation order generation.

[0062] Optionally, the above switching operation order generation method based on the power large model can further include: obtaining the execution feedback result of the optimal operation order; if the execution feedback result is successful, storing the switching operation order and the corresponding dispatching instruction to the sample order knowledge base; if the execution feedback result is a failure, recording the error information of the current failure operation; updating the parameters of the large model based on the new data and / or error information in the sample order knowledge base.

[0063] The updating strategy of the above sample ticket knowledge base relies on the execution feedback closed loop mechanism, and realizes the dynamic evolution of the knowledge system through the double-channel data backflow. For the successfully executed operation ticket, it is automatically encapsulated as a high-quality positive sample with the corresponding scheduling instruction, execution timestamp and success mark, and after de-duplication verification and format normalization, it is injected into the sample ticket knowledge base, and the execution success rate and historical execution times of the sample are updated, and the weights of the two parameters are updated, so as to continuously expand the typical operation mode library. For the failed case, the detailed error log, failure reason classification and verification conflict point are captured, and a structured negative sample record is constructed and associated with the original scheduling instruction to form a negative sample knowledge base to identify high-risk operation modes. The bidirectional accumulation of positive and negative samples makes the sample ticket knowledge base not only have the reference function of static templates, but also become a self-updating knowledge carrier carrying operation experience and error lessons, and the data distribution adapts to the change of the power grid operation mode.

[0064] The above large model parameter updating strategy is based on the incremental learning paradigm, and uses the newly added positive and negative samples in the sample ticket knowledge base to construct a dynamic training data set, and performs lightweight fine-tuning or low-rank adaptive optimization on the model. The newly added successful sample ticket is used to enhance the probability distribution of the model in similar operation tasks, and the memory and reproduction ability of high-confidence operation modes are strengthened through supervised fine-tuning; and the failure case and error information are used as difficult negative samples, and a contrast learning or ranking loss function is introduced in the training process to improve the discrimination ability and risk avoidance awareness of the model to illegal operation sequences. This updating strategy can reduce the high cost of full-retraining, realize local optimization and knowledge updating of the model parameter space through continuous injection of incremental data, and ensure that the model performance is improved synchronously with the expansion of the power grid topology, the commissioning of equipment and the evolution of scheduling rules.

[0065] The above scheme realizes the dynamic evolution of the knowledge base based on the actual operation effect by establishing a bidirectional closed loop mechanism of execution feedback and knowledge base accumulation, automatically deposits the successfully executed operation ticket into the sample ticket knowledge base to continuously expand the sample size, and records the error information of the failed case to form a negative sample library, improves the self-improvement ability of the above switching operation ticket generation method to the power grid operation scene; on the other hand, through the continuous parameter updating and incremental training of the large model based on the newly added sample ticket data and error information, the model can quickly adapt to new scenes such as the change of power grid topology, the commissioning of equipment and the evolution of scheduling rules, avoiding the low efficiency problem of traditional offline full-retraining, effectively enhancing the timeliness and generalization ability of the model, and reducing the long-term maintenance cost.

[0066] Step S140: inputting the operation elements, real-time state information and sample ticket data into the large model to generate the switching operation ticket.

[0067] Optionally, step S140 may include: decomposing the switching operation ticket generation task into multiple subtasks; wherein each subtask is used to generate an operation sequence for a single device; calling the subtask agents deployed on multiple computing nodes to execute the subtasks and generate corresponding sub-operation sequences; and integrating each sub-operation sequence into a complete switching operation ticket according to the device operation order.

[0068] Understandably, decomposing the switching operation ticket generation task into multiple subtasks can circumvent the token length limitation of large models and improve the computational efficiency of generating long sequences. A complete switching operation ticket typically involves dozens of equipment operation steps. Generating the entire sequence at once can easily exceed the model context window, leading to information truncation and loss of logical coherence. Furthermore, a single equipment operation sequence has relatively independent logical boundaries. Decoupling it into atomic-level subtasks enables parallel processing of subtask agents on multiple computing nodes, reducing overall generation latency and enhancing throughput. When decomposing the switching operation ticket generation task, the large model can perform topology analysis on operation elements and equipment information. Based on safety constraints in power operation procedures (such as disconnecting circuit breakers before opening isolating switches and verifying voltage before grounding), the model performs topological sorting of all equipment involved in the operation, generating an ordered list of equipment operations. Subsequently, the model maps each device in the list to an independent subtask. The subtask definition includes equipment identifier, operation type, initial and target states, and associated topology predecessor constraints.

[0069] An operation sequence refers to a set of atomic, time-dependent operation steps defined for a specific electrical device in a switching operation ticket. It is the basic logical unit constituting a complete operation ticket. The operation sequence strictly adheres to the five-prevention constraints and equipment state transition rules of the power operation regulations, typically consisting of three functional segments: pre-operation checks (such as confirming the current device location, verifying the status of relevant circuits, and judging safety conditions), specific operation instructions (such as standardized verb-object structures like "disconnect XX circuit breaker" and "open XX disconnector"), and post-operation confirmations (such as resetting the device location remote signal, receiving SOE events for state changes, and verifying mechanical indications). Each operation sequence is device instance-specific; its content parameters (device name, number, status) are uniquely determined by the target device information in the operation elements, while the operation logic template originates from standardized step patterns matched in the sample ticket knowledge base. By decomposing complex operation tickets into device-level operation sequences, task parallelization and token length optimization are achieved, while ensuring clear pre- and post-departure dependencies between operation steps and complete safety verification points. This maintains the coherence and compliance of the global operation logic based on step-by-step generation.

