Power task processing method and device, storage medium and electronic equipment

Abstract

The invention discloses an electric power task processing method and device, a storage medium and electronic equipment. The method comprises the following steps: acquiring electric power task data, and inputting the electric power task data into a to-be-trained model, so that the to-be-trained model processes the electric power task data according to an inference strategy of the to-be-trained model, and generates a plurality of candidate answer data for the electric power task data; analyzing each piece of candidate answer data based on a preset reference model to obtain character-level-by-character probability distribution data corresponding to each piece of candidate answer data; performing compliance verification on each piece of candidate answer data based on a preset discrimination model to obtain a binary score of each piece of candidate answer data; and on the basis of the binary score and the character-level probability distribution data one by one, parameters of the to-be-trained model are updated through a dynamic constraint mechanism, and the to-be-trained model with the updated parameters serves as an electric power task execution model and is used for processing input electric power task data on the edge computing equipment and outputting a task execution result.

Description

Technical Field

[0001] This application relates to the field of power grid technology, and in particular to a power task processing method, apparatus, storage medium and electronic device. Background Technology

[0002] In edge computing scenarios at power grid stations, core tasks such as generating power equipment maintenance plans and rapidly responding to dispatch instructions place extremely high demands on the real-time performance and reliability of models. Edge node computing resources typically only support models with fewer than 500M parameters, and task response times must be kept below 500ms to meet the immediate needs of on-site maintenance and dispatch decisions. However, current models capable of accurately handling these power tasks are mostly large-scale language models with more than 10B parameters. These models require hundreds of GB of storage resources, and single-task inference often takes more than 2 seconds. Their storage requirements and computing power consumption far exceed the capacity limits of edge devices, making them unsuitable for direct deployment at power grid station edge nodes. Existing lightweight models adapted for edge deployment, lacking sufficient learning of power industry expertise (such as maintenance procedures and dispatch standards), generally have an accuracy rate below 75% when handling tasks such as equipment fault analysis and compliant instruction generation. This makes it difficult to meet the stringent requirements of the power industry for operational safety and instruction compliance, leading to a predicament in intelligent task processing at power grid station edge nodes. Summary of the Invention

[0003] In view of the above problems, this application provides a power task processing method, apparatus, storage medium and electronic device.

[0004] To solve the above-mentioned technical problems, this application proposes the following solution:

[0005] In a first aspect, this application provides a method for processing power tasks. The method includes: acquiring power task data; inputting the power task data into a model to be trained, so that the model to be trained can process the power task data according to its own reasoning strategy to generate multiple candidate response data for the power task data; parsing each candidate response data according to a preset benchmark model to obtain character-level probability distribution data corresponding to each candidate response data; performing compliance verification on each candidate response data according to a preset discrimination model to obtain a binary score for each candidate response data; updating the parameters of the model to be trained through a dynamic constraint mechanism based on the binary score and the character-level probability distribution data; the updated model to be trained serves as a power task execution model, used to process the input power task data on an edge computing device and output the task execution result.

[0006] Secondly, this application provides an electric task processing device, which includes: a processing module for acquiring electric task data and inputting the electric task data into a model to be trained, so that the model to be trained can process the electric task data according to its own reasoning strategy and generate multiple candidate response data for the electric task data.

[0007] The parsing module is used to parse each candidate answer data according to a preset benchmark model, and obtain the character-level probability distribution data corresponding to each candidate answer data.

[0008] The verification module is used to perform compliance verification on each candidate answer data based on a preset discrimination model, and obtain a binary score for each candidate answer data.

[0009] The update module is used to update the parameters of the model to be trained based on binary scores and character-by-character probability distribution data through a dynamic constraint mechanism. The updated model to be trained serves as a power task execution model, which is used to process the input power task data on edge computing devices and output the task execution results.

[0010] To achieve the above objectives, according to a third aspect of this application, a storage medium is provided, the storage medium including a stored program, wherein, when the program is executed, the device where the storage medium is located is controlled to perform the power task processing method of the first aspect described above.

[0011] To achieve the above objectives, according to a fourth aspect of this application, an electronic device is provided, the device including at least one processor, and at least one memory and bus connected to the processor; wherein the processor and memory communicate with each other via the bus; the processor is used to call program instructions in the memory to execute the power task processing method of the first aspect described above.

[0012] By employing the above-described technical solution, the technical solution provided in this application has at least the following advantages:

[0013] This application acquires power task data and inputs it into a training model, allowing the model to generate multiple candidate responses for that power task data based on its own reasoning strategy. This process enables the training model to autonomously explore different response logics and expressions around a single power task, providing a diverse learning sample base for subsequent model parameter updates and avoiding the inability of the model to fully adapt to the actual processing needs of power tasks due to a single sample. Next, leveraging the knowledge accumulated by the benchmark model in the power field, its judgment on the character generation logic in the candidate responses is passed to the training model in the form of a fine-grained probability distribution, providing precise supervision and compensating for the training model's deficiencies in power field expertise and character-level semantic understanding. Then, based on a pre-set discriminant model, each candidate response is subjected to compliance verification, obtaining a binary score for each candidate response. Starting from the compliance standards of the power industry, the quality of candidate responses is clearly distinguished, ensuring that the training model subsequently learns only from high-quality responses that meet domain requirements, avoiding misleading model training by erroneous or non-compliant response logic. Finally, based on the binary score and character-by-character probability distribution data, the parameters of the model to be trained are updated through a dynamic constraint mechanism. This allows the model to absorb professional knowledge in the power field by relying on the character-by-character probability distribution of the benchmark model during the parameter update process, while also avoiding erroneous logic in non-compliant answers based on the binary score. The resulting power task execution model can be deployed on edge computing devices, satisfying the computing power and storage constraints of edge devices while ensuring the processing effect of power task data.

[0014] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0016] Figure 1 A flowchart illustrating a power task processing method provided in an embodiment of this application is shown;

[0017] Figure 2 This illustration shows an example schematic diagram of a power task processing method provided in an embodiment of this application;

[0018] Figure 3A schematic diagram of the structure of a power task processing device provided in an embodiment of this application is shown;

[0019] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0020] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0021] In the embodiments of this application, the terms "first," "second," etc., do not have a logical or temporal dependency, nor do they limit the quantity or execution order. It should also be understood that although the following description uses the terms "first," "second," etc., to describe various elements, these elements should not be limited by the terms. These terms are merely used to distinguish one element from another.

