Power industry large model optimization method and device

Abstract

The embodiment of the invention provides a power industry large model optimization method and device, and the method comprises the steps: obtaining set role information, input data, task information and output information, generating an input template according to the role information, the input data, the task information and the output information and a preset template rule, carrying out the analysis of the input template, obtaining an analysis result, and carrying out the optimization of a large model in the power industry. And the analysis result is converted into a feature vector, the feature vector is input into a pre-constructed electric power industry large model, and the electric power industry large model outputs an output result corresponding to the output information based on the feature vector. According to the invention, the accuracy and reliability of model processing power business can be improved, and multi-field power business scenes are supported.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and apparatus for optimizing large-scale power industry models. Background Technology

[0002] With the development of artificial intelligence technology, large-scale models have been widely applied in various industries. For example, in the power industry, large-scale models can be used to achieve functions such as fault diagnosis and resource scheduling, ensuring the stable operation of the power system and improving energy efficiency. However, because large-scale models do not deeply integrate power industry knowledge and lack a professional power perspective, they are prone to misjudgment and bias when processing power industry data, affecting the accuracy of business decisions. Furthermore, the lack of targeted prompt word templates causes the models to fail to accurately understand power industry data, resulting in output results that are disconnected from power business operations. Summary of the Invention

[0003] In view of this, the purpose of this application is to propose a method and apparatus for large-scale model optimization in the power industry.

[0004] Based on the above objectives, embodiments of this application provide a large-scale model optimization method for the power industry, including: Retrieve the set character information, input data, task information, and output information; Based on the character information, input data, task information, and output information, an input template is generated according to preset template rules; The input template is parsed to obtain the parsing result, and the parsing result is converted into a feature vector; The feature vector is input into a pre-constructed large-scale power industry model, and the large-scale power industry model outputs an output result corresponding to the output information based on the feature vector.

[0005] Optionally, before obtaining the set role information, input data, task information, and output information, the method further includes: Based on the role information, the corresponding professional knowledge base is invoked; Based on the aforementioned professional knowledge base, the pre-trained general large model is trained and adjusted to obtain the power industry large model.

[0006] Optionally, the template rules include role name tags, role domain tags, power parameter constraint tags, raw input data tags, input data type tags, output type tags, output keyword tags, and verification control tags; the role information includes role name and role domain, the task information includes power parameters and verification switch status, and the output information includes output type and output keywords; Based on the role information, input data, task information, and output information, an input template is generated according to preset template rules, including: Based on the character name and the character domain, set the character name tag and the character domain tag respectively; Based on the input data, set the original input data label and the input data type label; wherein, the original input data label is the input data, and the input data type label is the type of the input data; Based on the power parameters, set the power parameter constraint labels; Based on the output type and output keywords, set the output type label and output keyword label respectively; The verification control tag is set according to the verification switch status.

[0007] Optionally, the input template is parsed to obtain a parsing result, and the parsing result is converted into a feature vector, including: When the input data type label is long text, the input data is segmented using a sliding window to obtain multiple short text paragraphs; Key information is extracted from each short text using a pre-trained language model for electricity. By integrating the key information from each short text, a semantic graph of the long text is constructed. The long text semantic graph is converted into a corresponding feature vector.

[0008] Optionally, the verification switch includes a parameter verification switch; the method further includes: When the verification control tag is enabled for parameter verification, the output result is verified according to the preset parameter rules to obtain the verified output result.

[0009] Optionally, the parameter rules include parameter threshold rules and business logic rules; The step of verifying the output result according to preset parameter rules to obtain the verified output result includes: The output result is validated according to the parameter threshold rules and business logic rules. If the output result does not conform to the parameter threshold rules and / or business logic rules, the output result is corrected according to the parameter threshold rules and / or business logic rules.

[0010] Optionally, the output result may be corrected according to the parameter threshold rules and / or business logic rules, including: If the output results contain parameter threshold errors, business logic inconsistencies, or process specification errors; Errors related to parameter thresholds, business logic contradictions, and process specifications are corrected in sequence.

[0011] Optionally, the verification switch includes a security verification switch; the method further includes: When the verification control tag is enabled for security verification, the input data and output results are checked for sensitive information according to preset security rules. If the sensitive information exists, it is desensitized to obtain the input data and output results after security verification.