[0070] The sub-task agent is a lightweight task execution entity deployed on a distributed computing node, serving as a parallel processing unit for the generation of switching operation tickets. Its function is to independently complete the operation sequence generation sub-task of a single device. Each sub-task agent receives a sub-task description through a standardized API interface, which includes target device identification, operation type, initial and target state, and other local operation elements. The agent integrates a large model calling module and a sample ticket knowledge base retrieval module. During sub-task execution, the agent first retrieves standardized operation knowledge fragments (covering pre-operation check items and post-operation confirmation items) matching the device from the sample ticket knowledge base, and then inputs the retrieval results and device parameters of the current sub-task into the large model to generate individualized operation sub-sequences for the device. The sub-task agent is deployed in multiple nodes in parallel to improve the task-level parallelism of operation ticket generation and avoid Token length restrictions. The output results are integrated in the order defined by the master node to form a complete operation ticket.

[0071] The above scheme breaks through the Token length limitation of a single large model when processing long sequence operation tickets by decomposing the switching operation ticket generation task into sub-tasks for a single device and calling sub-task agents deployed on multiple computing nodes to execute in parallel, improving the feasibility and response speed of large-scale switching operation ticket generation. The modular task decomposition mechanism allows each sub-task agent to run independently without interference, enhancing the scalability and fault tolerance of the above switching operation ticket generation method. Meanwhile, by integrating sub-operation sequences in the order of device operation, the logical coherence and execution timing accuracy of the final complete operation ticket are ensured.

[0072] Optionally, the sub-task agent executes the sub-task to generate a sub-operation sequence, including: retrieving a sample ticket knowledge fragment matching the corresponding device in the sample ticket knowledge base; wherein the sample ticket knowledge fragment includes pre-operation check items and post-operation confirmation items; and replacing device information in the sample ticket knowledge fragment with actual device information in the operation elements to generate a sub-operation sequence.

[0073] The sample ticket knowledge fragment is a standardized operation logic template for a single electrical device stored in the sample ticket knowledge base. Its content is indexed by device type and operation type, and includes general pre-operation check items and post-operation confirmation items for the device in a specific state transition scenario. Pre-operation check items usually include remote verification of the current position state of the device, logical judgment of the charged state of the closed loop, availability check of the control power supply and energy storage mechanism, and other safety prerequisites; post-operation confirmation items include device position displacement SOE event reception confirmation, mechanical indicator state check, related protection signal reset, and operation timestamp recording, and other integrity verification steps.

[0074] The sub-task agent can construct a composite query key value according to the device type code (such as a circuit breaker, a disconnector) and the operation type identifier (such as opening, closing, switching) specified in the current sub-task, and perform retrieval in the sample ticket knowledge base. The retrieval process can first perform strict equality matching, and if no hit is found, the generalization rule is enabled, such as falling back to a general circuit breaker template when a specific type of circuit breaker is not matched, to ensure a balance between accuracy and coverage. After a successful match, the complete content of the knowledge fragment is returned, including the conditional judgment logic, the verification item list, and the state confirmation steps, providing a reusable logic skeleton for subsequent device information replacement and individualization generation.

[0075] The above-mentioned sub-task agent parses the template structure of the sample ticket knowledge fragment, identifies the reserved device-related placeholder fields, including the device standardized naming slot, the device number variable, the voltage level parameter, and the state description macro. Then, the actual device information extracted from the operation elements, such as the specific device name, the station unique number, the target voltage level, and the initial / target state value, is bound to the corresponding actual value through key-value mapping. The replacement process supports two modes of string template replacement and abstract syntax tree node rewriting, where string template replacement is suitable for operation instructions in pure text format, and regular expressions or template engines are used for variable interpolation; abstract syntax tree node rewriting is suitable for structured sample ticket fragments, and device attribute nodes are located and replaced by traversing the abstract syntax tree. The final generated sub-operation sequence completely retains the operation logic framework of the sample ticket knowledge fragment, and all device information has been instantiated as the target object of this operation, realizing accurate adaptation from a general template to a specific device operation.

[0076] The above-mentioned scheme retrieves the sample ticket knowledge fragment matching the corresponding device in the sample ticket knowledge base, directly reuses the standardized operation logic containing pre-operation check items and post-operation confirmation items, avoids the uncertainty caused by the complete reliance of the sub-task agent on model generation, and improves the standardization and accuracy of sub-operation sequence generation; on the other hand, the device information replacement mechanism replaces the device information in the sample ticket knowledge fragment with the actual device information in the operation elements, realizes accurate adaptation from a typical template to a specific device instance, maintains the consistency of operation logic, ensures the accurate matching of generated content and target devices, and enhances the applicability and reliability of the operation ticket.

