Large model-based electric power operation problem solving method and system, and electronic equipment

Abstract

The invention relates to the technical field of large models, and particularly discloses an electric power operation problem solving method and system based on a large model and electronic device.An electric power field knowledge base is constructed by collecting enterprise multi-source data, and a knowledge enhancement type electric power large language model is formed; the method comprises the following steps: acquiring original field operation data, generating a fault feature value, and fusing the fault feature value with equipment basic information, historical operation data and environmental parameters to form a complete fault description data set; by implementing a hierarchical diagnosis mechanism of end-side cloud cooperation, fault type identification is realized, a diagnosis result is output, a model is called to generate a multi-dimensional solution, and an optimal execution scheme is determined through machine learning scoring model evaluation; and an execution result, actual time consumption, a fault type and a final scheme are automatically collected as feedback information for updating and optimizing the model, so that continuous learning is realized. According to the invention, the fault identification accuracy and response speed can be improved, the solution selection is optimized, and the power operation and maintenance efficiency and safety are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large models, more particularly, the present application relates to a large model-based power operation problem solving method and system and electronic equipment. BACKGROUND

[0002] In the daily operation and maintenance of the power industry, field technicians often face various complex equipment failures and technical problems. The traditional problem solving method mainly relies on the personal experience of technicians, paper manual query and remote expert telephone guidance. This method has many limitations. First, the number of experienced technical experts is limited, and they cannot respond to all on-site needs in a timely manner, resulting in delayed fault handling. Second, paper manuals are updated slowly, and the query efficiency is low, and it is difficult to cover all possible fault scenarios. Third, telephone communication has the problem of inaccurate information transmission, and experts cannot intuitively understand the on-site situation, and the suggestions given may deviate from the actual situation. Finally, a large amount of valuable fault handling experience is scattered in the minds of various technicians, making it difficult to form systematic knowledge accumulation and inheritance. These problems seriously hinder the improvement of power system operation and maintenance efficiency, and increase the economic loss and safety risk caused by equipment failure.

[0003] Therefore, it is necessary to provide a large model-based power operation problem solving method, system and electronic equipment to solve the above technical problems. In order to solve the above problems, a technical solution is provided. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a large model-based power operation problem solving method, system and electronic equipment, which is used to solve the problem that the existing power operation relies on experience, paper manual and telephone guidance, resulting in slow response, inaccurate information transmission, difficulty in systematizing fault experience, limiting operation and maintenance efficiency, and increasing the risk of equipment failure and economic loss.

[0005] To achieve the above purpose, the present application provides the following technical solution: The large model-based power operation problem solving method comprises the following steps: Generating a power field knowledge base by collecting multi-source data of power enterprises, training an industry large language model based on the power field knowledge base, and outputting a knowledge-enhanced power large language model; Deploying an intelligent acquisition terminal at the operation site, integrating multiple sensing devices to obtain raw field operation data, pre-processing and feature extraction of the raw field operation data based on edge computing nodes, outputting fault feature values, and fusing the fault feature values with device basic information, historical operation data and environmental parameters to form a fault description data set; Based on the knowledge enhanced power large language model, a hierarchical diagnosis mechanism of end-edge-cloud collaboration is implemented to identify the fault type and output the fault identification result; According to the fault identification result, the knowledge enhanced power large language model is called to generate a multi-dimensional solution, the multi-dimensional solution is scored through a scoring model based on machine learning, and the final solution is determined; Based on the final solution, the execution result is determined, the execution result, actual time consumption, fault type and final solution are automatically collected as feedback information, and the knowledge enhanced power large language model is updated and optimized.

[0006] As a further scheme of the present application, an intelligent acquisition terminal is deployed at the work site to integrate multiple sensing devices to obtain original field work data, wherein the multiple sensing devices include a high-definition camera, an infrared sensor, a vibration sensor and a sound sensor, and the original field work data includes device appearance image, temperature distribution data, mechanical state data and noise monitoring data.

[0007] As a further scheme of the present application, the original field work data is preprocessed and feature extracted by the edge computing node to output a work structured feature vector, and the specific steps are as follows: The original field work data is preprocessed by the edge computing module embedded in the intelligent acquisition terminal; Set the sampling frequency , the original field work data in the time window is continuously sampled to obtain an original time series data matrix , wherein is the sampling value of the i-th sensing device at time t, is the sampling value of the n-th sensing device at time t, is the number of sensing devices; Based on the original time series data matrix, the sampling features of each sensing device are obtained, including sampling mean, sampling variance and sampling peak, and the fault feature vector of the sensing device is obtained by splicing the sampling features , wherein is the sampling mean of the i-th sensing device, is the sampling variance of the i-th sensing device, is the sampling peak of the i-th sensing device; The fault feature vectors of each sensing device are normalized to splice the work structured feature vector , wherein is the fault feature vector of the high-definition camera, is the fault feature vector of the infrared sensor, is the fault feature vector of the vibration sensor, a fault feature vector of the sound sensor; perform feature fusion based on the job structured feature vector to obtain a fault feature value.

[0008] As a further scheme of the present application, based on the knowledge enhanced power large language model, a hierarchical diagnosis mechanism of edge-cloud collaboration is implemented to identify the fault type and output the fault identification result, and the specific steps are as follows: Based on the knowledge distillation of the knowledge enhanced power large language model, an edge diagnosis model is obtained, and based on the edge diagnosis model, primary fault identification is performed. According to the primary identification result, the fault type is determined, and the fault type includes a standard fault type and a rare fault type. For the rare fault type, secondary fault identification is performed through the knowledge enhanced power large language model to identify the fault root cause and associated impact.

[0009] As a further scheme of the present application, based on the edge diagnosis model, primary fault identification is performed, and the specific steps include: According to the historical fault description data set, the edge diagnosis model is trained, and the real-time acquired fault description data set is input for primary fault identification. If the edge diagnosis model can match the same data set in the historical fault description data set, the device fault type is determined as a standard fault type. If the edge diagnosis model cannot match the same data set in the historical fault description data set, the device fault type is determined as a rare fault type.

[0010] As a further scheme of the present application, for the rare fault type, secondary fault identification is performed through the knowledge enhanced power large language model to identify the fault root cause and associated impact, and the specific steps are as follows: Obtain the fault description data set corresponding to the rare fault type, upload the fault feature vector, preliminary diagnosis result and confidence to the cloud end through the edge end; Perform secondary fault identification through the knowledge enhanced power large language model deployed in the cloud end, and perform deep analysis combining the historical cases of the whole network, real-time operation data and weather information to identify the fault root cause and associated impact.