[0012] Optionally, the power industry large model outputs an output result corresponding to the output information based on the feature vector, including: According to the output type corresponding to the output type label, output standardized output results including output keywords corresponding to the output keyword label are output.

[0013] This application also provides a large-scale power industry model optimization device, including: The acquisition module is used to acquire the set role information, input data, task information, and output information; The generation module is used to generate an input template according to the role information, input data, task information, and output information, and in accordance with preset template rules. The parsing module is used to parse the input template, obtain the parsing result, and convert the parsing result into a feature vector; The output module is used to input the feature vector into a pre-constructed large-scale power industry model, and the large-scale power industry model outputs the output result corresponding to the output information based on the feature vector.

[0014] As can be seen from the above description, the power industry large-scale model optimization method and apparatus provided in this application obtains set role information, input data, task information, and output information; generates an input template according to preset template rules based on the role information, input data, task information, and output information; parses the input template to obtain the parsing result; converts the parsing result into a feature vector; inputs the feature vector into a pre-constructed power industry large-scale model; and the power industry large-scale model outputs an output result corresponding to the output information based on the feature vector. This application can improve the accuracy and reliability of the model in processing power business and supports power business scenarios in multiple fields. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the method flow of an embodiment of this application; Figure 2 This is a schematic diagram of the output result of one embodiment of this application; Figure 3 This is a schematic diagram of the output result of another embodiment of this application; Figure 4 This is a block diagram of the device structure according to an embodiment of this application; Figure 5 This is a block diagram of the electronic device structure according to an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0018] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0019] like Figure 1 As shown in the embodiments of this application, a large-scale model optimization method for the power industry is provided, including: S101: Obtain the set role information, input data, task information, and output information; In this embodiment of the application, specific professional roles can be assigned to the large model according to different business scenarios in the power industry, clarifying the business positioning and professional perspective of the model. Specific roles include, but are not limited to, power equipment diagnostic engineers, power grid dispatching experts, power grid planning analysts, and operation and maintenance specialists.

[0020] In some embodiments, before obtaining the set role information, input data, task information, and output information, the following steps are also included: Based on the role information, access the corresponding professional knowledge base; Based on the professional knowledge base, the pre-trained general large model is trained and adjusted to obtain a large model for the power industry.

[0021] In this embodiment, different roles correspond to exclusive professional knowledge bases. The large model executes corresponding business processing logic based on the professional knowledge base of a specific role. The professional knowledge base includes core regulations and standards of the power industry, historical case databases, and parameter threshold databases. For example, authoritative regulations and documents such as the "Power Equipment Maintenance Regulations," "Power Grid Dispatch Management Regulations," and "Power Safety Work Regulations" cover historical cases of equipment failures at different voltage levels such as 500kV, 220kV, and 110kV, typical power grid dispatch scenarios, planning scheme examples, and safety threshold rules for key parameters such as voltage, current, oil temperature, H2 concentration, and partial discharge. Various professional knowledge bases support real-time updates by accessing relevant power systems (such as equipment management systems and dispatch systems) through API interfaces to ensure the timeliness and accuracy of knowledge.

[0022] After setting specific roles, the corresponding professional knowledge base for each role is determined. This professional knowledge base is then invoked and loaded. The pre-trained general-purpose model is trained and adjusted based on this professional knowledge base to obtain a power industry-specific model suitable for handling the business logic of the specific role. For example, for a power grid dispatching expert, the general-purpose model, based on learning and training from the relevant professional knowledge base, yields a power industry-specific model capable of power resource dispatching; for a power equipment diagnostic engineer, the general-purpose model, based on learning and training from the relevant professional knowledge base, yields a power industry-specific model capable of fault diagnosis, and so on. In this way, by assigning roles and loading professional knowledge bases, the large model acquires a professional perspective from the power industry, improving the accuracy of the power industry-specific model in handling power business and achieving adaptive adaptation across multiple scenarios.

[0023] In some embodiments, configuration information such as role information, input data, task information, and output information can be configured from the input interface. For role information, the role type, the domain to which the role belongs, and the corresponding professional knowledge base can be selected. For input data, the type of input data can be configured, and the original input data can be imported. For task information, power parameters related to power business and additional verification switch states can be configured. For output information, the output type and the output keywords to be included in the expected output results of the large power industry model can be configured.