[0077] It can be understood that in addition to the distributed parallel architecture, a monolithic loop execution mode can also be used to implement subtask processing. In this mode, only a single subtask agent instance can be deployed to run on an independent computing node, and the master node can sequentially deliver the decomposed subtasks to the agent entrance according to the device operation order, forming a serial processing pipeline. The subtask agent traverses the device list through a loop control structure, calls the large model to generate the operation subsequence of the current device in each iteration, and caches the result to a temporary queue.

[0078] Optionally, the power-based large model switching operation ticket generation method can further include: performing an error prevention check on the switching operation ticket; wherein the error prevention check includes at least one of device name verification, operation range verification, pre-operation device state verification, post-operation device state verification, missing item verification, operation terminology verification, and incorrect operation verification; and / or performing an error prevention simulation rehearsal on the switching operation ticket; and / or pushing an artificial audit prompt to the user and obtaining the artificial audit result of the user.

[0079] The above anti-misoperation check is the core safety verification link in the generation process of the switching operation ticket, which performs multi-dimensional logical review on the generated operation ticket by calling the anti-misoperation system rule set and the device knowledge base, to ensure that the operation sequence meets the safety, compliance and integrity constraints of the power system operation. The anti-misoperation check system includes seven key checks: device name check, operation range check, pre-operation device state check, post-operation device state check, missing item check, operation terminology check, and misoperation check. Each check item cooperates to form a defense-in-depth mechanism. The device name check verifies whether the device name and number referred to in the operation ticket are consistent with the standard naming in the device knowledge base through string exact matching and fuzzy matching algorithm, to prevent misoperation object risks caused by alias, abbreviation or spelling errors. The operation range check relies on the dispatching instruction authorization boundary and the device topology relationship library, and uses a graph traversal algorithm to check whether the operation object exceeds the plant, voltage level and interval range covered by the current dispatching instruction, to intercept the illegal operation of unauthorized or misselected devices. The device state check performs double verification of the device state before and after operation: the pre-operation device state check compares the initial state field of the operation ticket with the state telesignal value synchronized in real time from the anti-misoperation system in the device knowledge base, to ensure that the pre-operation condition is consistent with the actual state on site, and to prevent misoperation in the device under voltage or abnormal state; the post-operation device state check simulates the operation execution result based on the state transition rule engine, to verify whether the predicted target state meets the requirements of the dispatching instruction and the device technical specification. The missing item check can use a directed acyclic graph topology sorting algorithm to detect the integrity of the key path of the operation step set in the operation ticket, to identify and alarm the missing safety steps such as voltage test, grounding, interlocking unlocking, etc. The operation terminology check can perform part-of-speech tagging and similarity matching based on the power standard terminology library, to verify whether the verbs and state description words in the operation ticket text meet the industry specifications, and to eliminate the execution misunderstanding caused by term ambiguity or unconventional expression. The misoperation check can call the built-in five-prevention rule set and production logic rule of the anti-misoperation system, to check whether the operation sequence triggers logical conflicts such as live ground knife, live load pulling of disconnecting switch, etc., and outputs the risk level and alarm information through the conflict detection algorithm of the rule engine. The above various check items can be executed in parallel or in series, and any link that does not pass triggers the operation ticket rollback process, and the generated check report records the violation type, conflict position and suggested correction measures in detail, providing decision basis for subsequent manual review or automatic optimization.

[0080] The above anti-misoperation simulation rehearsal can be a deep verification link triggered by the system after the anti-misoperation check passes, aiming to perform full-sequence timing simulation on the switching operation order in a virtual electrical topology environment by calling the digital twin simulation engine of the anti-misoperation system, to verify its feasibility and safety under real physical constraints. When performing the anti-misoperation simulation rehearsal, the final operation order text can be first serialized into a standard control command timing stream, which is injected into the virtual execution sandbox of the anti-misoperation system through a standard API interface. The sandbox constructs a mirror of the initial state of the equipment consistent with the field based on the real-time section data of the substation, and performs state transition simulation according to the steps of the operation order. Each step triggers dynamic logical judgment of the anti-misoperation rule set, monitors whether there are hidden risks such as device state conflict (such as attempting to open a circuit breaker that is already in an open state), operation sequence violation (such as operating the associated disconnecting switch without disconnecting the circuit breaker), electrical interlocking condition damage (such as mistakenly closing the grounding knife switch when the bus is live), or insufficient safety distance. If any step in the simulation rehearsal process triggers anti-misoperation logic locking or state abnormality, the execution is immediately terminated and precise error positioning information (including the violating device ID, the conflicting rule number, and the recommended correction strategy) is returned, allowing the user to directly modify the dispatching instructions or optimize the prompt words based on this feedback and then trigger the intelligent order generation process again. If the full-sequence simulation execution is successful and there is no logical conflict, a rehearsal pass flag is output, and the operation order enters the subsequent manual review stage. This mechanism realizes high-fidelity verification of the logical correctness of the operation order through deep coupling of virtual execution and rule checking, reducing the risk of field misoperation caused by model reasoning errors or knowledge base lag.