[0022] In this application, the term "at least one" means one or more, and the term "multiple" means two or more.

[0023] It should also be understood that the term “if” can be interpreted as “when” or “upon”, or “in response to determination” or “in response to detection”. Similarly, depending on the context, the phrase “if determination…” or “if detection [the stated condition or event]” can be interpreted as “when determination…” or “in response to determination…” or “when detection [the stated condition or event]” or “in response to detection [the stated condition or event]”.

[0024] The power task processing method, apparatus, storage medium, and electronic equipment disclosed in this application are mainly applied to edge computing scenarios with limited computing power, such as power grid stations. They enable efficient training of lightweight models and accurate task processing for specific power tasks such as generating power dispatch instructions, querying equipment maintenance plans, verifying the compliance of invoices, and answering power knowledge questions.

[0025] Taking the edge computing node deployment project of a 110kV substation in a certain region as an example, this scenario imposes clear practical constraints on the model: the computing resources of the edge nodes can only support the operation of models with fewer than 500M parameters. While existing large-scale power industry models have the ability to accurately answer maintenance-related questions (with parameter counts generally exceeding 10B), they are limited by the massive computing power and storage requirements, making them unsuitable for edge deployment scenarios. Therefore, it is urgent to build a lightweight model that can quickly respond to equipment maintenance work order query needs. The power task processing method is explained in detail below with reference to the attached diagram. Figure 1 This is a flowchart illustrating a power task processing method provided in this application. Specifically, it includes the following steps:

[0026] Step 110: Obtain power task data and input the power task data into the model to be trained so that the model can process the power task data according to its own reasoning strategy and generate multiple candidate response data for the power task data.

[0027] The power task data here comes from the substation's historical equipment maintenance work orders over the past three years, relevant clauses in the current "Power Equipment Maintenance Regulations" for querying, and real-time maintenance questions input by on-site maintenance personnel through edge node terminals. Examples include "maintenance procedures when the oil level of main transformer No. 1 is abnormal" and "passing standards for circuit breaker mechanical characteristic tests." This power task data is then input into the model to be trained (a Transformer architecture model with an initial parameter count of 400M, used as the student model π). s The model to be trained independently processes each power task data based on its basic power knowledge and reasoning strategies learned during the pre-training phase, generating multiple candidate response data. Taking "maintenance steps when the oil level of main transformer No. 1 is abnormal" as an example, the model to be trained generates 3 candidate responses (K=3, i.e., 3 rollouts): Candidate response 1 details the steps of "shutdown - drain oil - check seals - replenish oil - trial operation"; candidate response 2 omits the "seals check" step and directly states "shutdown - replenish oil - trial operation"; candidate response 3 includes the complete process of "shutdown - check oil tank - investigate leaks - handle defects - replenish oil - withstand voltage test - trial operation".

[0028] The generation process of the aforementioned candidate answers is essentially a preliminary response of the model to the task requirements based on its initial knowledge of electricity. However, there are significant differences in the completeness and compliance of the steps among the different candidate answers. This difference reflects both the limitations of the model's current reasoning ability and highlights the potential problem of insufficient sample coverage when relying solely on existing task data for training. To improve the effectiveness of subsequent model training, it is necessary to first expand the electricity task dataset through data augmentation techniques to build more comprehensive training data support. The specific processing flow is as follows.

[0029] First, obtain multi-source heterogeneous raw data in the power field. In the data source system of this provincial power grid company, the raw data covers four categories of core data: The first category is structured equipment parameter data, which comes from the monitoring and data acquisition systems of each substation and the equipment ledger management platform, including structured fields such as the model, rated voltage, rated current, operation years, and historical maintenance records of core equipment such as 110kV - 500kV transformers, circuit breakers, and disconnectors. It is stored in the form of Excel tables and relational database tables (MySQL), and the data volume of a single table reaches hundreds of thousands of records. The second category is semi-structured maintenance work order texts, which come from the power operation and maintenance management system. Each work order contains fields such as work order number, equipment ID, fault phenomenon description, maintenance personnel, processing steps, and acceptance results. The field format is presented in the form of being wrapped by HTML tags or key-value pairs. For example, "<Fault phenomenon> The oil temperature of the No. 1 main transformer is too high, exceeding 85°C< / Fault phenomenon>", and the cumulative stock of work orders exceeds 500,000. The third category is unstructured dispatching voice transcription texts, which are collected from the voice call records of the power grid dispatching center and are converted into texts through voice recognition technology (such as a voice transcription model based on Transformer). The content includes real-time instruction interactions between dispatchers and substation duty officers (such as "Notify the 220kV Chengdong Substation to transfer the load of #2 main transformer to #1 main transformer"), fault emergency handling communications, etc. The text format is a dialogue record without a fixed structure, and the length of a single record ranges from 10 to 500 characters, and the cumulative data volume exceeds 100,000. The fourth category is power system alarm logs, which come from the fault wave recording devices and relay protection devices of each substation and record the alarm information triggered by equipment abnormalities, including alarm time, alarm type (such as overcurrent alarm, gas alarm), associated equipment ID, alarm level, etc. It is stored in the form of TXT log files, and the daily new log exceeds 10,000. Through data interfaces (such as database JDBC interfaces, file FTP transfer protocols), the above multi-source data is uniformly collected into a distributed data storage platform (such as Hadoop HDFS) to form a raw data pool.

[0030] Secondly, perform unified format conversion and noise filtering on the original data. For the format differences of different types of data, adopt customized conversion strategies: For structured device parameter data, use an ETL tool (such as Kettle) to uniformly convert the data in Excel and MySQL into JSON format, ensuring the standardization of field names (such as converting "device model" to "device_model" and "service life" to "service_life"). For semi-structured maintenance work order texts, use regular expressions to extract the core content in HTML tags or key-value pairs, remove tag symbols and redundant format characters (such as converting "<Maintenance personnel>Zhang San< / Maintenance personnel>" to "Maintenance personnel: Zhang San"), and convert it to plain text format. For unstructured dispatching voice transcription texts, perform sentence segmentation through natural language processing tools (such as jieba分词), and remove meaningless characters generated during the speech recognition process (such as filler words like "um" and "ah", and incorrect recognition garbled characters like "#"). For power system alarm logs, parse the field delimiters (such as spaces and commas) in the TXT file, convert the log content into structured JSON format, and supplement missing fields (such as automatically labeling logs without an "alarm level" as "general", "important", and "urgent" based on the alarm type). In the noise filtering stage, focus on removing invalid data and redundant information: Screen out invalid data through data integrity verification (such as records missing the core field "device ID") and logical consistency verification (such as device parameters with a negative "service life"), and the deletion ratio is approximately 3% of the total data volume. Remove duplicate maintenance work orders and duplicate alarm logs (such as the same alarm triggered repeatedly by the same device within a short period) through text deduplication algorithms (such as the SimHash algorithm), and the data volume after deduplication is reduced by approximately 15%. Manually assist in screening and deleting records with ambiguous semantics in the dispatching voice transcription texts (such as single texts only containing content unrelated to tasks like "Received" and "Understood"), and finally obtain preprocessed data with a unified format and cleared noise.