[0024] S102: Generate an input template according to the preset template rules based on the character information, input data, task information, and output information; In this embodiment, after configuring various information such as role information, input data, task information, and output information, the system automatically generates an input template according to the set template rules. The input template is a structured prompt template in XML format, including tags for role name, role domain, power parameter constraint, raw input data, input data type, output type, output keyword, and validation control. The configured information is filled into the corresponding tag items of the input template to generate the input template, specifically including: Set character name tags and character domain tags respectively, based on character name and character domain; Based on the input data, set the original input data label and the input data type label; where the original input data label is the input data, and the input data type label is the type of the input data; Set power parameter constraint labels based on power parameters; Set output type tags and output keyword tags according to the output type and output keywords respectively; Set the verification control label according to the verification switch status.

[0025] In this embodiment, based on the configured role type and its domain, the role name corresponding to the role type is entered in the role name tag (e.g., power grid dispatch expert), and the domain of the role is entered in the role domain tag. Based on the configured input data, the original input data is entered in the original input data tag, and the type of input data is entered in the input data type tag. Based on the configured task information, the power parameters are entered in the power parameter constraint tag, and the verification switch status is entered in the verification control tag. Based on the configured output information, the output type of the output result is entered in the output type tag, and the output keywords required to be included in the output result are entered in the output keyword tag. Standardized tags enable the structured organization of input information, constructing a prompt word template with power industry knowledge, ensuring that the large-scale power industry model can accurately capture the core requirements and constraints of the task.

[0026] In some methods, the input template's tags cover the core elements of the power business, specifically including: role name tags. <role>Role-related tags<business_domain> (Limited business areas, such as equipment diagnostics, dispatching decisions, etc.); power parameter constraint labels, including: voltage level<voltage_level> Safety threshold<safety_threshold> Equipment type<equipment_type> Equipment number<equipment_id> etc.; raw input data labels<input_text> (e.g., inspection reports, etc.), Input data type labels<input_type> (e.g., long text, tables, etc.); Output type tags<output_format> (Output format, such as Markdown, JSON, table, etc.), Output keyword tags.<output_elements> (e.g., fault symptoms, cause analysis, handling suggestions, etc.); verification control labels, including parameter verification control labels.<grid_param_verification> (Parameter verification switch, set to on or off according to the configured parameter verification switch status), security verification control label<safety_check> (Security verification switch, set to on or off according to the configured security verification switch status).

[0027] In this embodiment, by designing a structured prompt word template, the configured information is converted into a unified format input template, which enables the output results of the power industry large model to fully comply with the power industry regulations and the format requirements of different business scenarios. The time spent on manual correction is reduced to less than 5%, and it can be directly applied to actual business operations, such as maintenance plan formulation and dispatch instruction execution, which greatly improves business processing efficiency.

[0028] S103: Parse the input template to obtain the parsing result, and convert the parsing result into a feature vector; S104: Input the feature vector into the pre-built large power industry model, and the large power industry model outputs the output result corresponding to the output information based on the feature vector.

[0029] In this embodiment, after constructing the input template, the input template is parsed, and the parsing results are converted into feature vectors suitable for input into the power industry large model. The power industry large model focuses on the core parameters in the template based on the input feature vectors, combines the role's professional knowledge base, and processes them according to the business logic adapted to the role. For example, for equipment diagnosis tasks, the process is processed according to parameter collection, threshold comparison, fault location, cause analysis, and processing suggestions to obtain the corresponding output results. For resource scheduling tasks, the process is processed according to status monitoring, load forecasting, constraint verification, and instruction generation to obtain the corresponding output results.

[0030] In some embodiments, the input template is parsed to obtain a parsing result, and the parsing result is converted into a feature vector, including: When the input data type label is long text, the input data is segmented using a sliding window to obtain multiple short text paragraphs; Key information is extracted from each short text using a pre-trained language model for electricity. By integrating the key information from each short text, a semantic graph of the long text is constructed. Convert the semantic graph of a long text into a corresponding feature vector.

[0031] In this embodiment, for input data that is long text, such as power inspection reports, power grid dispatch instructions, system operation logs, etc., which are more than 1,000 characters long, a sliding window segmentation and domain model extraction method is adopted to fully retain key information and improve processing efficiency and accuracy.