[0081] The aforementioned manual review and prompting mechanism serves as the final safety confirmation step. Through a human-computer interaction interface, the generated switching operation ticket is pushed to qualified professional reviewers, and their manual review results are collected to complete the closed-loop decision-making process. After the error prevention verification and simulation pre-run pass, the review prompting process can be automatically triggered. Based on the token-level confidence level output by the model or the fuzzy matching results of the rule engine, questionable content in the operation ticket is identified and highlighted, such as low-confidence equipment names, novel operation steps not covered in the sample ticket knowledge base, or conditional branches with multiple possibilities in the state transition logic. An review request message containing uncertainty markers and risk warnings is generated. This message is pushed to the user terminal through the visualization interface of the scheduling monitoring system or a dedicated review client, presenting the generated operation ticket and recommended typical sample tickets in a column comparison format, along with real-time equipment status snapshots, topology diagrams, and proof of successful error prevention verification as auxiliary decision-making basis. After receiving the prompt, the reviewers can check the logical coherence, terminology standardization, and safety compliance of the operation ticket item by item, outputting the manual review results through three interaction modes: click confirmation, annotation modification, or rejection and regeneration. If the review is approved, the operation ticket will be officially marked as an executable version and written into the sample ticket knowledge base; if the review is not approved, the modification opinions will be associated with the original scheduling instructions and fed back to the large model to trigger the regeneration process. At the same time, the correction records during the review process will be stored as supervision signals in the training dataset for subsequent parameter updates and knowledge base optimization of the large model.

[0082] The above solution achieves comprehensive cross-verification of generated operation tickets by configuring a multi-dimensional AI anti-error verification mechanism that covers device name consistency, compliance of operation scope, logical integrity of state before and after operation, standardization of terminology, and detection of anti-error logic conflicts. This effectively intercepts complex logical errors that cannot be identified by single rule verification, improving the detection rate and accuracy of automatic verification. On the other hand, by introducing a dual-verification mechanism of anti-error simulation pre-run and manual review prompts, a defense-in-depth system of intelligent verification, virtual pre-run, and manual confirmation is constructed. This ensures that operation tickets undergo multiple verifications before entering the execution stage, minimizing security risks caused by model illusions or incomplete rule coverage, and ensuring the security and reliability of operation tickets in real-world scenarios.

[0083] To facilitate understanding of the working principle of the above-described method for generating switching operation tickets, a specific application example in a particular scenario is provided below. In this scenario, the large-scale model uses the DeepSeek model, and the above-described method for generating switching operation tickets mainly includes: Phase 1: Pre-training the large model; Please see Figure 2The pre-training step of the large model mainly includes: (1) collecting multi-source heterogeneous power operation data as a pre-training corpus, which includes historical operation tickets, typical operation tickets, power operation procedures, safety procedures, ticket building rules and professional term library. (2) Perform comprehensive data cleaning operation on the original data, remove redundant records, noise interference items and field missing samples, then split the long text operation ticket into atomic operation steps through semantic segmentation algorithm, and manually comb and label the difficult samples, supplement the missing context constraints and ambiguity resolution marks, and finally build structured training materials with good structure. (3) Based on the cleaned data, construct a multi-task mixed training data set for the power vertical field, the core components of which include question and answer pairs composed of dispatching instructions and operation ticket texts, operation reasoning procedure with explicitly annotated operation premise and state transition logic, operation step dependency relationship encoded by directed acyclic graph, and power-specific knowledge fields such as device state enumeration value and topology association index stored in key-value structure. Each sample follows a unified specification and contains a complete input-reasoning-target label chain.

[0084] At the model architecture level, DeepSeek power large model is used as the basic base, and its Mixture-of-Experts (MoE) sparse activation architecture is used for vertical pre-training in the power field. During training, the MoE layer dynamically routes the input sample to the expert subnetwork highly related to the power operation logic through the gating network, achieving the coordinated optimization of high capacity and low power consumption. The model integrates Multi-Head Latent Attention (MLA) mechanism to reduce the key-value cache overhead of long operation sequence, and introduces Multi-Token Prediction (MTP) strategy to generate multiple subsequent operation items simultaneously, improving the training signal density and generation efficiency. Training uses FP8 mixed precision calculation framework to accelerate gradient update and reduce memory occupancy while maintaining numerical stability. By inputting the above training data set into the DeepSeek model for deep supervision learning and instruction compliance fine-tuning, the model parameter space deeply integrates the distributed representation of power dispatching terminology, device state transition probability distribution and five-prevention rule constraint mode, and finally obtains a vertical domain large model with the ability of accurately understanding dispatching instructions, generating compliant operation sequences and reasoning complex operation logic, providing a core generation engine for subsequent intelligent ticket building.