[0031] Note: "jieba分词" is a Chinese term and is left as is in the translation as it's likely a specific tool name in the original context. If it's a misspelling and should be something like "jieba tokenizer", the translation should be adjusted accordingly.Subsequently, the data was categorized and labeled according to power task type to form a basic task dataset. Based on the core business scenarios of the power industry, the preprocessed data was divided into five major power task types: equipment maintenance query tasks (corresponding to maintenance work order text and equipment parameter data), dispatch instruction generation tasks (corresponding to dispatch voice-to-text transcription), alarm cause analysis tasks (corresponding to power system alarm logs), equipment status assessment tasks (corresponding to equipment parameter data and alarm logs), and power knowledge Q&A tasks (corresponding to maintenance procedure text and equipment manual data). For each piece of preprocessed data, its corresponding task type label was assigned, and a task data structure was constructed. Taking the equipment maintenance query task as an example, the "fault phenomenon description" in the maintenance work order was used as the "input question" of the power task data (e.g., "Main transformer oil temperature is too high, exceeding 85℃"), and the "handling steps" were used as the "reference answer" of the task (e.g., "1. Reduce the main transformer load to 60% of the rated load; 2. Check the cooling fan operation status; 3. Check if the oil temperature sensor is normal; 4. If the oil temperature continues to rise, shut down the main transformer"), and the label was "Maintenance Query". For the dispatch instruction generation task, the "load adjustment requirement" in the dispatch speech-to-text is used as the "input question" (e.g., "Regional load increases by 10%, requiring adjustment of the main transformer load at the 220kV Chengdong Substation"), and the "dispatch instruction" in the speech recording is used as the "reference answer" (e.g., "220kV Chengdong Substation: Transfer the load of main transformer #2 to main transformer #1, controlling load fluctuation within 5% during the transfer process"). The label is "dispatch generated". Through a combination of manual and automatic labeling (automatic labeling accounts for 80%, and manual review and correction accounts for 20%), the classification and labeling of 5 million preprocessed data points are completed, forming the basic task dataset, which is stored in the data labeling platform.

[0032] Next, for each power task data in the basic task dataset, semantically equivalent extended task data is generated. First, core entities are extracted from the power task data using entity recognition algorithms: a BERT-based power domain entity recognition model (pre-trained on a power industry corpus, capable of recognizing "equipment entities" such as "main transformer #1" and "circuit breaker #2", "parameter entities" such as "oil temperature 85℃" and "load 10%", and "operation entities" such as "load transfer" and "outage") is used to extract entities from the "input question" of each task data. Taking the "input question" "main transformer #1 oil temperature too high, exceeding 85℃" in the basic task data as an example, the core entities extracted are the equipment entity "main transformer #1" and the parameter entity "oil temperature 85℃". Subsequently, the attribute information and associated entities related to the core entity are queried from the power industry knowledge graph. This knowledge graph covers entity nodes such as power grid equipment, parameters, operations, and fault types, as well as relationship edges such as "equipment-parameter", "equipment-associated equipment", and "parameter-fault cause". The associated attributes (such as model "SSZ11-120000 / 220" and cooling method "forced oil circulation air cooling") and associated equipment (such as "#1 cooling fan" and "oil temperature sensor #1") of "No.1 main transformer" are queried through the graph database query language. The associated fault causes (such as "cooling system failure" and "overload") of "oil temperature 85℃" are also queried. Based on the core entity, associated attribute information, and associated entity reconstruction expression logic, extended task data with semantic equivalent to the original power task data is generated: the core semantics of the original problem ("main transformer oil temperature too high") are retained, and associated information is replaced or supplemented. For example, the original problem is reconstructed as "The No. 1 main transformer with model SSZ11-120000 / 220 has an oil temperature exceeding 85℃, which may be related to the cooling system. How should it be handled?" "The No. 1 cooling fan of the No. 1 main transformer is faulty, causing the oil temperature to rise to 85℃. What maintenance steps should be taken?" etc. Each basic task data generates 3-5 extended task data, and the amount of task data after expansion increases from 5 million to 20 million.

[0033] Then, power scenario disturbance factors are introduced to dynamically adjust non-core fields in the extended task data. Power scenario disturbance factors include four categories: equipment model disturbance, parameter value disturbance, scenario environment disturbance, and description style disturbance. Equipment model disturbance is based on a database of similar equipment models (e.g., collecting common 220kV main transformer models "SSZ11-120000 / 220" and "S13-100000 / 220"), replacing the equipment models in the extended task data with other models of the same type. For example, "main transformer No. 1 with model SSZ11-120000 / 220" is adjusted to "main transformer No. 1 with model S13-100000 / 220". Parameter value disturbance involves adjusting the values ​​of parameter entities within a reasonable range (e.g., oil temperature of 85℃ can be adjusted to 82℃-88℃, load of 10% can be adjusted to 8%-12%), ensuring that the adjusted values ​​conform to the operating rules of the power equipment (e.g., oil temperature does not exceed the equipment's rated maximum oil temperature of 95℃). The system incorporates environmental perturbations to supplement information on different operating scenarios, such as "In low-temperature winter conditions, the oil temperature of main transformer No. 1 exceeds 85℃" and "During peak grid load periods, the oil temperature of main transformer No. 1 exceeds 85℃." It also perturbs the presentation style by converting the task data to different styles (e.g., changing declarative sentences to interrogative sentences like "The oil temperature of main transformer No. 1 exceeds 85℃, how should it be inspected?", and concise statements to detailed statements like "During operation, maintenance personnel found that the oil temperature of main transformer No. 1 at a certain substation reached 85℃, exceeding the normal operating range (40℃-80℃), and a maintenance plan needs to be developed"). These perturbation adjustments further enrich the scenario diversity of the expanded task data, avoiding overfitting during model training. After the adjustments, the expanded task data volume remains at 20 million records, but the scenario coverage is improved by 40%.