[0032] Specifically, the sliding window size is adaptively adjusted based on the length and content density of the long text, automatically dividing it into multiple shorter texts. The length of the sliding window ranges from 300 to 1000 characters, with a preferred value of 500 characters. The overlap ratio between adjacent windows is 15% to 25%, preferably 20%, to avoid information gaps caused by segmentation. Paragraphs containing key information such as device numbers and failure times are automatically marked as important paragraphs, prioritizing the preservation of complete semantics.

[0033] After dividing the long text into multiple shorter texts, a pre-trained power industry language model is invoked. This model is used to extract key entities and generate structured summaries for each short text. Extracted key entities include, but are not limited to, equipment number, equipment type, fault occurrence time, fault location, operating parameters (including voltage, current, oil temperature, oil level, H2 concentration, SF6 gas pressure, partial discharge, etc.), operation instruction keywords, and maintenance personnel numbers. The structured summaries contain key information such as the core events of each segment, parameter change trends, and abnormal situations. Optionally, the pre-trained power industry language model is implemented based on the general BERT model and fine-tuned using a power industry corpus containing over 100,000 power-related texts.

[0034] After extracting key information from each short text, this key information is fused together, that is, the key entities extracted from each short text and the structured summary are merged to obtain a complete semantic graph of the long text. This semantic graph, along with the content of other tags in the input template, is then converted into feature vectors. These feature vectors are input into a large-scale power industry model for processing, ensuring that the model can process based on complete global information and avoiding misjudgments caused by partial local information. By combining adaptive sliding window segmentation of long text with key information extraction technology, the key information of long texts can be fully preserved, improving the processing speed of long texts by 30%, solving the problems of information loss and low efficiency in power long text processing. Simultaneously, the misjudgment rate of core business operations such as fault diagnosis and parameter judgment is reduced to below 10%, significantly improving the accuracy and reliability of the model output.

[0035] In some embodiments, the verification switch includes a parameter verification switch; the method further includes: When the verification control tag is enabled for parameter verification, the output result is verified according to the preset parameter rules to obtain the verified output result.

[0036] In this embodiment, to further improve the accuracy of the output results, parameter verification is performed based on the output results of the large-scale power industry model, according to the relevant rules of the power industry, to obtain the parameter-verified output results. Optionally, parameter verification can be selected by configuring the verification switch status; when the parameter verification switch is on, parameter verification is performed on the model's output results.

[0037] In some embodiments, parameter rules include parameter threshold rules and business logic rules; The output results are validated according to preset parameter rules, and the validated output results are obtained, including: The output results are validated according to parameter threshold rules and business logic rules. If the output results do not conform to the parameter threshold rules and / or business logic rules, the output results are corrected according to the parameter threshold rules and / or business logic rules.

[0038] In this embodiment, parameter threshold rules and business logic rules are pre-constructed based on relevant rules in the power industry. When parameter verification of the output results is required, the output results are verified according to the parameter threshold rules and business logic rules respectively. The power grid parameters, fault diagnosis conclusions, operation suggestions, etc. in the output are compared with the corresponding rules to detect whether there are problems such as parameter exceeding the standard, logical contradictions, or process violations. When it is determined that there is an error in the output results, the output results are corrected according to the corresponding rules to reduce the risk of misjudgment caused by model illusion.

[0039] The parameter threshold rules include safety thresholds for core operating parameters under different equipment types and voltage levels, such as the 500kV main transformer oil temperature safety threshold ≤85℃, the H2 concentration safety threshold ≤100μL / L, and the 110kV line current overload threshold of 115% of the rated current. The business logic rules include power operation process specifications, such as the dispatch instruction issuance process, equipment maintenance process, and fault association logic, such as the correlation between H2 concentration exceeding the standard and winding insulation faults.

[0040] In some embodiments, the output results are corrected according to parameter threshold rules and / or business logic rules, including: If the output results contain parameter threshold errors, business logic inconsistencies, or process specification errors; Errors related to parameter thresholds, business logic contradictions, and process specifications are corrected in sequence.