[0085] Second stage: knowledge base initialization; The knowledge base initialization phase relies on the Dify workflow management platform to establish and optimize the device knowledge base and sample ticket knowledge base, and to build the data foundation for intelligent ticket building. Please refer to Figure 3, the device knowledge base construction process mainly includes: (1) receiving the original wiring diagram file of the substation, (2) triggering the power large model driven image analysis pipeline through the Dify platform: the power large model calls the vector conversion tool to translate the wiring diagram into SVG format, uses the path element recognition and graph semantic reasoning technology to extract the device name, number, type (circuit breaker, disconnector, grounding switch, etc.) and electrical connection topology relationship (node connection sequence, bus-bar-section attribution) from the SVG file, and encodes the extraction results into a standardized JSON structured description file. The JSON data can be written in layers into the device knowledge base, and stored in the primary device model library (main device attributes), the secondary device model library (protection and control device logic) and the topology relationship library (device connection matrix) respectively, forming a three-dimensional knowledge base with graph structure expression capability. To ensure the dynamic nature of the knowledge, a real-time data channel can be established with the anti-misoperation system through the RESTful API interface, and the device position remote signal, state value and logic node data can be pulled or subscribed regularly, and an event-driven monitoring agent can be deployed, which automatically triggers API calls to incrementally update the corresponding state fields in the device knowledge base when receiving the position change event or fault alarm pushed by the anti-misoperation system, to realize millisecond-level synchronization of device information and operating conditions.

[0086] See Figure 4 , the main steps of the power large model calling the vector conversion tool to translate the wiring diagram into SVG format include: (1) establishing a metadata list: extracting basic information such as file attributes, layer structure and graph object quantity, etc. to provide an index framework for subsequent analysis. (2) structure recognition: the tool identifies the geometric graph elements (such as rectangle, circle, line, polygon) in the wiring diagram and their combination modes through graph algorithms, locates the symbolized graph elements representing power devices such as circuit breakers, disconnectors, transformers and bus bars, and analyzes the spatial connection relationship between the graph elements, to preliminarily construct the geometric topology network of the device nodes. (3) labeling semantic labels: using the pre-trained visual-linguistic alignment capability in the power field, performing semantic label labeling on the identified graph elements, and assigning a standardized device type code (such as CBR for circuit breaker, DIS for disconnector) to each device symbol. (4) labeling function, logic path and state: labeling the function attributes (such as main incoming line, tie-in, feeder), logic path (such as current flow direction, control loop connection) and operating state (such as closed, open, grounded) of the devices. After completing the labeling, a structured data object is established for each device element, encapsulating its geometric coordinates, size parameters, associated terminal list and topology connection index, forming a hierarchical device model tree. (5) output XML data: the vector conversion tool serializes all the labeled graph elements and metadata into a standardized XML format file.

[0087] See Figure 5The sample ticket knowledge base initialization process mainly includes: extracting the audited and archived typical switching operation tickets from the power grid dispatching control system or production management system through an ETL tool, covering different voltage levels, device types and operation task types, each record containing a standardized operation step sequence, device state transition path, operation timestamp and associated device ID, and after vectorization coding, storing in a vector database to form a basic template library. The sample ticket knowledge base has a dynamic update mechanism. After the intelligent ticket generation process is completed, the operation tickets that have been successfully executed and passed the verification, as well as the optimized sequences corrected by manual audit, are added to the sample ticket knowledge base through an incremental writing interface, and the metadata such as the execution frequency and success rate of the ticket sample are updated, realizing self-evaluation of the quality of the ticket sample based on the running effect and reordering of the priority.

[0088] The above-mentioned device knowledge base and sample ticket knowledge base can be associated and mapped in both directions through device topology indexing and operation object coding. Each operation step in the ticket sample is bound to a unique device identifier and real-time state constraint in the device knowledge base, forming a knowledge system that is integrated, correlated and dynamically updated among devices, topology and ticket samples, providing structured, searchable and high-confidence data support for subsequent intelligent ticket generation.

[0089] Third stage: intelligent ticket generation; The intelligent ticket generation stage is the core function, which generates high-quality switching operation tickets through the collaborative work of large model agents, knowledge bases and anti-misoperation checks. Please refer to Figure 6 , which mainly includes: (1) Dispatching instruction input: the user inputs the power dispatching instruction (for example, "the XX main transformer of a certain power plant needs to be switched from the running state to the maintenance state"), which is then transmitted to the large model after being received.

[0090] (2) Feature value extraction: the large model analyzes the dispatching instruction and extracts key feature values, including but not limited to the ordered power plant, main device, voltage level, operation device, operation type, device type, initial state and target state of operation.

[0091] (3) Device information retrieval: according to the operation device and the main device, the detailed information of the current device is retrieved from the device knowledge base, such as device name, number, real-time state, topology relationship, etc.

[0092] (4) Sample ticket knowledge retrieval: retrieve similar typical operation ticket data from the sample ticket knowledge base as a reference for generating. The main steps include: (a) input the order text into the special embedding model for Embedding vectorization processing to generate a 1024-dimensional vector; (b) calculate the cosine similarity between the order vector and all typical operation ticket vectors in the knowledge base, and find the 5 closest typical operation tickets. (c) According to the extracted feature values, filter the 5 typical operation tickets based on device type, voltage level, risk level, etc., and then weight average according to ticket update time, historical execution times, execution success rate, matching degree, etc. After sorting, the optimal typical operation ticket is selected. If the automatically selected typical operation ticket has an error in subsequent operation, a human way can be used to reselect the optimal typical operation ticket to continue the following operation. If the operation ticket is executed successfully and highly matched with the typical ticket, the weight of the historical execution times and the weight of the execution success rate of the typical ticket can be increased; if it fails or deviates, it can be marked down or discarded.