[0011] As a further scheme of the present application, according to the fault identification result, the knowledge enhanced power large language model is called to generate a multi-dimensional solution, and a scoring model based on machine learning is constructed to score the multi-dimensional solution to determine the final solution, and the specific steps are as follows: Based on the historical successful solution cases in the knowledge enhanced power large language model, a Monte Carlo tree search algorithm is used to explore the multi-dimensional solution; A mapping relationship between the case features and the success rate is learned by constructing a scoring model based on machine learning by taking historical successful solution cases as training samples, and the multi-dimensional solutions are scored; The multi-dimensional solutions are sorted according to the scoring results, the multi-dimensional solution ranked first is taken as the final solution, and the multi-dimensional solution ranked second is taken as the alternative solution.

[0012] As a further scheme of the present application, the execution result is determined based on the final solution, the execution result, the actual time consumption, the fault type and the final solution are automatically collected as feedback information, the knowledge-enhanced power large language model is updated and optimized, and the specific steps are as follows: By automatically collecting the execution result, the actual time consumption, the fault type and the final solution as feedback information, a solution exceeding an expected difference threshold is identified through an active learning algorithm, triggering expert review and labeling to form new training samples; An incremental learning technology is adopted to integrate the new samples into the training process of the knowledge-enhanced power large language model, and the parameters and knowledge graph of the model are updated, and a causal reasoning analysis method is used to extract fault causal relationships from the feedback information.

[0013] The power operation problem solving system based on a large model comprises a power knowledge construction module, a multi-source data acquisition and feature fusion module, an end-edge-cloud collaborative hierarchical diagnosis module, a multi-dimensional solution generation and scoring module and a model self-optimization feedback module. The power knowledge construction module is used to generate a power domain knowledge base by collecting multi-source data of power enterprises, train an industrialized large language model based on the power domain knowledge base, and output a knowledge-enhanced power large language model. The multi-source data acquisition and feature fusion module is used to integrate multiple sensing devices to obtain original field operation data by deploying intelligent acquisition terminals on the operation site, pre-process and extract features from the original field operation data based on edge computing nodes, output fault feature values, and fuse the fault feature values with device basic information, historical operation data and environmental parameters to form a fault description data set. The end-edge-cloud collaborative hierarchical diagnosis module is used to implement an end-edge-cloud collaborative hierarchical diagnosis mechanism based on the knowledge-enhanced power large language model, identify the fault type, and output the fault identification result. The multi-dimensional solution generation and scoring module is used to call the knowledge-enhanced power large language model according to the fault identification result, generate multi-dimensional solutions, score the multi-dimensional solutions by constructing a scoring model based on machine learning, and determine the final solution. The model self-optimization feedback module is used for determining an execution result based on a final solution, automatically collecting the execution result, actual time consumption, fault type and final solution as feedback information, and updating and optimizing the knowledge-enhanced power large language model.

[0014] An electronic device comprises a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes the steps of the power operation problem solving method based on a large model as claimed in any one of the above.

[0015] The technical effects and advantages of the power operation problem solving method, system and electronic device based on a large model of the present application are as follows: the present application constructs a power field knowledge base by collecting enterprise multi-source data, trains an industry large language model based thereon, and forms a knowledge-enhanced power large language model; in the operation site, an intelligent acquisition terminal is deployed, multi-sensors are integrated to obtain original data, and pre-processing and feature extraction are performed through edge computing to generate fault feature values, which are then fused with device basic information, historical operation data and environmental parameters to form a complete fault description data set; based on the model, a hierarchical diagnosis mechanism of end-edge-cloud collaboration is implemented to realize fault type identification and output diagnosis results; then, a multi-dimensional solution is generated by calling the model, and the optimal execution scheme is determined through the evaluation of a machine learning scoring model; the execution result, actual time consumption, fault type and final solution are automatically collected as feedback information for updating and optimizing the model to realize continuous learning. The present application can improve fault identification accuracy and response speed, optimize solution selection, reduce dependence on experts, accumulate systematic operation and maintenance knowledge, and significantly improve power operation and maintenance efficiency and safety.

[0016] Through multi-source data fusion and fault feature extraction, combined with the knowledge-enhanced power large language model, the present application can accurately identify complex device fault types and reduce human misjudgment; the hierarchical diagnosis mechanism of end-edge-cloud collaboration and the on-site intelligent acquisition terminal enable real-time collection and processing of fault information, shortening the time from fault occurrence to solution; the machine learning scoring model is used to evaluate multi-dimensional solutions to ensure that the final solution is efficient, safe and feasible; by automatically collecting execution results and feedback information and updating the model through incremental learning, the system can be continuously optimized to gradually form a systematic power operation and maintenance knowledge base; the combination of a large model and an automatic mechanism can provide scientific decision-making directly on site, alleviate the problem of limited expert resources, and make the overall process intelligent and data-driven, thereby reducing device fault risk and economic loss while ensuring the safety of operation and maintenance personnel. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A flowchart of the power operation problem solving method based on a large model provided by the embodiments of the present application is shown in the figure. Figure 2A system block diagram of the power operation problem solving system based on a large model provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0018] The technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described technical solutions are only a part of the present application, not the whole. Based on the technical solutions in the present application, all other technical solutions obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0019] As Figure 1 shown is a flowchart of the power operation problem solving method based on a large model provided by an embodiment of the present application, Figure 1 The execution subject of the method shown can be a software and / or hardware device. The execution subject of the present application can include but is not limited to at least one of the following: a user device, a network device, etc. The user device can include but is not limited to a computer, a smart phone, a personal digital assistant (PDA), and the above-mentioned electronic devices, etc. The network device can include but is not limited to a single network server, a server group composed of multiple network servers, or a cloud composed of a large number of computers or network servers based on cloud computing, wherein cloud computing is a kind of distributed computing, which is a super virtual computer composed of a loose-coupled computer group. The present embodiment does not make any limitation. It includes steps S1 to S5, which are as follows: S1, generating a power field knowledge base by collecting multi-source data of a power enterprise, training an industrial large language model based on the power field knowledge base, and outputting a knowledge-enhanced power large language model; S2, deploying an intelligent acquisition terminal at a work site, integrating multiple sensing devices to obtain original field operation data, pre-processing and feature extraction of the original field operation data according to an edge computing node, outputting fault feature values, fusing the fault feature values with device basic information, historical operation data, and environmental parameters, and forming a fault description data set; S3, implementing a hierarchical diagnosis mechanism based on the knowledge-enhanced power large language model, identifying a fault type, and outputting a fault identification result; S4, calling the knowledge-enhanced power large language model according to the fault identification result, generating a multi-dimensional solution, scoring the multi-dimensional solution by constructing a scoring model based on machine learning, and determining a final solution; S5, determining an execution result based on the final solution, automatically collecting the execution result, actual time consumption, fault type, and final solution as feedback information, and updating and optimizing the knowledge-enhanced power large language model.