[0041] In this embodiment, parameter verification identifies misjudgments in the output results, such as incorrect parameter values, misjudged fault causes, and violations of operational suggestions. These errors are corrected based on corresponding baseline rules, with the correction priority from highest to lowest: parameter threshold errors, business logic inconsistencies, and process specification errors. Specifically, parameter threshold errors are corrected first, followed by business logic inconsistencies, and then process specification errors. Optionally, for complex misjudgments that cannot be automatically corrected, correction prompts can be output to guide manual review. The overall error correction rate can reach over 90%, effectively improving the accuracy of the model's output results.

[0042] In some embodiments, the verification switch includes a security verification switch; the method further includes: When the verification control tag is enabled for security verification, it checks whether there is sensitive information in the input data and output results according to the preset security rules. If the sensitive information exists, it is de-identified to obtain the input data and output results after security verification.

[0043] In this embodiment, to improve the security of input data and output results, the input data and output results of the large-scale power industry model are subjected to security verification according to relevant security rules, resulting in security-verified input data and output results. Optionally, security verification can be selected by configuring a verification switch status. When the security verification switch is enabled, security verification is performed on the model's input data and output results, reducing the risk of data leakage and unauthorized instruction injection.

[0044] In some methods, sensitive information in the power industry includes, but is not limited to, grid topology, core substation locations, maintenance personnel identification information, equipment passwords, and dispatch command encryption keys. When sensitive information is detected in input data or output results, de-identification processing is performed using methods such as replacement (e.g., replacing a specific location with a specific area), masking (e.g., replacing part of a number with a specific character), and deletion (e.g., deleting irrelevant sensitive ancillary information) to ensure that the de-identified data does not affect business processing and to prevent information leakage.

[0045] In some methods, security rules containing sensitive information are constructed based on regular expressions. The sensitive information includes keywords related to illegal operations such as forced shutdown, illegal power outage, modification of protection settings, and bypassing security checks, as well as keywords related to the State Grid core topology and military area power supply schemes.

[0046] In some approaches, considering that while the input data and output results may not contain obvious sensitive information, they may contain potentially leaky semantics (e.g., stopping power supply to substation XX under the guise of equipment maintenance), to ensure comprehensive and accurate detection and interception, a pre-built power industry semantic recognition model is used to perform semantic analysis on the input data and output results to identify potential risks such as forged dispatch orders, intentions for illegal operations, and intentions to steal sensitive information. Thus, through a two-layer security filtering mechanism of security rules and the power industry semantic recognition model, the output of illegal operation commands can be completely blocked, avoiding the risk of information leakage, meeting the high security requirements of the power industry, and ensuring the safe operation of the power grid and national energy security. Optionally, the power industry semantic recognition model is implemented based on an LSTM model and trained using a corpus of illegal commands.

[0047] In some embodiments, the power industry large model outputs the corresponding output results based on the feature vectors, including: According to the output type corresponding to the output type tag, the output includes the standardized output results of the output keywords corresponding to the output keyword tag.

[0048] In this embodiment, in order to conform to the standard format and business requirements of the power industry, the output results of the large power industry model can be standardized by configuring the output type and output keywords. This ensures that the output results are standardized results that conform to the output type and contain output keywords, thus ensuring the consistency and usability of the output.

[0049] In some methods, when the output of the large power industry model is a power grid analysis report (such as a fault analysis report or an operation status analysis report), the output type can be set to Markdown format. The output keywords must include core elements such as fault phenomena, operation status, cause analysis, handling suggestions, and optimization schemes, and additional parameter comparison tables, trend chart reference labels, etc.

[0050] In some methods, when the output of the large power industry model is an operation and maintenance instruction, the output type can be set to JSON format, and the output keywords can include instruction number, instruction type, operation object (equipment number), operation content, execution time window, safety precautions, and the person who issued the instruction.

[0051] In some methods, when the output of the power industry large model is a power grid planning scheme, the output type can be set to text description and multi-dimensional table format. The output keywords of the text description part include planning background, planning objectives, overall scheme, etc., while the output keywords of the table part include load forecast data, equipment additions, renovation list, investment estimate, construction schedule, benefit analysis, etc.

[0052] In some methods, when the output of the large-scale power industry model is a dispatch instruction, the output type can be set to a standardized text format. The output keywords include fields such as instruction number, receiving unit, execution time, instruction content, instruction basis, issuer, and reviewer, which comply with the format requirements stipulated in the "Regulations on Power Grid Dispatch Management".