[0093] (5) Operation sequence generation: input the extracted feature values, device information and sample ticket knowledge into the power large model, and the model generates a preliminary switching operation ticket sequence through deep learning and reasoning. Here, the segmented output method can be used to avoid the token limit of the power large model. The main steps include: (a) retrieve the associated devices in the device knowledge base through the extracted device information, such as all devices in the same interval, and then sort the devices according to the operation procedures, such as disconnecting the circuit breaker first, then disconnecting the isolating switch, opening the cabinet door operation, which needs to be pre-electricity, no electricity, then execute the next operation, and then hang the ground wire or ground knife. (b) Split the text to avoid exceeding the token limit of the power large model due to long context, in the sorted device list, a sub-task agent can be called for each device to generate the operation sequence of the corresponding operation device, and the MapReduce algorithm can be used to improve the generation efficiency. (c) The sub-task agent matches the sample ticket knowledge segment to find the pre-operation check items and post-operation confirmation items required for the operation, and uses the learning and reasoning ability of the power large model to replace the devices in the sample ticket with the actual operation devices. This result is a device operation sub-sequence. This operation ensures that each input + output is within the token limit of the model, ensuring stable operation of the model. Finally, the results are spliced according to the device list order, which is the complete operation sequence corresponding to this operation ticket.

[0094] (6) AI Anti-misoperation Verification: The generated preliminary operation ticket is subjected to multi-step AI anti-misoperation verification to ensure the safety and accuracy of the operation ticket. The verification includes the following 7 key steps, and each step is considered successful after passing: (a) Device Name Verification: Verify whether the device name involved in the operation ticket is consistent with the record in the device knowledge base to prevent name errors or confusion. (b) Operation Range Verification: Confirm whether the operation range meets the dispatch instructions and operation procedures to avoid unauthorized or incorrect operations. (c) Pre-operation Device State Verification: Check whether the device state before operation is consistent with the real-time state in the device knowledge base to ensure that the preconditions are met. (d) Post-operation Device State Verification: Predict whether the device state after operation meets the target state and compare it with the target state of the dispatch instructions. (e) Missing Item Check: Check whether the operation ticket misses key steps or device state records. (f) Operation Term Check: Verify whether the terms in the operation ticket conform to the power industry standards and operation procedures to avoid ambiguity or non-standard expressions. (g) Anti-misoperation Verification: Combine the rule set of the anti-misoperation system to check whether the operation ticket may cause misoperation risks.

[0095] (7) Anti-misoperation Simulation Pre-performance: If the AI anti-misoperation verification is successful, the simulation pre-performance function of the anti-misoperation system is called to execute the operation ticket in a virtual environment to verify the feasibility and safety of the operation. If the simulation pre-performance fails, specific error information (such as device state conflict, operation sequence error, etc.) is prompted, and the user is allowed to modify the dispatch instructions or optimize the prompt words to restart the intelligent ticket generation process.

[0096] (8) Artificial Audit Stage: If the simulation pre-performance is successful, the generated operation ticket enters the artificial audit stage. Uncertain words, device states, etc. are highlighted, and artificial confirmation is provided to guide the auditors to make the final confirmation on the logic, compliance, and safety of the operation ticket. If the audit is passed, the operation ticket is officially generated and written into the sample ticket knowledge base for subsequent continuous learning and optimization; if the audit is not passed, feedback information is recorded and returned to the intelligent ticket generation stage for correction.

[0097] In the above application scenarios, the switching operation ticket generation method can also achieve continuous learning through dynamic updating of the knowledge base and optimization of the power large model: (1) Knowledge Accumulation: Each successfully generated and audited operation ticket is recorded in the sample ticket knowledge base to form historical data accumulation. (2) Model Iteration: Based on the newly accumulated data, the power large model is periodically incrementally trained to improve its accuracy and adaptability in generating and verifying power operation tickets. (3) Abnormality Handling: For cases where the verification fails or the simulation pre-performance fails, the error reasons and user feedback for modification are recorded as part of the training data to improve the model and the coverage of the knowledge base.

[0098] See Figure 7, based on the same inventive concept, the embodiment of the present application also provides a switching operation order generation system based on a power large model. The system adopts a hierarchical architecture design and is composed of an application layer and a platform layer. The application layer deploys an intelligent order generation module, a misoperation prevention system interface, and a monitoring system interface. The intelligent order generation module serves as the core interaction entrance and is responsible for receiving user input scheduling instructions and driving the whole-process order generation. The misoperation prevention system interface provides real-time device state data subscription and simulation rehearsal calling capabilities. The monitoring system interface is used to obtain auxiliary decision-making information of the substation operating condition. The platform layer carries the intelligent decision-making and knowledge management functions of the system, is built based on a Dify low-code workflow platform, and internally has a dispatching instruction analysis intelligent agent and a misoperation prevention checking intelligent agent, which respectively realize scheduling instruction semantic analysis and multi-dimensional safety checking of the operation order. The knowledge base system is composed of a sample order knowledge base and a device knowledge base. The former stores typical operation order templates and dynamically updated execution feedback data, and the latter solidifies the device topology relationship generated by the large model from the electrical wiring diagram and the real-time state data synchronized with the misoperation prevention system.