[0034] Finally, the basic task data and the adjusted extended task data are merged to form the final power task dataset. Data fusion tools (such as Apache Spark) are used to perform format alignment (unifying the two types of data into a triplet structure of "input question-reference answer-task type"), field validation (ensuring no missing fields and consistent data format), and hierarchical storage according to task type (e.g., "maintenance query" task data is stored in a separate dataset file, and "scheduling generation" task data is stored separately).

[0035] Step 120: Based on the preset benchmark model, analyze each candidate answer data separately to obtain the character-level probability distribution data corresponding to each candidate answer data.

[0036] First, the candidate answer data is processed by character sequence splitting. For candidate answer 1, a specialized text processing tool for the power industry is used to split it character by character, while preserving the order of the characters in the original answer, forming an ordered character sequence. The character sequence splitting result of candidate answer 1 is denoted as τ1=(y1,1 , y 1,2 , …, y 1,48 ), where y 1,1 = "#", y 1,2 = "3", y 1,3 = "broken", y 1,4 = "circuit", y 1,5 = "breaker", y 1,6 = "disconnecting", y 1,7 = "switch", …, y 1,48 = "function", i = 1 (representing the 1st candidate answer data), T = 48 (representing the total length of the character sequence corresponding to this candidate answer data). During the splitting process, synchronously and associatively save the original power task data x = "#3 breaker disconnecting switch coil burned out" corresponding to this candidate answer data, ensuring that subsequent probability calculations can combine the task context and avoid isolated character probability analysis that is divorced from the fault scenario.

[0037] Secondly, for each character position t in the ordered character sequence, construct the combined input data for character position t. For each character position t (1 ≤ t ≤ 48), extract all the characters before this position to form a prefix character sequence y1,<t (i.e., the character set from y1,1 to y1,t−1), and combine it with the original power task data x to form the combined input data for this character position. Taking t = 10 (corresponding to the character y 1,10 = "old", which is located in the statement "Remove the old disconnecting switch coil") as an example, the prefix character sequence y1,<10 is "#3 breaker disconnecting switch coil burned out repair plan: 1. Disconnect the power supply of the breaker; 2. Remove", and combine this prefix sequence with the original power task data x = "#3 breaker disconnecting switch coil burned out" to obtain the combined input data: "Power task: #3 breaker disconnecting switch coil burned out; Current generated content: #3 breaker disconnecting switch coil burned out repair plan: 1. Disconnect the power supply of the breaker; 2. Remove". Through this combination method, ensure that when the baseline model calculates the probability of character position t, it can simultaneously refer to the original task requirements and the generated context content, meeting the logical coherence requirements for generating power repair plans (e.g., after "Remove", characters such as "old" and "damaged" that modify the coil should be connected preferentially, rather than irrelevant characters).

[0038] Next, the combined input data is input into the baseline model to generate the initial character-level probability distribution at character position t. The combined input data at each character position t constructed above is input into the baseline model in the power field through the API interface. The baseline model calculates the character-level probability based on its pre-trained inference logic in the power field (such as the sequence of maintenance steps, collocation rules of professional terms, etc.): for character position t, the model will output the initial probability value of each character in the power field character set (including Chinese characters, numbers, punctuation marks, power professional symbols such as "#", "℃", etc., about 5,000 characters in total) at this position. Still taking t = 10 as an example, the baseline model combines the context logic after "dismantle" and the fault scenario of "shunt coil burned out", and calculates that the initial probability value of the character "old" is 0.89, the initial probability value of the character "bad" is 0.06, the initial probability value of the character "new" is 0.02, the initial probability value of the character "damaged" is 0.01, and the sum of the initial probability values of the remaining characters (such as "electricity", "water", etc.) is 0.02. These probability values together constitute the initial character-level probability distribution at character position t = 10. For other positions in the character sequence, such as t = 15 (corresponding to the character "install", in the sentence "install a new coil"), the baseline model calculates that the initial probability value of "install" is 0.92, the initial probability value of "mount" is 0.05 (because "install" is a fixed collocation, the probability of "mount" following "install" needs to be calculated at subsequent positions), and the sum of the probabilities of the remaining characters is 0.03, forming the initial probability distribution at this position.

[0039] Subsequently, based on the preset correction factor library for power field characters, the initial character-level probability distribution is calibrated. This correction factor library is constructed based on power industry standards, maintenance regulations, and historical high-quality maintenance cases, and contains the correction factor β of each character in the character set under different power task scenarios (such as circuit breaker maintenance, transformer maintenance, dispatching order generation). Its core function is to strengthen the probability weight of power professional term characters and suppress the interference of non-professional characters. For example, in the "equipment maintenance" scenario, the correction factors β of professional term characters such as "brake", "coil", "insulation", and "voltage withstand" are 1.3, the correction factors β of general characters such as "of", "already", and "in" are 0.9, and the correction factors β of irrelevant characters such as "flower", "bird", and "car" are 0.1. For punctuation marks, the correction factors β of commonly used ones such as ";", "1.", "2." in the maintenance plan are 1.2, and the correction factors β of special symbols such as "@" and "&" are 0.05. During the calibration process, for each character y in the initial probability distribution at each character position t, the calibrated probability value is calculated through the formula where π t (y) is the initial probability value of character y, β y is the correction factor corresponding to character y, and ∑ y′ π t (y′)·βy′ It is the sum of the probabilities of all characters weighted by the correction factor (used for normalization to ensure that the sum of the probabilities of all characters after calibration is 1). Still taking t = 10 as an example, the initial probability of the character "旧" is 0.89, and the corresponding correction factor β = 1.1 ("旧" is often used to modify the parts to be replaced in the maintenance scenario, such as "旧线圈", "旧零件"), then the weighted probability is 0.89×1.1 = 0.979; the initial probability of the character "坏" is 0.06, the correction factor β = 1.2, and the weighted probability is 0.06×1.2 = 0.072; the sum of the weighted probabilities of the remaining characters is approximately 0.021. Divide these weighted probabilities by the sum (0.979 + 0.072 + 0.021 = 1.072), and the probability value of "旧" after calibration is 0.979 / 1.072≈0.914, the probability value of "坏" is 0.072 / 1.072≈0.067, and the sum of the probabilities of the remaining characters≈0.019, completing the calibration of the initial probability distribution at this position. Through this step, the domain adaptability of the probability distribution output by the baseline model is further improved, ensuring that the probability weights of professional characters at key positions are more in line with the requirements of power tasks.