[0053] In some methods, the output results can also be format-validated according to the output type to ensure the integrity of the fields and the standardization of the format. The pass rate of the validation must reach more than 99%.

[0054] In some embodiments, to achieve continuous iterative optimization of the performance of the large-scale power industry model, reinforcement learning is used to fine-tune the model based on real-time acquired power data. Specifically, real-time power grid operation status data, real-time equipment monitoring data, and the latest fault cases are acquired from power grid SCADA systems, equipment online monitoring systems, and dispatch automation systems. The model training data and case library are updated based on the acquired data to improve the model's adaptability to real-time business scenarios. Using diagnostic accuracy, format compliance rate, and safety compliance rate as a comprehensive reward function, historical output results (including correctly output cases, incorrectly judged and corrected cases, and manually reviewed cases) are collected from actual applications of the model. Reinforcement learning is used to fine-tune the model based on these historical output results, continuously optimizing the model weight parameters and improving the model's performance stability and accuracy in specific tasks within the power industry. After a predetermined period of fine-tuning iterations, the model's diagnostic accuracy can be improved to over 95%, adapting to the business development and technological upgrades of the power industry.

[0055] In some approaches, the large-scale power industry model can be directly deployed on core business systems of power companies, such as intelligent operation and maintenance platforms, power grid dispatch centers, and equipment management systems. It supports localized and private deployments, adapts to the security environment of the power industry's intranet, ensures the security of data transmission and storage, supports containerized deployment and cluster expansion, and meets the needs of high-concurrency business.

[0056] In some specific embodiments, the user selects the role of a power equipment diagnostic engineer through the user interface. The system automatically loads the "500kV Transformer Maintenance Regulations," the "Power Transformer Fault Diagnosis Guidelines," a 500kV main transformer historical fault case library (containing more than 1,000 similar equipment fault records), and a parameter threshold library (500kV main transformer oil temperature safety threshold ≤85℃, H2 concentration safety threshold ≤100μL / L, partial discharge ≤100pC, etc.). The user inputs a 2,000-word inspection report for the #2 main transformer (including information such as 1200 hours of operation, today's inspection oil temperature of 95℃, H2 concentration of 150μL / L, normal operation of the cooling system, and no obvious abnormal noise). Based on the role information and input data, a structured prompt word template is automatically generated, as shown in the example below: <role> Power Equipment Diagnostic Engineer< / role> <business_domain> Equipment Fault Diagnosis< / business_domain> <voltage_level> 500kV< / voltage_level> <equipment_type> Main change< / equipment_type> <equipment_id> #2 Main Transformer< / equipment_id> <safety_threshold> Oil temperature ≤ 85℃; H2 concentration ≤ 100 μL / L; Partial discharge ≤ 100 pC< / safety_threshold> <input_type> Long text< / input_type> <input_text> [Inspection Report] #2 main transformer has been running for 1200 hours. Today's inspection showed an oil temperature of 95℃, H2 concentration of 150μL / L, normal cooling system operation with no obvious abnormal noise, load rate of 85%, and ambient temperature of 32℃...< / input_text> <output_format> Markdown< / output_format> <output_elements> Fault symptoms, cause analysis, and troubleshooting suggestions< / output_elements> <grid_param_verification> Open< / grid_param_verification> For long text input data, a 500-character sliding window was used to segment the 2000-word inspection report, dividing the long text into 4 short texts with adjacent segments overlapping by 100 characters. A pre-trained power industry language model was used to extract key entities from each short text, including: #2 main transformer, 500kV, operating time 1200h, oil temperature 95℃, H2 concentration 150μL / L, cooling system normal, load rate 85%, and ambient temperature 32℃. After generating structured summaries for each short text, the key information from each short text was fused to construct a complete semantic graph of the long text. The contents of other fields parsed from the input template and the semantic graph were converted into feature vectors. These feature vectors were input into a large power industry model, which output: "#2 main transformer cooling system malfunction, causing oil temperature and H2 concentration to exceed standards; replacement of cooling system components is recommended."