[0020] Preferably, the industry large language model is trained based on the power field knowledge base, and the specific steps are as follows: Collect multi-source data of power enterprises, including work order data, equipment manuals, fault cases, industry standards accumulated by power enterprises over the years, and preprocess the multi-source data, extract fault phenomena, device basic information, historical operation data, environmental parameters, solutions, etc. from the multi-source data through natural language processing technology; Based on multi-source data, entity recognition, relationship extraction and attribute extraction output knowledge triple set, and generate knowledge graph according to knowledge triple set; The pre-trained industry large language model is adapted to the field by using parameter efficient fine-tuning technology, and the power professional knowledge is injected while maintaining the original ability of the industry large language model through low-rank matrix decomposition method, and the knowledge enhanced power large language model is output; Extract industry standard specifications in multi-source data into a logic rule base to provide rule constraints for subsequent fault diagnosis and solution verification.

[0021] It should be noted that based on multi-source data, entity recognition, relationship extraction and attribute extraction output knowledge triple set, wherein the entity recognition includes: device class, such as transformer, circuit breaker, bus, insulator, etc. ; state class, such as overheating, discharge, abnormal vibration; parameter class, such as temperature, current, voltage, frequency; event class, such as maintenance, trip, alarm, recovery. The relationship extraction adopts a relationship extraction model based on semantic dependency tree and attention mechanism to identify the relationship between entities, such as "circuit breaker", "cause", "short circuit fault". The attribute extraction extracts the feature attributes of each entity from the corpus, such as: voltage level, model, installation time, geographic location; fault probability, maintenance record.

[0022] The embodiment of the application takes a provincial power company as a demonstration object, and aims at the problems of scattered knowledge and high experience dependence in the process of device operation and maintenance and fault disposal. First, the multi-source heterogeneous data in the enterprise is centralized. The data sources include work order repair data accumulated for nearly ten years, equipment operation and maintenance manual, fault and maintenance case records over the years, industry standards and specifications, and environmental monitoring and meteorological data. Through the automatic data access interface, the above multi-source data is formatted, repeated items are cleaned and semantic unified coding is performed, and the natural language processing technology is used to perform entity recognition and information extraction on the text content. The device basic information (such as device type, rated voltage, installation date), running state parameters (such as temperature, current, vibration amplitude), fault phenomenon description, environmental factors and corresponding solutions contained therein are recognized, and a structured corpus set is formed.

[0023] On this basis, entity recognition, relation extraction and attribute extraction operations are performed on the corpus to generate a set of knowledge triples such as ("transformer", "occurrence", "temperature anomaly"), ("temperature anomaly", "possible cause", "cooling oil blockage"), ("cooling oil blockage", "solution", "cleaning of cooling circuit") and the like, and a power domain knowledge graph is constructed based on the set to realize the association modeling between devices, faults, causes and solutions. Subsequently, an industry-oriented large language model pre-trained on general corpus is selected, and a parameter-efficient fine-tuning technique is used to update the model parameters locally through a low-rank matrix decomposition method to inject professional knowledge from the power knowledge graph into the model in a way that minimizes the calculation cost, so that the model has both general language understanding ability and power domain semantic reasoning and professional question answering ability, and outputs a knowledge-enhanced power large language model.

[0024] In addition, industry standard items related to device operation, inspection, maintenance and safety specifications are extracted from multi-source data and converted into a machine-readable logical rule library, for example, "when the temperature of the transformer exceeds 85℃, it should be immediately operated in a reduced load mode" is converted into a logical rule form, such as: if temperature > 85℃, then perform a load reduction operation. The logical rule library participates in decision generation and solution verification as a constraint condition in the model reasoning stage, ensuring that the diagnostic results and solutions generated by the model meet the safety standards and operation specifications of the power industry, thereby realizing the deep integration and professional enhancement of the knowledge graph and the large language model.

[0025] Preferably, an intelligent acquisition terminal is deployed at the work site to integrate multiple sensing devices to obtain raw field work data, wherein the multiple sensing devices include a high-definition camera, an infrared sensor, a vibration sensor, and a sound sensor, and the raw field work data includes device appearance images, temperature distribution data, mechanical state data, and noise monitoring data.

[0026] It should be noted that the integration of multiple sensing devices to obtain raw field work data is as follows: the high-definition camera is used to obtain device appearance images, the infrared sensor is used to obtain temperature distribution data, the vibration sensor is used to obtain mechanical state data, and the sound sensor is used to obtain noise monitoring data. At the same time, the integration of multiple sensing devices not only enables the acquisition of raw field work data, but also enables the monitoring of abnormalities in raw field work data, i.e., through the high-definition camera to capture device appearance image abnormalities, through the infrared sensor to monitor temperature distribution abnormalities, through the vibration sensor to monitor mechanical state abnormalities, and through the sound sensor to monitor noise monitoring data abnormalities. Specifically, the means for monitoring abnormalities is to compare the raw field work data with a preset threshold range, if the raw field work data exceeds the preset threshold range, the raw field work data is abnormal, otherwise, the raw field work data is not abnormal.

[0027] The embodiment of the application takes the online inspection scene of a high-voltage circuit breaker in a substation as an example, deploys an intelligent collection terminal at the work site, and integrates various types of sensing devices on the terminal, including a high-definition camera, an infrared sensor, a vibration sensor, and a sound sensor, to realize real-time monitoring and abnormality identification of the running state of the equipment. The high-definition camera is used to collect appearance image data of the circuit breaker, covering the front, side, and key joint parts of the equipment, so as to identify surface abnormal conditions such as appearance structure deformation, oil leakage, and insulation porcelain sleeve rupture; the infrared sensor is used to obtain temperature distribution data of the circuit breaker and its terminal in real time, and by analyzing the thermal image, it can be judged whether there is a local overheating or uneven temperature rise phenomenon; the vibration sensor is installed at the equipment base and transmission mechanism, and is used to monitor the vibration spectrum characteristics of the mechanical parts to identify mechanical state abnormalities caused by looseness, wear, or bearing imbalance; the sound sensor is arranged near the operating mechanism of the circuit breaker, and is used to collect noise signals of the equipment during the opening and closing process, so as to detect abnormal impact sound, friction sound, or arc discharge sound.

[0028] During the on-site operation, the intelligent collection terminal periodically collects the original on-site operation data of each sensing device, including equipment appearance image, temperature distribution data, mechanical state data, and noise monitoring data, and performs real-time abnormality detection through the built-in edge computing module. Specifically, a threshold or feature distribution interval is preset for each type of sensing data: for example, the temperature distribution under normal running state is limited to 40℃-70℃, the mechanical vibration acceleration amplitude is limited to 0.2g-0.8g, and the noise sound pressure level is limited to 50dB-80dB. When the collected original data exceeds the corresponding preset threshold range, it is automatically determined that the data is abnormal and the alarm mechanism is triggered.