[0040] Finally, integrate the calibrated probability distributions of all character positions to obtain the character-by-character probability distribution data corresponding to the candidate answer data. After sequentially completing the generation and calibration processing of the initial probability distributions of all 48 character positions in the ordered character sequence, integrate the calibrated probability distributions of each position in the order of character positions (from t = 1 to t = 48) to form the complete character-by-character probability distribution data corresponding to candidate answer 1. This data is stored in JSON format, and each record contains the character position t, the character at this position, and the calibrated probability value of this character at this position, along with the probability rankings of other characters (such as t = 10: {"position": 10, "char": "旧", "probability": 0.914, "other_chars": [{"char": "坏", "probability": 0.067}, {"char": "损", "probability": 0.012}, …]}). For other candidate answer data (such as candidate answer 2, candidate answer 3) generated by the model to be trained, repeat the above steps of character splitting, joint input construction, initial probability calculation, calibration processing, and integration to obtain the corresponding character-by-character probability distribution data respectively.

[0041] Step 130: Perform compliance verification on each candidate answer data based on a preset discriminant model to obtain the binary score of each candidate answer data.

[0042] Perform compliance verification on each candidate answer data respectively based on a preset discrimination model to obtain binary scores. The discrimination model here is a classification model constructed based on the "Regulations for the Maintenance of Electrical Equipment" and the "Safety Operating Procedures for Substations", and is trained by labeled compliant / non-compliant maintenance step samples. It can perform binary determination (score r i ∈{0,1}, 1 indicates compliance and correctness, and 0 indicates non-compliance and error) on the semantic correctness and compliance of the candidate answer according to discrimination bases such as domain regulations and standard answers. After inputting the above 3 candidate answers into the discrimination model, the discrimination results are as follows: The score of candidate answer 1 (including key steps, compliant) is 1; the score of candidate answer 2 (missing the core inspection link, with potential safety hazards) is 0; the score of candidate answer 3 (the steps are complete and in line with the regulations) is 1.

[0043] Step 140: Update the parameters of the model to be trained through a dynamic constraint mechanism based on the binary scores and character-by-character probability distribution data.

[0044] After completing the character-by-character probability distribution analysis and binary score verification of the candidate answer data, enter the model parameter update stage, that is, iteratively optimize the network parameters of the current model to be trained through a dynamic constraint mechanism. For different situations of binary scores, different parameter update logics are adopted respectively.

[0045] When the binary score is the first preset value (here the first preset value is 1, corresponding to candidate answer 1 and candidate answer 3 which are compliant and correct), construct target distribution data based on the character-by-character probability distribution data output by the model to be trained and the benchmark model, and update the parameters of the model to be trained with this target distribution data as the optimization benchmark. Specifically, first extract the probability distribution data output by the model to be trained for each character position in each candidate answer, and at the same time extract the character-by-character probability distribution data output by the benchmark model for the same character position. Taking the position t = 10 (corresponding to the character "leak") in candidate answer 3 as an example, the probability of the character "leak" output by the model to be trained at this position is 0.75, the probability of "leakage" is 0.15, and the sum of the probabilities of other characters is 0.10. The probability of the character "leak" output by the benchmark model at this position is 0.91, the probability of "leakage" is 0.06, and the sum of the probabilities of other characters is 0.03. Select the maximum value of the two probability distributions at each character position as the target probability value at that character position, that is, the target probability value of "leak" is taken as max(0.75, 0.91) = 0.91, the target probability value of "leakage" is taken as max(0.15, 0.06) = 0.15, and the target probability values of other characters are taken as the maximum value of the corresponding probabilities of the two, forming the initial target distribution at that character position. Subsequently, perform normalization processing on the initial target distribution to ensure that the sum of the target probability values of all characters is 1, and obtain the target distribution data q i,t (y∣y i,<t ,x).

[0046] Next, calculate the loss value between the output of the model to be trained and the target distribution data through the loss function. The loss function is where T is the total length of the character sequence of the candidate answer (T = 45 for candidate answer 3), t is the character position index (1 ≤ t ≤ 45), y is a single character in the character set of the power field (including Chinese characters, numbers, punctuation marks, etc.), q i,t is the normalized target distribution data, and π s is the probability distribution data output by the model to be trained. After calculating the loss values for all character positions of candidate answer 1 and candidate answer 3, adjust the network parameters of the model to be trained (including attention weights, fully connected layer weights, etc.) through the backpropagation algorithm. After 100 rounds of iterative training, the loss value converges to 0.08 (the preset convergence condition is that the loss value ≤ 0.1), and the parameter update corresponding to the compliant samples in this round is completed. This process realizes the self-enhancement of knowledge by strengthening the learning of the correct character sequence in high-quality candidate answers by the model to be trained.

[0047] When the binary score is the second preset value (here the second preset value is 0, corresponding to the non-compliant candidate answer 2), determine the difference between the character-level probability distribution data output by the model to be trained and the benchmark model, and update the parameters of the model to be trained based on this difference and the probability distribution data of the benchmark model. First, extract the character-level probability distribution data output by the model to be trained for each character position in candidate answer 2, and synchronously extract the probability distribution data output by the benchmark model for the same character position. Taking the position t = 6 (corresponding to the character "补") in candidate answer 2 as an example, the probability of the character "补" output by the model to be trained is 0.88, the probability of the character "检" is 0.05, and the sum of the probabilities of other characters is 0.07. In the probability distribution of the benchmark model at this position, the probability of the character "检" is 0.90, the probability of the character "补" is 0.06, and the sum of the probabilities of other characters is 0.04, and there is a significant difference between the two. Calculate this difference through the relative entropy (KL divergence). The KL divergence formula is The KL divergence value calculated for this character position is 1.23, that is, the difference quantization value.

[0048] The probability distribution data based on the baseline model is used as the optimization baseline. Combining this difference quantization value, a relative entropy loss function is constructed, and the sum of the KL divergences at all character positions of candidate answer 2 is used as the total loss value. After calculating the difference loss between the output of the model to be trained and the output of the baseline model through this loss function, the network parameters of the model to be trained are also adjusted through the backpropagation algorithm, with a focus on correcting the inference logic that leads to the omission of key steps in the answer. For example, at the position of t = 6, by adjusting the parameters, the probability weight of the output "repair" of the model to be trained is reduced, and the probability weight of "inspection" is increased, enabling the model to learn the necessity of the "inspection" link in the "oil level abnormal" scenario. After 80 rounds of training iterations in this round, the loss value converges to 0.12 (meeting the preset convergence condition), and the parameter correction corresponding to the non-compliant samples is completed.