[0057] For the output of the large-scale power industry model, the parameters "oil temperature 95℃, H2 concentration 150μL / L" were compared with the parameter threshold rules. It was found that both exceeded the safety threshold. At the same time, according to the business logic rules, the conclusion that the cooling system was operating normally contradicted the conclusion of "cooling system failure". Therefore, the output result was determined to be a misjudgment. Based on the related cases of "excessive H2 concentration + excessive oil temperature + normal cooling system" in the historical case library, the cause of the failure was automatically corrected. The corrected output result was: "#2 main transformer oil temperature and H2 concentration exceed the standard, cooling system is operating normally, the cause of failure is likely abnormal winding insulation."

[0058] like Figure 2 As shown, following the requirements of the output type label, the parameter-corrected output is converted to Markdown format. The output includes three elements: fault phenomenon, cause analysis, and handling suggestions. The output is checked for sensitive information; if any is found, it is anonymized, including removing the specific installation location of the main transformer and the inspection personnel number. Keyword filtering reveals no sensitive keywords, and semantic analysis detects no intention of unauthorized operation. The security check passes. The parameter-corrected output is then output in standard Markdown format.

[0059] Simultaneously, the system is connected to the power grid SCADA system, and the case library is updated based on the oil temperature and H2 concentration changes after the subsequent load rate adjustment of the #2 main transformer. The model weights are optimized using diagnostic accuracy as a reward function (if subsequent detection confirms winding insulation abnormalities, the diagnosis is considered accurate and a positive reward is given), thereby improving the diagnostic accuracy for similar faults.

[0060] In another specific embodiment, the power grid dispatch expert role is selected, and the system loads the "Power Grid Dispatch Management Regulations," "Dispatch Instruction Operation Specifications," power grid load forecasting case library, and 110kV line dispatch rule library, etc. The user input data is: the load on the 110kV XX line is continuously increasing, the current load rate is 110%, and a dispatch instruction needs to be generated to transfer part of the load to the adjacent 110kV YY line. Based on the role information and input data, a structured input template is generated. <role> Power grid dispatching experts< / role> <business_domain> Power grid dispatch< / business_domain> <voltage_level> 110kV< / voltage_level> <input_text> The load on the 110kV XX line continues to rise, currently at 110% load rate. It is necessary to transfer some load to the adjacent 110kV YY line to ensure the safe operation of the line.< / input_text> <output_format> JSON< / output_format> <output_elements> Instruction number, receiving unit, execution time, instruction content, safety precautions, and the person who issued the instruction.< / output_elements> <safety_check> Open< / safety_check> like Figure 3 As shown, after the large-scale power industry model outputs dispatch instructions, it verifies whether the output dispatch instructions conform to the dispatch process specifications (such as load transfer operation sequence and safety threshold constraints). After the verification is passed, the dispatch instructions are converted into standard JSON format. The dispatch instructions are then subjected to security verification. If there is sensitive information, it is desensitized. Finally, the dispatch instructions in JSON format are output, which can be directly imported into the dispatch operation terminal for execution by operation and maintenance personnel.

[0061] The power industry large-scale model optimization method provided in this application generates an input template according to template rules based on various configuration information. The parsing results of the input template are converted into feature vectors, which are then input into a pre-constructed power industry large-scale model. The large-scale model outputs results corresponding to the output information based on the feature vectors. This method supports multiple business scenarios such as power equipment diagnosis, grid dispatching, and operation and maintenance planning. Through role switching, knowledge base updates, and format adaptation, it can quickly respond to different business needs. Through standardized and professional optimization strategies, it significantly enhances the application value of the large-scale model in the power industry, reduces labor costs, decreases unplanned downtime, improves the scientific and timely nature of grid dispatching decisions, and ensures the safe and stable operation of the grid. It has significant economic and social benefits (estimated to reduce power company operation and maintenance costs by 15%-20% and improve fault diagnosis efficiency by more than 40%) and broad industrial application prospects, and can be widely promoted and applied in various power companies.

[0062] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0063] It should be noted that the above description describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims may be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0064] like Figure 4 As shown in the figure, this application provides a large-scale model optimization device for the power industry, including: The acquisition module is used to acquire the set role information, input data, task information, and output information; The generation module is used to generate input templates according to preset template rules based on role information, input data, task information, and output information. The parsing module is used to parse the input template, obtain the parsing result, and convert the parsing result into a feature vector; The output module is used to input feature vectors into a pre-built large-scale power industry model, and the large-scale power industry model outputs output results corresponding to the output information based on the feature vectors.