[0029] For example, in a certain inspection, the infrared sensor detects that the temperature peak of the contact point on the circuit breaker reaches 92℃, which exceeds the upper limit threshold of 70℃ set by the system; at the same time, the appearance image captured by the high-definition camera shows that the contact point area has obvious discoloration marks. After comprehensively analyzing the temperature and image features, the edge computing node immediately determines that the equipment has a potential risk of increased contact resistance, and marks this abnormal state as "temperature distribution abnormality - local overheating". Similarly, when the vibration sensor detects an abnormal high-frequency component in the mechanical spectrum, and the sound sensor simultaneously captures intermittent high-frequency friction noise, the system determines that the transmission mechanism of the equipment has mechanical wear or component looseness.

[0030] Through the above multi-sensor cooperative collection and abnormality monitoring method, the appearance abnormality, thermal abnormality, mechanical abnormality, and acoustic abnormality of the equipment can be identified in real time on site, providing accurate data support for subsequent feature extraction, fault type identification, and solution generation, thereby significantly improving the automation monitoring level and fault warning accuracy in the operation process of power equipment.

[0031] Preferably, the original field operation data is preprocessed and features are extracted based on the edge computing nodes to output a structured feature vector of the operation. The specific steps are as follows: The raw field operation data is preprocessed by the edge computing module embedded in the intelligent acquisition terminal; Set sampling frequency For the time window The original field operation data within the data center was continuously sampled to obtain the original time-series data matrix. In the formula, Let be the sampled value of the i-th sensor at time t. Let be the sampled value of the nth sensor at time t. The number of sensing devices; Feature extraction is performed based on the original time-series data matrix to obtain the sampling features of each sensor device, including the sampling mean, sampling variance, and sampling peak value. The fault feature vector of the sensor device is obtained by concatenating the sampling features. ,in, Let be the sampling mean of the i-th sensor. Let be the sampling variance of the i-th sensor. The peak value of the sampling of the i-th sensor; The fault feature vectors of each sensor are normalized and then concatenated to obtain the operational structured feature vector. ,in, This represents the fault feature vector of a high-definition camera. This represents the fault feature vector of the infrared sensor. This represents the fault feature vector of the vibration sensor. This represents the fault feature vector of the sound sensor. Fault feature values ​​are obtained by feature fusion based on the structured feature vectors of the operation. The calculation formula for feature fusion is as follows: ; In the formula: These are fault characteristic values. These are the weighting coefficients of the fault feature vector of the sound sensor. This represents the fault feature vector of a high-definition camera. These are the weighting coefficients of the fault feature vector of the infrared sensor. This represents the fault feature vector of the infrared sensor. These are the weighting coefficients of the fault feature vector of the vibration sensor. This represents the fault feature vector of the vibration sensor. These are the weighting coefficients of the fault feature vector of the sound sensor. This is the fault feature vector of the sound sensor.

[0032] It should be noted that feature extraction is performed based on the original time-series data matrix to obtain the sampling features of the sensing device, including the sampling mean, sampling variance, and sampling peak value. The steps for feature extraction of the device appearance image are as follows: by extracting the pixel values ​​of the corresponding appearance image within each sampling frequency, the mean, variance, and peak value are calculated based on the pixel values, thereby obtaining the sampling mean, sampling variance, and sampling peak value respectively.

[0033] This invention takes the operational status monitoring of a circuit breaker in a 220kV substation as an example. The system integrates a high-definition camera, infrared sensor, vibration sensor, and sound sensor through an intelligent data acquisition terminal installed on-site to acquire raw operational data of the equipment in real time. The intelligent data acquisition terminal has an embedded edge computing module for real-time preprocessing and feature extraction of the acquired multi-source heterogeneous data.

[0034] Set sampling frequency =1kHz, within the time window =Continuous sampling of the outputs of various sensors within 10 seconds to form an original time-series data matrix. For example, temperature data sequence recorded by infrared sensor, acceleration signal sequence output by vibration sensor, sound pressure level data sequence captured by sound sensor, and continuous frame appearance image captured by high-definition camera.

[0035] Based on the aforementioned raw time-series data, the edge computing module performs feature extraction operations on the data from each type of sensor. Taking the infrared sensor as an example, it calculates the mean, variance, and maximum peak value of the temperature data within the current sampling window, reflecting the overall temperature level, temperature fluctuation amplitude, and local maximum temperature rise of the equipment, respectively. After performing a Fourier transform on the acceleration signal collected by the vibration sensor, it extracts the mean and variance of the frequency domain energy to characterize the vibration stability of the mechanical system. For the noise signal from the sound sensor, it calculates the mean, variance, and peak value of the sound pressure level to determine whether there is any abnormal impact or discharge sound.

[0036] For the sequence of appearance images captured by the high-definition camera, pixel-level feature extraction is performed on each frame within the sampling window. Specifically, the pixel value set of the target area, such as the contact point or the ceramic sleeve surface, is extracted, and its gray-level mean, gray-level variance, and maximum gray-level peak value are calculated respectively: the gray-level mean reflects the overall brightness change of the equipment surface, the gray-level variance reflects the unevenness of texture or dirt distribution, and the gray-level peak value corresponds to potential bright or ablated areas. In this way, the sampling mean, sampling variance, and sampling peak value corresponding to the high-definition camera are obtained, which ultimately constitute the fault feature vector of the sensor.

[0037] Similarly, the infrared, vibration, and sound sensors output their corresponding feature vectors. The feature vectors of each sensor are normalized to eliminate differences in dimensions, and then spliced ​​together to obtain a complete operational structured feature vector. Subsequently, the multi-source features are fused and calculated according to preset weight coefficients.

[0038] For example, during a certain operation, the temperature variance of the infrared sensor increased significantly, the acceleration peak of the vibration sensor exceeded the threshold, and the grayscale peak of the image extracted by the high-definition camera deviated from the normal range, indicating that the equipment experienced localized overheating accompanied by mechanical loosening. Based on this, a set of feature vectors was generated and fused to output fault feature values. The edge diagnosis model immediately determined that the equipment had a potential risk of "localized overheating caused by loose mechanical support".

[0039] This approach enables automatic preprocessing, feature extraction, and feature fusion of multi-source sensor data at the edge, providing accurate and structured fault feature inputs for subsequent intelligent diagnostic modules, thereby improving the real-time performance and diagnostic accuracy of power equipment status identification.