[0099] The model layer of the system adopts a hierarchical coupling architecture of a vertical power large model and a base power large model. The base model selects DeepSeek, which provides high-capacity semantic representation capabilities relying on its mixed expert (MoE) architecture and multi-head latent attention (MLA) mechanism. The vertical power large model is specialized in scheduling instruction understanding, operation logic reasoning, and order text generation by pre-training and incremental fine-tuning on the DeepSeek base through power corpus. During system operation, the dispatching instruction analysis intelligent agent calls the power large model to extract structured operation elements, the device knowledge base retrieval module queries topology and state information based on the device identifier in the elements, the sample order knowledge base retrieval module obtains the optimal reference template through vector similarity calculation and business attribute filtering, the generation stage generates the operation sequence by the power large model combined with the RAG mechanism, and the misoperation prevention checking intelligent agent performs name consistency, state logic, term specification, and misoperation rule conflict detection. After verification, it calls the misoperation prevention system interface to complete the simulation rehearsal, and finally forms an executable operation order after manual review and confirmation, and feeds back to the sample order knowledge base to realize continuous learning and closed loop.

[0100] It can be understood that the switching operation order generation system based on the power large model provided by the embodiment of the present application can be used to execute the switching operation order generation method based on the power large model provided by the embodiment of the present application. The implementation principle and the technical effects generated have been introduced in the foregoing method embodiment. For brief description, the part of the system embodiment not mentioned can be referred to the corresponding content in any one of the foregoing method embodiments.

[0101] Figure 8 A schematic diagram of an electronic device provided by the embodiment of the present application is shown in FIG. 1. Referring to FIG. 1, Figure 8The electronic device 300 includes a processor 310, a memory 320, and a communication interface 330, which are interconnected and communicate with each other through a communication bus 340 and / or other forms of connection mechanism (not shown). Among them, the memory 320 includes one or more (only one is shown in the figure), which can be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The processor 310 and other possible components can access, read and / or write data in the memory 320. The processor 310 includes one or more (only one is shown in the figure), which can be an integrated circuit chip with signal processing capability. The above-mentioned processor 310 can be a general-purpose processor, including a central processing unit (CPU), a micro controller unit (MCU), a network processor (NP) or other conventional processors; it can also be a special-purpose processor, including a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The communication interface 330 includes one or more (only one is shown in the figure), which can be used for direct or indirect communication with other devices to interact with data. For example, the communication interface 330 can be an Ethernet interface; it can be a mobile communication network interface, such as a 3G, 4G, 5G network interface; or it can be another type of interface with data transceiver function. One or more computer program instructions can be stored in the memory 320, and the processor 310 can read and run these computer program instructions to implement the power-based large model-based switching operation ticket generation method provided by the embodiments of the present application and other desired functions.

[0102] It can be understood that,Figure 8 The illustrated structure is merely schematic, and the electronic device 300 can include more or fewer components than shown in FIG. 1, or have a different configuration from that shown. Figure 8 The components shown in FIG. 1 can be implemented in hardware, software, or a combination thereof. For example, the electronic device 300 can be a single server (or other device having computational processing capabilities), a combination of multiple servers, a cluster of a large number of servers, etc., and can be a physical device or a virtual device. Figure 8 Figure 8 The components shown in FIG. 1 can be implemented in hardware, software, or a combination thereof. For example, the electronic device 300 can be a single server (or other device having computational processing capabilities), a combination of multiple servers, a cluster of a large number of servers, etc., and can be a physical device or a virtual device.

[0103] In the embodiments provided by the present application, it should be understood that the disclosed system and method can be implemented in other manners. The embodiments described above are merely schematic, and the division of the units is merely logical function division. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices, or units, and can be electrical, mechanical, or in other forms.

[0104] In addition, the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units. That is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present application.

[0105] In addition, the functional modules in the various embodiments of the present application can be integrated together to form a separate part, or each module can exist independently, or two or more modules can be integrated to form a separate part.

[0106] It should be noted that if the functions are realized in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such an understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art, or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and various other media capable of storing program codes.​

[0107] In this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.

[0108] The above description is merely illustrative of the application and not intended to limit the scope of the application. Various modifications and changes can be made by those of ordinary skill in the art without departing from the spirit and scope of the application. Any modification, equivalent replacement, improvement, and the like made within the spirit and principle of the application should be included in the scope of the application.