[0065] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware.

[0066] The apparatus described above is used to implement the corresponding methods in the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0067] Figure 5 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0068] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0069] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0070] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0071] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0072] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0073] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0074] The electronic devices described above are used to implement the corresponding methods in the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0075] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0076] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0077] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0078] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0079] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this disclosure.< / role>

Claims

1. A method for optimizing a large-scale power industry model, characterized in that, include: Retrieve the set character information, input data, task information, and output information; Based on the character information, input data, task information, and output information, an input template is generated according to preset template rules; The input template is parsed to obtain the parsing result, and the parsing result is converted into a feature vector; The feature vector is input into a pre-constructed large-scale power industry model, and the large-scale power industry model outputs an output result corresponding to the output information based on the feature vector.

2. The method according to claim 1, characterized in that, Before obtaining the set role information, input data, task information, and output information, the process also includes: Based on the role information, the corresponding professional knowledge base is invoked; Based on the aforementioned professional knowledge base, the pre-trained general large model is trained and adjusted to obtain the power industry large model.

3. The method according to claim 1, characterized in that, The template rules include role name tags, role domain tags, power parameter constraint tags, raw input data tags, input data type tags, output type tags, output keyword tags, and verification control tags; the role information includes role name and role domain, the task information includes power parameters and verification switch status, and the output information includes output type and output keywords; Based on the role information, input data, task information, and output information, an input template is generated according to preset template rules, including: Based on the character name and the character domain, set the character name tag and the character domain tag respectively; Based on the input data, set the original input data label and the input data type label; wherein, the original input data label is the input data, and the input data type label is the type of the input data; Based on the power parameters, set the power parameter constraint labels; Based on the output type and output keywords, set the output type label and output keyword label respectively; The verification control tag is set according to the verification switch status.

4. The method according to claim 3, characterized in that, The input template is parsed to obtain the parsing result, and the parsing result is converted into a feature vector, including: When the input data type label is long text, the input data is segmented using a sliding window to obtain multiple short text paragraphs; Key information is extracted from each short text using a pre-trained language model for electricity. By integrating the key information from each short text, a semantic graph of the long text is constructed. The long text semantic graph is converted into a corresponding feature vector.

5. The method according to claim 3, characterized in that, The verification switch includes a parameter verification switch; the method further includes: When the verification control tag is enabled for parameter verification, the output result is verified according to the preset parameter rules to obtain the verified output result.

6. The method according to claim 5, characterized in that, The parameter rules include parameter threshold rules and business logic rules; The step of verifying the output result according to preset parameter rules to obtain the verified output result includes: The output result is validated according to the parameter threshold rules and business logic rules. If the output result does not conform to the parameter threshold rules and / or business logic rules, the output result is corrected according to the parameter threshold rules and / or business logic rules.

7. The method according to claim 6, characterized in that, The output result is corrected according to the parameter threshold rules and / or business logic rules, including: If the output results contain parameter threshold errors, business logic inconsistencies, or process specification errors; Errors related to parameter thresholds, business logic contradictions, and process specifications are corrected in sequence.

8. The method according to claim 3, characterized in that, The verification switch includes a security verification switch; the method further includes: When the verification control tag is enabled for security verification, the input data and output results are checked for sensitive information according to preset security rules. If the sensitive information exists, it is desensitized to obtain the input data and output results after security verification.

9. The method according to any one of claims 3-8, characterized in that, The power industry large model outputs output results corresponding to the output information based on the feature vector, including: According to the output type corresponding to the output type label, output standardized output results including output keywords corresponding to the output keyword label are output.

10. A large-scale model optimization device for the power industry, characterized in that, include: The acquisition module is used to acquire the set role information, input data, task information, and output information; The generation module is used to generate an input template according to the role information, input data, task information, and output information, and in accordance with preset template rules. The parsing module is used to parse the input template, obtain the parsing result, and convert the parsing result into a feature vector; The output module is used to input the feature vector into a pre-constructed large-scale power industry model, and the large-scale power industry model outputs the output result corresponding to the output information based on the feature vector.

Images