[0040] Preferably, based on a knowledge-enhanced power big data language model, a hierarchical diagnostic mechanism involving edge-cloud collaboration is implemented to identify fault types and output fault identification results. The specific steps are as follows: An edge diagnosis model is obtained by knowledge distillation of a knowledge-enhanced power big language model, and a first-level fault identification is performed based on the edge diagnosis model. The fault type is determined based on the primary identification results. Fault types include standard fault types and rare fault types. For rare fault types, a knowledge-enhanced power big data language model is used for secondary fault identification to identify the root cause of the fault and its associated effects.

[0041] This invention takes the monitoring scenario of the main transformer in a 110kV substation as an example. Based on the fault feature values ​​obtained in the aforementioned steps, the edge-cloud collaborative hierarchical diagnosis mechanism of the knowledge-enhanced power big data language model is activated to realize the automatic identification and hierarchical analysis of equipment fault types.

[0042] First, a knowledge distillation method is used to extract core diagnostic knowledge and classification capabilities from a cloud-deployed knowledge-enhanced power big data model. While maintaining the integrity of the model's reasoning logic, the parameter scale is compressed and computational efficiency is optimized to obtain a lightweight edge diagnostic model. This edge model is deployed on the edge computing nodes of substations, enabling real-time reasoning and first-level fault identification locally. When the intelligent acquisition terminal uploads structured feature vectors or fused fault feature values, the edge diagnostic model immediately performs pattern matching and rapid classification, outputting preliminary fault judgment results.

[0043] In practical applications, when the transformer monitoring system detects abnormal temperature and high-amplitude vibration signals, the edge diagnostic model automatically identifies the "abnormal temperature rise - cooling failure" pattern based on previous training samples, assigns a high confidence level, and marks the result as a standard fault type. The system directly calls the local knowledge base to generate a processing solution, such as "check the cooling oil circulation loop, clean the heat dissipation channel, and verify the operating status of the temperature control device," and pushes the fault handling information to the duty terminal.

[0044] If the confidence level of the edge diagnostic model's identification results is low or the feature matching degree is insufficient, the current feature vector, preliminary diagnostic conclusion, and confidence level are uploaded to the cloud. The knowledge-enhanced power big data model deployed in the cloud then enters the secondary fault identification process. The cloud model performs deep reasoning analysis based on the entire network's historical case library, knowledge graph, and real-time operational data. For example, when frequent voltage fluctuations, high-frequency arc characteristics in the noise spectrum, and abnormally high humidity recorded in on-site meteorological data are detected, the model uses graph path reasoning to determine that the fault is highly correlated with "winding insulation aging." Through causal chain inference, the model further identifies the possible root cause as "partial discharge caused by insulation dampness," and combines this with the equipment structure graph to identify its potential associated effects as "increased coil temperature rise and shortened insulation life."

[0045] Finally, the identified fault types, root causes, and associated impacts are output in a structured format for maintenance personnel to view or automatically transmitted to the dispatch center. Compared with traditional diagnostic methods based on a single model, this edge-cloud collaborative hierarchical diagnostic mechanism achieves a two-layer linkage of "rapid on-site identification + deep cloud-based inference," ensuring both the real-time nature of the diagnosis and improving the accuracy and interpretability of complex fault identification, thus providing efficient support for the intelligent operation and maintenance of power equipment.

[0046] Preferably, the first-level fault identification is based on an edge diagnostic model, and the specific steps include: The edge diagnostic model is trained based on the historical fault description dataset. The real-time acquired fault description dataset is input for first-level fault identification. If the edge diagnostic model can match the same dataset in the historical fault description dataset, the equipment fault type is determined to be a standard fault type. If the edge diagnostic model cannot match the same dataset in the historical fault description dataset, the equipment fault type is determined to be a rare fault type.

[0047] This invention takes the main transformer monitoring system in a 220kV substation as an example. An edge diagnostic model, obtained through knowledge distillation from a knowledge-enhanced power big data model, is used to perform first-level fault identification on-site. First, the edge diagnostic model is trained using a historical fault description dataset accumulated by the enterprise. This dataset includes equipment operation records, maintenance work orders, abnormal event logs, and environmental parameter data from previous years, which are then encoded into a structured sample set. Each sample data entry contains basic equipment information (such as model and voltage level), fault phenomenon characteristics (such as temperature rise and abnormal vibration), operating condition parameters, and corresponding standard fault type labels to guide the edge model's classification learning.

[0048] During on-site operation, when the intelligent acquisition terminal collects new fault description datasets in real time—for example, an infrared sensor detects abnormal temperature, a vibration sensor detects a sudden increase in high-frequency amplitude, or a sound sensor captures intermittent impact noise—this data is input into the edge diagnostic model for rapid matching and identification. The edge model first encodes the real-time fault description data into feature vectors and compares their similarity with locally stored historical fault description samples. When the similarity calculated by the model is higher than a set threshold, it determines that the fault has a corresponding type in the historical sample set, i.e., it is identified as a "standard fault type." For example, in a certain detection, the model identifies that this feature is extremely similar to the historical sample "cooling fan jamming causing temperature rise," so it directly outputs the diagnostic result as "cooling system malfunction" and calls the local knowledge base to generate a corresponding processing solution.

[0049] If the model fails to find a record in the historical sample set with a matching degree exceeding a threshold, the fault is classified as a "rare fault type." For example, the system detects an asymmetric hotspot in the temperature distribution recorded by the infrared sensor, while the sound sensor detects low-frequency resonance features, but this feature combination does not appear in the historical sample set or has low similarity. In this case, the edge model marks the preliminary diagnosis result as a "rare fault" and uploads the fault feature vector, preliminary conclusion, and corresponding confidence level to the cloud, where a knowledge-enhanced power big data language model deployed in the cloud performs secondary recognition.

[0050] Through the above methods, the edge diagnostic model can quickly identify and respond to standard fault types on-site, while effectively diverting unknown or rare anomalies, allowing cloud resources to be concentrated on in-depth reasoning analysis of complex problems, thereby constructing an end-edge-cloud collaborative diagnostic system that combines real-time performance and intelligence.

[0051] Preferably, for rare fault types, a knowledge-enhanced power big data language model is used for secondary fault identification to identify the root cause of the fault and its associated effects. The specific steps are as follows: Obtain the fault description dataset corresponding to rare fault types, and upload the fault feature vector, preliminary diagnosis results and confidence scores to the cloud via the edge device; Secondary fault identification is performed by using a knowledge-enhanced power big data language model deployed in the cloud. In-depth analysis is conducted by combining historical cases from the entire network, real-time operation data, and meteorological information to identify the root causes and related impacts of the fault.