Claims

1. A method for generating a switching operation order based on a power large model, characterized in that, The method comprises: receiving a scheduling instruction and performing semantic analysis on the scheduling instruction by using a power large model pre-trained by a power corpus to extract structured operation elements; based on the operation elements, retrieving real-time state information of related equipment from an equipment knowledge base; wherein the equipment knowledge base stores equipment topology relationship data generated by analyzing an electrical wiring diagram by the power large model, and the equipment knowledge base is linked with a misoperation prevention system to synchronize equipment state data; based on the operation elements and the real-time state information, retrieving matching sample ticket data from a sample ticket knowledge base; wherein the sample ticket knowledge base stores typical operation ticket data, and the sample ticket knowledge base is configured to receive execution feedback results for dynamic updating; inputting the operation elements, the real-time state information and the sample ticket data into the power large model to generate a switching operation ticket. 2.The power model based switching operation ticket generation method of claim 1, wherein, The operation of inputting the operation elements, the real-time state information and the sample ticket data into the power large model to generate a switching operation ticket comprises: decomposing a switching operation ticket generation task into a plurality of subtasks; wherein each subtask is used to generate an operation sequence of a single device; respectively invoking subtask agents deployed on a plurality of computing nodes to execute the subtasks and generate corresponding sub-operation sequences; integrating each sub-operation sequence into a complete switching operation ticket according to the device operation sequence. 3.The power model based switching operation ticket generation method of claim 2, wherein, The operation of the subtask agent executing the subtask to generate the sub-operation sequence comprises: retrieving a sample ticket knowledge segment matching the corresponding device from the sample ticket knowledge base; wherein the sample ticket knowledge segment includes pre-operation inspection items and post-operation confirmation items; replacing the device information in the sample ticket knowledge segment with the actual device information in the operation elements to generate the sub-operation sequence. 4.The power model based switching operation ticket generation method of claim 1, wherein, The operation of retrieving matching sample ticket data from the sample ticket knowledge base based on the operation elements and the real-time state information comprises: performing vectorization processing on the operation elements to generate an instruction feature vector; calculating the similarity of the instruction feature vector with each typical operation ticket data in the sample ticket knowledge base, and obtaining a candidate sample ticket set based on the similarity; filtering the typical operation tickets in the candidate sample ticket set based on the task attributes in the operation elements to obtain a target sample ticket set; wherein the task attributes include device type, voltage level and risk level; weighting and sorting the typical operation tickets in the target sample ticket set based on update time, historical execution times, execution success rate and similarity with the instruction feature vector to obtain an optimal operation ticket.

5. The power model based switching order generation method of claim 4, wherein, The method further comprises: obtaining an execution feedback result of the optimal operation ticket; if the execution feedback result is failure, pushing an operation ticket reselection prompt to the user and obtaining the optimal operation ticket reselected by the user from the target sample ticket set to continue the switching operation ticket generation process; if the execution feedback result is success, increasing the weight of the historical execution times and the execution success rate in weighting sorting. 6.The power model based switching operation ticket generation method of claim 5, wherein, The method further comprises: obtaining an execution feedback result of the optimal operation ticket; If the execution feedback result is success, the switching operation order and the corresponding dispatching instruction are stored in the sample order knowledge base; If the execution feedback result is failure, error information of the failure operation is recorded; Based on the new data in the sample order knowledge base and / or the error information, the parameters of the large model are updated.

7. The power model based switching order generation method according to any one of claims 1 to 6, characterized in that, The method further comprises: constructing the device knowledge base, and / or, constructing the sample order knowledge base, wherein: The way of constructing the device knowledge base comprises: obtaining an electrical wiring diagram; using the power large model to analyze and reason the electrical wiring diagram, extracting device information and topological relationship; generating structured data and storing it to the device knowledge base; interacting with the anti-misoperation system to obtain device state information and store it to the device knowledge base; The way of constructing the sample order knowledge base comprises: obtaining the typical operation order data and storing it to the sample order knowledge base; based on the execution feedback result, dynamically updating the newly generated switching operation order to the sample order knowledge base; using the new data in the sample order knowledge base to iteratively train the large model. 8.The power model based switching operation ticket generation method according to any one of claims 1-6, wherein, The method further comprises: Anti-misoperation verification is performed on the switching operation order, wherein the anti-misoperation verification comprises at least one of device name verification, operation range verification, pre-operation device state verification, post-operation device state verification, omission item verification, operation terminology verification and misoperation verification; And / or, anti-misoperation simulation rehearsal is performed on the switching operation order; And / or, an artificial audit prompt is pushed to the user and the artificial audit result of the user is obtained.

9. The power model based switching order generation method according to any one of claims 1-6, wherein, The method further comprises: Obtaining historical data related to power operation; wherein the historical data comprises historical operation order, operation procedure, safety procedure and professional terminology library; Pretreating the historical data to construct a training data set containing operation logic, device state change and operation rules; Based on the training data set, the large model is subjected to domain enhancement training to obtain the trained large model.

10. An electronic device, comprising: Comprise: A processor, a memory and a communication bus, wherein the processor and the memory complete communication with each other through the communication bus; The memory stores program instructions executable by the processor, and the processor calling the program instructions can execute the method of any one of claims 1-9.

Images