[0049] Through the above cyclic iterative training (selecting 100 power task data in each round of training, generating 3 candidate answers for each data, and continuously training for 50 rounds), the response time of the finally obtained power task execution model on the edge computing device is stable at about 450 ms, the device resource occupancy rate is less than 60%, the answer accuracy rate for the maintenance work order query task reaches 92%, the accuracy rate is increased by 15% compared with the model trained by the traditional hard distillation method, and the response speed is increased by 60% compared with the quantized large model, fully meeting the deployment and application requirements of the substation edge node.

[0050] In other power task scenarios, such as the power dispatch instruction generation task, the power task data is the load adjustment requirement issued by the dispatch center (such as "the transformer tap adjustment instruction when the regional load increases by 10%"). The model to be trained generates multiple candidate texts of dispatch instructions, the baseline model (a large model dedicated to power dispatch) provides the character-level probability distribution, and the discriminant model performs compliance verification according to the "Power System Dispatching Regulations" (such as whether the instruction format is standard and whether the adjustment range meets the safety threshold), and then updates the model parameters through the dynamic constraint mechanism. The finally trained model can quickly generate compliant and accurate dispatch instructions, effectively supporting the real-time dispatching decision-making of the power grid. The processing logics of the above various power tasks can all be passed through Figure 2The diagram clearly outlines the complete closed loop of "task input - candidate generation - probability prediction - compliance judgment - parameter update": When a power task is given as input, the small-sized training model first autonomously generates multiple candidate answers (including correct and incorrect samples) based on its own reasoning strategy. The large-sized baseline model outputs the corresponding predicted probability for each candidate answer generation process. At the same time, reference materials (such as domain regulations and standard answers) provide the judgment basis for the discriminant model. The discriminant model performs binary labeling (distinguishing between right and wrong) on ​​the semantic correctness and compliance of each candidate answer. Finally, the loss is calculated by combining the judgment result and the probability distribution output by the baseline model. The parameters of the small-sized model are optimized through backpropagation of the loss, realizing the efficient transfer and accurate adaptation of domain knowledge from the large model to the lightweight model. This intuitively demonstrates the universality and core execution logic of the proposed method in different power task scenarios.

[0051] It is understood that, in order to achieve the functions in the above embodiments, the computer device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and method steps described in conjunction with the embodiments disclosed in this application, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application scenario and design constraints of the technical solution.

[0052] Furthermore, as a response to the above Figure 1 The implementation of the method embodiment shown in this application provides a power task processing device. The embodiment of this device corresponds to the foregoing method embodiments. For ease of reading, this embodiment will not repeat the details of the foregoing method embodiments one by one, but it should be clear that the device in this embodiment can correspondingly implement all the contents of the foregoing method embodiments. Specifically, as shown... Figure 3 As shown, the power task processing device 300 includes:

[0053] The processing module 310 is used to acquire power task data and input the power task data into the model to be trained, so that the model to be trained can process the power task data according to its own reasoning strategy and generate multiple candidate answer data for the power task data.

[0054] The parsing module 320 is used to parse each candidate answer data according to a preset benchmark model to obtain the character-level probability distribution data corresponding to each candidate answer data.

[0055] The verification module 330 is used to perform compliance verification on each candidate answer data based on a preset discrimination model, and obtain a binary score for each candidate answer data.

[0056] The update module 340 is used to update the parameters of the model to be trained based on binary scores and character-by-character probability distribution data through a dynamic constraint mechanism. The model to be trained after parameter update serves as a power task execution model, which is used to process the input power task data on the edge computing device and output the task execution results.

[0057] Furthermore, such as Figure 3 As shown, the update module 340 is specifically used to construct target distribution data based on the character-by-character probability distribution data output by the model to be trained and the benchmark model if the binary score is the first preset value, and update the parameters of the model to be trained using the target distribution data as the optimization benchmark.

[0058] Furthermore, such as Figure 3 As shown, the update module 340 is specifically used to extract the probability distribution data output by the model to be trained for each character position in the candidate answer data, and to extract the probability distribution data output by the benchmark model for the same character position in the same candidate answer data; select the maximum value of the probability distribution data output by the benchmark model and the model to be trained at each character position as the target probability value of the character position, integrate the target probability values ​​of all characters at the same character position to form the initial target distribution of the character position; normalize the initial target distribution to obtain the target distribution data that meets the probability distribution requirements; and through... Calculate the loss value (loss) between the character-by-character probability distribution data output by the model to be trained and the target distribution data, where T represents the total length of the character sequence of the candidate response data, t is the character position index in the candidate response data, and y is a single character in the character set. In the normalized target distribution data, for the i-th candidate answer data and the t-th character position, the character sequence y preceding the t-th character is used. i,＜t And the probability value of power task data x as a condition and character y as π s (y∣y i,<t (x) represents the sequence of characters y preceding the t-th character in the i-th candidate answer data and at the t-th character position, which is the model to be trained. i,<t Given power task data x as a condition and character y as a probability value; based on the loss value, adjust the network parameters of the model to be trained using the backpropagation algorithm until the loss value meets the preset convergence condition, thus completing the update of the model parameters.

[0059] Furthermore, such as Figure 3 As shown, the update module 340 is specifically used to determine the difference between the character-by-character probability distribution data output by the model to be trained and the benchmark model if the binary score is the second preset value, and update the parameters of the model to be trained based on the difference and the character-by-character probability distribution data output by the benchmark model.

[0060] Furthermore, such as Figure 3 As shown, the update module 340 is specifically used to extract the character-by-character probability distribution data output by the model under test for each character position in the candidate answer data, and simultaneously extract the character-by-character probability distribution data output by the benchmark model for the same character position in the same candidate answer data; calculate the difference between the character-by-character probability distribution data output by the model under test and the benchmark model at the same character position using relative entropy, and obtain the difference quantification value; construct a relative entropy loss function based on the character-by-character probability distribution data output by the benchmark model to optimize the benchmark and the difference quantification value; calculate the loss value between the character-by-character probability distribution data output by the model under test and the benchmark model using the relative entropy loss function; and adjust the network parameters of the model under test using the backpropagation algorithm based on the loss value until the loss value meets the preset convergence condition, thus completing the update of the parameters of the model under test.