[0052] It should be noted that after a fault is identified as a rare type and the edge device reports the fault feature vector, preliminary diagnosis results, and confidence level to the cloud, the cloud initiates a secondary fault identification process based on a knowledge-enhanced power big data language model. The cloud first performs integrity and security checks on the reported data, and then uses the fault feature vector as the query key to search for similar historical cases in the vector retrieval database. Simultaneously, it searches the knowledge graph for subgraphs related to equipment type, fault representation, and operational semantics, including entities, relationships, and causal edges. Subsequently, the top K most similar cases to the current fault, knowledge graph triples, equipment topology, and recent real-time operating data (such as temperature, current, and load curves for the past 24 hours), along with relevant meteorological information, are combined to form an enhanced context.

[0053] Based on this enhanced context, the knowledge-enhanced power big data language model initiates multimodal semantic reasoning through retrieval-enhanced generation: first, text summaries, time-series data, and visual / thermal imaging fragments are encoded separately, then aligned using a cross-modal attention mechanism to extract the evidence fragments most relevant to the current manifestation. The model generates a preliminary set of root cause hypotheses, each accompanied by natural language arguments and source citations, and invokes a knowledge graph to perform causal path search. A graph search algorithm quantifies the path score from observed symptoms to candidate root causes in the knowledge graph, with path confidence based on historical co-occurrence rate and expert weighting. For each candidate root cause, a multi-source evidence consistency score is further calculated: combining the confidence score, path score, similarity to the most similar historical cases, and the matching degree between real-time data and the hypothesis generated by the knowledge-enhanced power big data language model, a comprehensive ranking score is obtained, thus outputting a ranked list of root cause candidates and labeling the evidence chain of each candidate, such as the cited case ID, knowledge graph path, key sensor curve snapshot, or anomaly heatmap.

[0054] After identifying the primary root cause candidates, the model simulates potential chain reactions by propagating outwards along the "impact" relationships in the knowledge graph. It then assesses the risk level and time window of affected components and services by combining power grid / equipment topology and real-time load data, generating a list of associated impacts, including affected equipment, the potential scope of service disruptions, estimated severity, and propagation paths. Based on the root cause and impact analysis, the knowledge-enhanced power big data model further generates multiple actionable mitigation and remediation suggestions, each accompanied by estimated resources, timeframes, risk descriptions, and verification rules. All outputs are distributed in structured report format, including root cause candidates, causal paths, impact lists, recommended solutions, evidence citations, and model / knowledge base version numbers. Low-confidence or high-risk suggestions automatically trigger a manual review process. The review results and execution feedback are then used to update the historical case library and knowledge graph, triggering incremental learning or distillation updates, thereby achieving a traceable, auditable, and continuously self-optimizing two-tiered identification closed loop.

[0055] Preferably, based on the fault identification results, a knowledge-enhanced power big data language model is invoked to generate multi-dimensional solutions. A machine learning-based scoring model is then constructed to score these multi-dimensional solutions, determining the final solution. The specific steps are as follows: By using historical successful solution cases in the knowledge-enhanced power big language model, we explore multi-dimensional solutions using the Monte Carlo tree search algorithm; By using historical successful solution cases as training samples to build a machine learning-based scoring model, the model learns the mapping relationship between case features and success rate, and scores multi-dimensional solutions. The multi-dimensional solutions are ranked based on the scoring results. The multi-dimensional solution ranked first is taken as the final solution, and the multi-dimensional solution ranked second is taken as the alternative solution.

[0056] It should be noted that the multi-dimensional solution includes a sequence of operational steps, as well as comprehensive information such as required tools and materials, estimated time, safety measures, and risk level.

[0057] It should also be noted that the scoring dimensions of the machine learning-based scoring model include technical feasibility, resource availability, scope of power outage impact, and operational complexity.

[0058] This invention takes a cooling system anomaly in a 110kV distribution transformer as an example. After completing the aforementioned primary and secondary fault identification, the fault type has been determined to be "local overheating caused by obstructed cooling oil circulation." At this point, based on the fault identification results, a knowledge-enhanced power big data language model is invoked to generate a multi-dimensional set of solutions. The knowledge-enhanced power big data language model searches the power knowledge graph and historical successful solution cases, and uses the Monte Carlo tree search algorithm to explore and optimize in the operation strategy space, forming multiple feasible solution branches.

[0059] During the search process, the Monte Carlo tree search algorithm uses the current fault characteristics as the root node and combines model reasoning to generate different operation paths as child nodes, such as "Solution A: Clean the cooling oil passage + replace the circulating pump", "Solution B: Drain the cooling oil + check the temperature control device + add new oil", and "Solution C: Shutdown and maintenance + clean the air duct + calibrate the sensor". Each node generates a detailed sequence of operation steps from the model, along with comprehensive information such as the required tools and materials (e.g., cleaning agent, oil pump, temperature control module), estimated time, safety measures (e.g., whether power outage is required), risk level, and environmental impact assessment, forming a complete set of multi-dimensional solutions.

[0060] Subsequently, using historical successful solution cases built into the knowledge-enhanced power big data language model as training samples, a machine learning-based scoring model is constructed. This scoring model extracts key features from historical cases through feature engineering, including the complexity of operation steps, the amount of resources required, environmental constraints, power outage duration, and final repair success rate. It learns the mapping relationship between solution features and actual success rates, thereby enabling quantitative scoring of newly generated multi-dimensional solutions.

[0061] During the scoring phase, each solution is comprehensively evaluated across multiple dimensions, including technical feasibility, resource availability, power outage impact range, operational complexity, and safety risk level. For example, "Solution A" receives a score of 0.92 based on its short operation steps, low resource consumption, and moderate risk level; "Solution B" receives a score of 0.81 due to the need for replenishing cooling oil and a longer downtime; and "Solution C" receives a score of 0.68 due to its involvement in power outage maintenance and higher safety risk. The scoring model sorts all solutions in descending order of their scores and automatically identifies "Solution A," ranked first, as the final solution, while setting "Solution B" as an alternative.

[0062] Once the solution is finalized, it will be distributed to the field terminals in the form of a structured task list, including operating instructions, a list of tools and materials, safety precautions, estimated operation time, and risk warnings. After the operation is completed, the actual execution time, resource usage, and repair results will be automatically recorded as feedback samples for subsequent model retraining and scoring model optimization.

[0063] This embodiment enables the automatic generation of multi-dimensional and structured solutions after fault diagnosis, and the optimization of solutions based on a data-driven machine learning scoring mechanism, which significantly improves the scientificity, rationality and execution efficiency of power operation and maintenance decisions, and realizes a fully intelligent closed loop from fault identification to decision output.

[0064] Preferably, based on the final solution, the execution result is determined, and the execution result, actual time consumption, fault type, and final solution are automatically collected as feedback information to update and optimize the knowledge-enhanced power big data language model. The specific steps are as follows: By automatically collecting execution results, actual time consumption, fault type and final solution as feedback information, the active learning algorithm identifies solutions that exceed the expected difference threshold, triggers expert review and annotation, and forms new training samples. Incremental learning technology is used to integrate new samples into the training process of the knowledge-enhanced power big data language model and update the model's parameters and knowledge graph. At the same time, causal reasoning analysis is used to extract fault causal relationships from feedback information.

[0065] In an intelligent operation and maintenance system of a power dispatch center, when an abnormal fault occurs in the distribution network, the system first generates multi-dimensional processing solutions based on a knowledge-enhanced power big data language model and automatically selects the final solution for execution. During execution, the system records the execution results, actual time consumption, fault type, and the final solution adopted in real time, and automatically collects this data as feedback information. Subsequently, an active learning algorithm analyzes this feedback information. When it finds that the actual effect of some solutions differs significantly from the expectation, it is automatically marked as an anomaly and triggers expert review. Experts then annotate and optimize the solutions, forming new training samples. Next, incremental learning technology is used to integrate these new samples into the training process of the power big data language model, while updating the model parameters and knowledge graph, enabling the model to generate subsequent solutions more accurately. Furthermore, causal reasoning analysis is used to extract the causal relationship of the fault from the feedback information, such as the tripping of a certain type of circuit breaker being caused by specific load fluctuations, further enriching the model's knowledge structure and improving its fault diagnosis and processing capabilities, thus realizing the self-optimization and continuous learning of the intelligent operation and maintenance system.

[0066] The power operation problem-solving system based on a large model includes a power knowledge construction module, a multi-source data acquisition and feature fusion module, an edge-cloud collaborative hierarchical diagnosis module, a multi-dimensional solution generation and scoring module, and a model self-optimization feedback module. The power knowledge construction module is connected to the multi-source data acquisition and feature fusion module, which is connected to the edge-cloud collaborative hierarchical diagnosis module. The edge-cloud collaborative hierarchical diagnosis module is connected to the multi-dimensional solution generation and scoring module, which is connected to the model self-optimization feedback module.

[0067] The power knowledge construction module is used to generate a power domain knowledge base by collecting multi-source data from power companies, train an industry-specific large language model based on the power domain knowledge base, and output a knowledge-enhanced power large language model. The multi-source data acquisition and feature fusion module is used to acquire raw field operation data by deploying intelligent acquisition terminals at the work site and integrating multiple sensing devices. It preprocesses and extracts features from the raw field operation data based on edge computing nodes, outputs fault feature values, and fuses the fault feature values ​​with equipment basic information, historical operating data, and environmental parameters to form a fault description dataset. The edge-cloud collaborative hierarchical diagnosis module is used to implement a hierarchical diagnosis mechanism based on the knowledge-enhanced power big data language model, identify fault types, and output fault identification results. The multi-dimensional solution generation and scoring module is used to call the knowledge-enhanced power big data language model based on the fault identification results to generate multi-dimensional solutions. The multi-dimensional solutions are scored by constructing a machine learning-based scoring model to determine the final solution. The model self-optimization feedback module is used to determine the execution result based on the final solution, automatically collect the execution result, actual time consumption, fault type and final solution as feedback information, and update and optimize the knowledge-enhanced power big data language model.

[0068] like Figure 2 The diagram shown is a system block diagram of a power operation problem-solving system based on a large model, according to an embodiment of the present invention. This system can be used to execute... Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0069] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the steps of the large-model-based power operation problem-solving method, system, and electronic device described in any of the preceding claims.

[0070] An electronic device includes: a processor, a memory, and a computer program; wherein, A memory is used to store the computer program, and the memory may also be flash memory. The computer program is, for example, an application program or functional module that implements the above method.

[0071] The processor is used to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0072] Alternatively, the memory can be either standalone or integrated with the processor.

[0073] When the memory is a device independent of the processor, the device may further include: A bus is used to connect the memory and the processor.

[0074] Through the above embodiments, this invention constructs a power industry knowledge base by collecting multi-source data from enterprises, and trains an industry-specific large language model based on this, forming a knowledge-enhanced power large language model. At the work site, intelligent acquisition terminals are deployed, integrating multiple sensors to acquire raw data. Edge computing is used for preprocessing and feature extraction to generate fault feature values, which are then fused with equipment basic information, historical operating data, and environmental parameters to form a complete fault description dataset. Based on this model, a hierarchical diagnostic mechanism involving edge, cloud, and end-to-end collaboration is implemented to identify fault types and output diagnostic results. Subsequently, the model is invoked to generate multi-dimensional solutions, which are evaluated using a machine learning scoring model to determine the optimal execution plan. Execution results, actual time consumption, fault type, and final solution are automatically collected as feedback information for updating and optimizing the model, enabling continuous learning. This invention can improve fault identification accuracy and response speed, optimize solution selection, reduce reliance on experts, accumulate systematic operation and maintenance knowledge, and significantly improve power operation and maintenance efficiency and safety.

[0075] This invention, through multi-source data fusion and fault feature extraction, combined with a knowledge-enhanced power big data language model, can accurately identify complex equipment fault types and reduce human error. The edge-cloud collaborative hierarchical diagnostic mechanism and on-site intelligent data acquisition terminals enable real-time collection and processing of fault information, shortening the time from fault occurrence to resolution. A machine learning scoring model is used to evaluate multi-dimensional solutions, ensuring the final solution is efficient, safe, and implementable. By automatically collecting execution results and feedback information and using incremental learning to update the model, the system can continuously optimize and gradually form a systematic power operation and maintenance knowledge base. Combining the big data model with automation mechanisms, it can provide scientific decision-making directly on-site, alleviating the problem of limited expert resources. The overall process is intelligent and data-driven, reducing equipment failure risks and economic losses while ensuring the safety of operation and maintenance personnel.

[0076] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

[0077] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A solution to the power operation problem based on a large model, characterized in that, Includes the following steps: By collecting multi-source data from power companies, a knowledge base for the power industry is generated. Based on the knowledge base, an industry-specific large language model is trained, and a knowledge-enhanced power large language model is output. By deploying intelligent data acquisition terminals at the work site, multiple sensing devices are integrated to acquire raw on-site work data. The raw on-site work data is preprocessed and features are extracted based on edge computing nodes, and fault feature values ​​are output. The fault feature values ​​are then fused with equipment basic information, historical operating data, and environmental parameters to form a fault description dataset. Based on a knowledge-enhanced power big data language model, a hierarchical diagnostic mechanism is implemented through edge-cloud collaboration to identify fault types and output fault identification results. Based on the fault identification results, a knowledge-enhanced power big data language model is invoked to generate multi-dimensional solutions. A machine learning-based scoring model is then constructed to score the multi-dimensional solutions and determine the final solution. The execution result is determined based on the final solution. The execution result, actual time consumption, fault type and final solution are automatically collected as feedback information to update and optimize the knowledge-enhanced power big data language model.