[0061] Furthermore, such as Figure 3 As shown, the parsing module 320 is specifically used to perform character sequence splitting on the candidate answer data, obtaining an ordered character sequence with individual characters as units. The ordered character sequence is denoted as (y i,1 ,y i,2 ,…,y i,T The input data is defined as follows: i represents the index of the candidate response data, T represents the total length of the character sequence corresponding to the candidate response data, and the original power task data x corresponding to the candidate response data is also stored. For each character position t in the ordered character sequence, the original power task data x is combined with the prefix character sequence before character position t to form the joint input data for character position t. The joint input data is input into the baseline model, which performs character-level probability calculations on the joint input data based on its own domain inference logic, outputting the initial probability values ​​corresponding to all characters in the character set at character position t, forming the initial character-level probability distribution for character position t. Based on a preset power domain character correction factor library, the correction factor β corresponding to each character is extracted and calculated using the formula... The initial character-level probability distribution is calibrated, where π t (y) represents the probability value of character y in the initial character-level probability distribution, π t,cal (y) represents the probability value of the calibrated character y. After generating and calibrating the initial character-level probability distributions for all character positions in the ordered character sequence, the calibrated probability distributions corresponding to all character positions are integrated in the order of the character sequence to obtain the character-level probability distribution data corresponding to the candidate answer data.

[0062] Furthermore, such as Figure 3As shown, the processing module 310 is specifically used to acquire multi-source heterogeneous raw data in the power field. The raw data includes, but is not limited to, structured equipment parameter data, semi-structured maintenance work order text, unstructured dispatch voice transcription text, and power system alarm logs. The raw data undergoes unified format conversion and noise filtering to remove invalid characters and redundant information. It is then categorized and labeled according to power task type to form a basic task dataset. For each power task data item in the basic task dataset, core entities are extracted using entity recognition algorithms. Attribute information and associated entities related to the core entities are queried from the domain knowledge graph. Based on the core entities, associated attribute information, and associated entities, the expression logic is reconstructed to generate extended task data that is semantically equivalent to the original power task data. A power scenario disturbance factor is introduced to dynamically adjust non-core fields in the extended task data. Finally, the basic task data and the adjusted extended task data are merged to form the final power task dataset.

[0063] Optionally, the power task processing device may be an electronic device with data processing capabilities, or a functional module within the electronic device, without limitation.

[0064] For example, the electronic device can be a server, which can be a single server or a server cluster consisting of multiple servers. As another example, the electronic device can be a mobile phone, tablet computer, desktop computer, laptop computer, handheld computer, notebook computer, ultra-mobile personal computer (UMPC), netbook, as well as cellular phones, personal digital assistants (PDAs), augmented reality (AR) devices, virtual reality (VR) devices, and other terminal devices. As yet another example, the electronic device can also be a recording device, video surveillance equipment, etc. This application does not impose any special limitations on the specific form of the electronic device.

[0065] The following example uses an electronic device as an example of a power task processing device. Figure 4 As shown, Figure 4 The hardware structure of an electronic device 400 provided in this application.

[0066] like Figure 4 As shown, the electronic device 400 includes a processor 410, a communication line 420, and a communication interface 430.

[0067] Optionally, the electronic device 400 may also include a memory 440. The processor 410, memory 440, and communication interface 430 can be connected via a communication line 420.

[0068] The processor 410 can be a central processing unit (CPU), a general-purpose processor, a network processor (NP), a digital signal processor (DSP), a microprocessor, a microcontroller, a programmable logic device (PLD), or any combination thereof. The processor 410 can also be any other device with processing capabilities, such as a circuit, device, or software module, without limitation.

[0069] In one example, processor 410 may include one or more CPUs, for example Figure 4 CPU0 and CPU1 in the CPU.

[0070] As an optional implementation, electronic device 400 may include multiple processors, for example, in addition to processor 410, it may also include processor 470. Communication line 420 is used to transmit information between the components included in electronic device 400.

[0071] Communication interface 430 is used for communicating with other devices or other communication networks. These other communication networks can be Ethernet, Radio Access Network (RAN), Wireless Local Area Networks (WLAN), etc. Communication interface 430 can be a module, circuit, transceiver, or any device capable of enabling communication.

[0072] Memory 440 is used to store instructions. These instructions can be computer programs.

[0073] The memory 440 can be a read-only memory (ROM) or other type of static storage device that can store static information and / or instructions; it can also be a random access memory (RAM) or other type of dynamic storage device that can store information and / or instructions; it can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, etc., without limitation.

[0074] It should be noted that the memory 440 can exist independently of the processor 410, or it can be integrated with the processor 410. The memory 440 can be used to store instructions, program code, or some data, etc. The memory 440 can be located inside or outside the electronic device 400, without restriction.

[0075] The processor 410 is configured to execute instructions stored in the memory 440 to implement the communication method provided in the following embodiments of this application. For example, when the electronic device 400 is a terminal or a chip in a terminal, the processor 410 can execute instructions stored in the memory 440 to implement the steps performed by the transmitting end in the following embodiments of this application.

[0076] As an optional implementation, the electronic device 400 also includes an output device 450 and an input device 460. The output device 450 can be a display screen, speaker, or other device capable of outputting data from the electronic device 400 to the user. The input device 460 can be a keyboard, mouse, microphone, joystick, or other device capable of inputting data into the electronic device 400.

[0077] It should be pointed out that, Figure 4 The structure shown does not constitute a limitation on the electronic device, except... Figure 4 In addition to the components shown, the electronic device may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0078] The power task processing device and application scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of power task processing devices and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0079] This application provides a storage medium storing a program that, when executed by a processor, implements the power task processing method.

[0080] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for processing power tasks, characterized in that, The method includes: Acquire power task data and input the power task data into the model to be trained, so that the model to be trained can process the power task data according to its own reasoning strategy and generate multiple candidate answer data for the power task data. Based on a preset benchmark model, each candidate answer data is analyzed to obtain character-level probability distribution data corresponding to each candidate answer data. Based on a preset discrimination model, each candidate answer data is subjected to compliance verification to obtain a binary score for each candidate answer data. Based on the binary score and the character-by-character probability distribution data, the parameters of the model to be trained are updated through a dynamic constraint mechanism. The model to be trained after parameter update serves as a power task execution model, which is used to process the input power task data on an edge computing device and output the task execution results.