2. The solution to the power operation problem based on a large model according to claim 1, characterized in that, By deploying intelligent data acquisition terminals at the work site, multiple sensors are integrated to acquire raw on-site work data. These sensors include high-definition cameras, infrared sensors, vibration sensors, and sound sensors. The raw on-site work data includes equipment appearance images, temperature distribution data, mechanical status data, and noise monitoring data.

3. The solution to the power operation problem based on a large model according to claim 1, characterized in that, The raw field operation data is preprocessed and features are extracted based on edge computing nodes to output a structured feature vector of the operation. The specific steps are as follows: The raw field operation data is preprocessed by the edge computing module embedded in the intelligent acquisition terminal; Set sampling frequency For the time window The original field operation data within the data center was continuously sampled to obtain the original time-series data matrix. In the formula, Let be the sampled value of the i-th sensor at time t. Let be the sampled value of the nth sensor at time t. The number of sensing devices; Feature extraction is performed based on the original time-series data matrix to obtain the sampling features of each sensor device, including the sampling mean, sampling variance, and sampling peak value. The fault feature vector of the sensor device is obtained by concatenating the sampling features. ,in, Let be the sampling mean of the i-th sensor. Let be the sampling variance of the i-th sensor. The peak value of the sampling of the i-th sensor; The fault feature vectors of each sensor are normalized and then concatenated to obtain the operational structured feature vector. ,in, This represents the fault feature vector of a high-definition camera. This represents the fault feature vector of the infrared sensor. This represents the fault feature vector of the vibration sensor. This represents the fault feature vector of the sound sensor; Fault feature values ​​are obtained by feature fusion based on the structured feature vectors of the operation.

4. The solution to the power operation problem based on a large model according to claim 1, characterized in that, Based on a knowledge-enhanced power big data language model, a hierarchical diagnostic mechanism involving edge-cloud collaboration is implemented to identify fault types and output fault identification results. The specific steps are as follows: An edge diagnosis model is obtained by knowledge distillation of a knowledge-enhanced power big language model, and a first-level fault identification is performed based on the edge diagnosis model. The fault type is determined based on the primary identification results. Fault types include standard fault types and rare fault types. For rare fault types, a knowledge-enhanced power big data language model is used for secondary fault identification to identify the root cause of the fault and its associated effects.

5. The solution to the power operation problem based on a large model according to claim 4, characterized in that, The steps for first-level fault identification based on the edge diagnostic model include: The edge diagnostic model is trained based on the historical fault description dataset. The real-time acquired fault description dataset is input for first-level fault identification. If the edge diagnostic model can match the same dataset in the historical fault description dataset, the equipment fault type is determined to be a standard fault type. If the edge diagnostic model cannot match the same dataset in the historical fault description dataset, the equipment fault type is determined to be a rare fault type.

6. The solution to the power operation problem based on a large model according to claim 4, characterized in that, For rare fault types, a knowledge-enhanced power big data language model is used for secondary fault identification to identify the root cause of the fault and its associated effects. The specific steps are as follows: Obtain the fault description dataset corresponding to rare fault types, and upload the fault feature vector, preliminary diagnosis results and confidence scores to the cloud via the edge device; Secondary fault identification is performed by using a knowledge-enhanced power big data language model deployed in the cloud. In-depth analysis is conducted by combining historical cases from the entire network, real-time operation data, and meteorological information to identify the root causes and related impacts of the fault.

7. The solution to the power operation problem based on a large model according to claim 1, characterized in that, Based on the fault identification results, a knowledge-enhanced power big data language model is invoked to generate multi-dimensional solutions. A machine learning-based scoring model is then constructed to score these multi-dimensional solutions, determining the final solution. The specific steps are as follows: Based on historical successful solution cases in the knowledge-enhanced power big language model, the Monte Carlo tree search algorithm is used to explore multi-dimensional solutions; By using historical successful solution cases as training samples to build a machine learning-based scoring model, the model learns the mapping relationship between case features and success rate, and scores multi-dimensional solutions. The multi-dimensional solutions are ranked based on the scoring results. The multi-dimensional solution ranked first is taken as the final solution, and the multi-dimensional solution ranked second is taken as the alternative solution.

8. The solution to the power operation problem based on a large model according to claim 1, characterized in that, Based on the final solution, the execution result is determined, and the execution result, actual time consumption, fault type, and final solution are automatically collected as feedback information to update and optimize the knowledge-enhanced power big data language model. The specific steps are as follows: By automatically collecting execution results, actual time consumption, fault type and final solution as feedback information, the active learning algorithm identifies solutions that exceed the expected difference threshold, triggers expert review and annotation, and forms new training samples. Incremental learning technology is used to integrate new samples into the training process of the knowledge-enhanced power big data language model and update the model's parameters and knowledge graph. At the same time, causal reasoning analysis is used to extract fault causal relationships from feedback information.

9. A power operation problem-solving system based on a large model, applied to the power operation problem-solving method based on a large model as described in any one of claims 1-8, characterized in that, The system includes: a power knowledge construction module, a multi-source data acquisition and feature fusion module, an edge-cloud collaborative hierarchical diagnosis module, a multi-dimensional solution generation and scoring module, and a model self-optimization feedback module. The power knowledge construction module is used to generate a power domain knowledge base by collecting multi-source data from power companies, train an industry-specific large language model based on the power domain knowledge base, and output a knowledge-enhanced power large language model. The multi-source data acquisition and feature fusion module is used to acquire raw field operation data by deploying intelligent acquisition terminals at the work site and integrating multiple sensing devices. It preprocesses and extracts features from the raw field operation data based on edge computing nodes, outputs fault feature values, and fuses the fault feature values ​​with equipment basic information, historical operating data, and environmental parameters to form a fault description dataset. The edge-cloud collaborative hierarchical diagnosis module is used to implement a hierarchical diagnosis mechanism based on the knowledge-enhanced power big data language model, identify fault types, and output fault identification results. The multi-dimensional solution generation and scoring module is used to call the knowledge-enhanced power big data language model based on the fault identification results to generate multi-dimensional solutions. The multi-dimensional solutions are scored by constructing a machine learning-based scoring model to determine the final solution. The model self-optimization feedback module is used to determine the execution result based on the final solution, automatically collect the execution result, actual time consumption, fault type and final solution as feedback information, and update and optimize the knowledge-enhanced power big data language model.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor runs the computer program stored in the memory, the processor performs the steps of the large-model-based power operation problem solution as described in any one of claims 1-8.

Images