2. The method according to claim 1, characterized in that, Based on the binary score and the character-by-character probability distribution data, the parameters of the model to be trained are updated through a dynamic constraint mechanism, including: If the binary score is a first preset value, target distribution data is constructed based on the character-by-character probability distribution data output by the model to be trained and the benchmark model, and the parameters of the model to be trained are updated using the target distribution data as the optimization benchmark.

3. The method according to claim 2, characterized in that, Based on the character-by-character probability distribution data output by the model to be trained and the benchmark model, target distribution data is constructed. The parameters of the model to be trained are then updated using the target distribution data as an optimization benchmark, including: Extract the probability distribution data output by the model to be trained for each character position in the candidate answer data, and extract the probability distribution data output by the benchmark model for the same character position in the same candidate answer data; The maximum value of the probability distribution data output by the baseline model and the model to be trained at each character position is selected as the target probability value of the character position. The target probability values ​​of all characters at the same character position are integrated to form the initial target distribution of the character position. The initial target distribution is normalized to obtain target distribution data that meets the probability distribution requirements; pass Calculate the loss value (loss) between the character-by-character probability distribution data output by the model to be trained and the target distribution data, where T represents the total length of the character sequence of the candidate response data, t is the character position index in the candidate response data, and y is a single character in the character set. In the normalized target distribution data, for the i-th candidate answer data and the t-th character position, the character sequence y preceding the t-th character is used. i,＜t And the probability value of power task data x as a condition and character y as π s (y∣y i,<t (x) represents the sequence of characters y preceding the t-th character in the i-th candidate answer data and at the t-th character position, which is the model to be trained. i,<t The probability value of power task data x as a condition and character y; Based on the loss value, the network parameters of the model to be trained are adjusted through the backpropagation algorithm until the loss value meets the preset convergence condition, thus completing the update of the parameters of the model to be trained.

4. The method according to claim 1, characterized in that, Based on the binary score and the character-by-character probability distribution data, the parameters of the model to be trained are updated through a dynamic constraint mechanism, which further includes: If the binary score is a second preset value, the difference between the character-by-character probability distribution data output by the model to be trained and the benchmark model is determined, and the parameters of the model to be trained are updated based on the difference and the character-by-character probability distribution data output by the benchmark model.

5. The method according to claim 4, characterized in that, Determine the difference between the character-by-character probability distribution data output by the model to be trained and the benchmark model, and update the parameters of the model to be trained based on the difference and the character-by-character probability distribution data output by the benchmark model, including: Extract the character-by-character probability distribution data output by the model to be trained for each character position in the candidate answer data, and simultaneously extract the character-by-character probability distribution data output by the benchmark model for the same character position in the same candidate answer data. The difference between the probability distribution data of the model to be trained and the benchmark model at the same character position is calculated by relative entropy, and the difference quantification value is obtained. Based on the character-by-character probability distribution data output by the benchmark model, a relative entropy loss function is constructed to optimize the benchmark and the difference quantification value; The loss value between the character-by-character probability distribution data output by the model to be trained and the character-by-character probability distribution data output by the benchmark model is calculated using the relative entropy loss function. Based on the loss value, the network parameters of the model to be trained are adjusted through the backpropagation algorithm until the loss value meets the preset convergence condition, thus completing the update of the parameters of the model to be trained.

6. The method according to claim 1, characterized in that, Based on a preset benchmark model, each candidate answer data is parsed to obtain character-by-character probability distribution data corresponding to each candidate answer data, including: The candidate answer data is split into character sequences to obtain an ordered character sequence consisting of individual characters, denoted as (y i,1 ,y i,2 ,…,y i,T ), where i represents the index of the candidate answer data, T represents the total length of the character sequence corresponding to the candidate answer data, and the original power task data x corresponding to the candidate answer data is also stored in association; For each character position t in the ordered character sequence, the original power task data x is combined with the prefix character sequence before character position t to form the joint input data for character position t; The joint input data is input into the baseline model so that the baseline model can perform character-level probability calculation on the joint input data according to its own domain reasoning logic, and output the initial probability values ​​corresponding to all characters in the character set at character position t, forming the initial character-level probability distribution at character position t. Based on a pre-defined character correction factor library for the power industry, the correction factor β corresponding to each character is extracted, and then calculated using a formula. The initial character-level probability distribution is calibrated, where π t (y) represents the probability value of character y in the initial character-level probability distribution, π t,cal (y) represents the probability value of the calibrated character y; After generating and calibrating the initial character-level probability distributions for all character positions in the ordered character sequence, the calibrated probability distributions corresponding to all character positions are integrated in the order of the character sequence to obtain the character-level probability distribution data corresponding to the candidate answer data.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: Acquire multi-source heterogeneous raw data in the power sector, including but not limited to structured equipment parameter data, semi-structured maintenance work order text, unstructured dispatch voice transcription text, and power system alarm logs; The raw data is subjected to uniform format conversion and noise filtering to remove invalid characters and redundant information. It is then classified and labeled according to power task type to form a basic task dataset. For each power task data in the basic task dataset, the core entity in the power task data is extracted by an entity recognition algorithm. The attribute information and related entities associated with the core entity are queried from the domain knowledge graph. Based on the core entity, the associated attribute information and related entities, the expression logic is reconstructed to generate extended task data that is semantically equivalent to the original power task data. A power scenario disturbance factor is introduced to dynamically adjust non-core fields in the extended task data; The basic task data is merged with the adjusted extended task data to form the final power task dataset.

8. A power task processing device, characterized in that, The device includes: The processing module is used to acquire power task data and input the power task data into the model to be trained, so that the model to be trained can process the power task data according to its own reasoning strategy and generate multiple candidate answer data for the power task data. The parsing module is used to parse each candidate answer data according to a preset benchmark model to obtain character-level probability distribution data corresponding to each candidate answer data. The verification module is used to perform compliance verification on each candidate answer data based on a preset discrimination model, and obtain a binary score for each candidate answer data. The update module is used to update the parameters of the model to be trained based on the binary score and the character-by-character probability distribution data through a dynamic constraint mechanism. The model to be trained after parameter update serves as a power task execution model, which is used to process the input power task data on the edge computing device and output the task execution results.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the storage medium to perform the power task processing method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the power task processing method as described in any one of claims 1-7.

Images