Work ticket auxiliary decision-making method and system

Abstract

The invention provides a work ticket auxiliary decision-making method and system, and relates to the technical field of electric power operation, and the method comprises the steps: collecting and fusing the multi-source data of an electric power operation system, and obtaining the multi-modal data of the electric power operation system; obtaining deep semantic features of the multi-modal data according to the multi-modal data through a large-scale language model, a data feature extraction model and an image feature extraction model; matching according to the deep semantic features through a preset knowledge graph to obtain knowledge graph nodes and relationships related to the multi-modal data; determining an initial auxiliary decision of the work ticket of the power operation by combining a preset deep reinforcement learning network; inputting the data into a power operation system model for simulation to obtain an execution effect; and optimizing the initial auxiliary decision according to the execution effect and the expected execution effect to obtain a final auxiliary decision of the work ticket. According to the invention, the accuracy and consistency of decision making are improved, and the intelligent degree of auxiliary decision making of the work ticket is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power operation, and in particular, to a work ticket auxiliary decision-making method and system. Background Art

[0002] A work ticket is a key document for ensuring work safety in the power industry. It covers information such as work content, time, location, personnel, and safety measures, and is the basis for work permit, guardianship, acceptance, etc. It can standardize the operation process, clarify responsibilities, and ensure safety. Therefore, through reasonable decisions contained in the work ticket, risks can be avoided, safety and efficiency can be ensured, and it is the core of the power operation safety system.

[0003] In the related art, the decisions contained in traditional work tickets usually require staff to spend a lot of time searching for relevant rules in a large amount of historical data to formulate, and it is difficult to deeply mine and accurately utilize historical data. At the same time, due to excessive reliance on manual experience, differences in understanding and judgment among different personnel will lead to unstable decision-making results of work tickets, and it is difficult to guarantee accuracy and consistency. Summary of the Invention

[0004] The problem solved by the present invention is how to improve the intelligence level of work ticket auxiliary decision-making.

[0005] To solve the above problems, the present invention provides a work ticket auxiliary decision-making method and system.

[0006] In a first aspect, the present invention provides a work ticket auxiliary decision-making method, which is applied to a distribution network. The distribution network includes a plurality of distribution facilities, and each distribution facility corresponds to a different power operation system. The method includes: when it is necessary to perform an operation on the distribution facility, collecting and fusing multi-source data of the power operation system of the distribution facility to obtain multi-modal data of the power operation system; through a large-scale language model, a data feature extraction model, and an image feature extraction model, according to the multi-modal data, obtaining deep semantic features of the multi-modal data; through a preset knowledge graph, matching according to the deep semantic features to obtain knowledge graph nodes and relationships related to the multi-modal data; according to the knowledge graph nodes and relationships, combining a preset deep reinforcement learning network to determine an initial auxiliary decision of a work ticket for power operation; inputting the initial auxiliary decision into a power operation system model for simulation to obtain an execution effect corresponding to the initial auxiliary decision; optimizing the initial auxiliary decision according to the execution effect and the expected execution effect to obtain a final auxiliary decision of the work ticket.

[0007] Optionally, when it is necessary to perform operations on the power distribution facilities, multi-source data of the power operation system of the power distribution facilities is collected and fused to obtain multi-modal data of the power operation system, including: when it is necessary to perform operations on the power distribution facilities, text data, sensing data, image data, and equipment operation data in the power operation system of the power distribution facilities are collected, and the text data, the sensing data, the image data, and the equipment operation data are used as the multi-source data; the multi-source data is converted into a preset format through data fusion technology, and the multi-source data in the unified format is fused to obtain the multi-modal data of the power operation system.

[0008] Optionally, obtaining deep semantic features of the multi-modal data through a large-scale language model, a data feature extraction model, and an image feature extraction model according to the multi-modal data includes: Inputting the multi-modal data into the large-scale language model, and performing semantic extraction on the text data in the multi-modal data through the large-scale language model to obtain semantic information of the multi-modal data; inputting the multi-modal data into the data feature extraction model, and performing feature extraction on the sensing data and the equipment operation data in the multi-modal data through the data feature extraction model to obtain data features of the multi-modal data; inputting the multi-modal data into the image feature extraction model, and performing feature extraction on the image data in the multi-modal data through the image feature extraction model to obtain image features of the multi-modal data; fusing the semantic information, the data features, and the image features to obtain the deep semantic features.

[0009] Optionally, matching according to the deep semantic features through a preset knowledge graph to obtain knowledge graph nodes and relationships related to the multi-modal data includes: performing vectorization processing on the deep semantic features to obtain a feature vector of the deep semantic features; embedding entities and relationships in the preset knowledge graph into a vector space through graph embedding technology to obtain knowledge graph embedding vectors; calculating the cosine similarity between the knowledge graph embedding vectors and the feature vector; determining the knowledge graph nodes and relationships related to the multi-modal data according to the cosine similarity.

[0010] Optionally, determining the initial auxiliary decision of the work ticket for power operation based on the knowledge graph nodes and relationships and in combination with a preset deep reinforcement learning network includes: taking the risk assessment level, safety measure selection, and work process arrangement of the power operation as the actions of the deep reinforcement learning network, and taking the knowledge graph nodes and relationships as the states of the deep reinforcement learning network; splicing the embedding vectors of the knowledge graph nodes and relationships with the feature vectors of the work site environment parameters to form a state vector and inputting the state vector into the deep reinforcement learning network; selecting to update the actions according to the state vector, and judging whether the update of the actions is completed according to a preset reward function; taking the updated actions as the initial auxiliary decision of the work ticket.

[0011] Optionally, inputting the initial auxiliary decision into a power operation system model for simulation to obtain the execution effect corresponding to the initial auxiliary decision includes: inputting the initial auxiliary decision into the power operation system model, and performing simulation execution according to the risk assessment level, the safety measure selection, and the work process arrangement; after the simulation execution is completed, obtaining the safety index, efficiency index, and specification index of the power operation system model after the simulation execution; taking the safety index, the efficiency index, and the specification index as the execution effect.

[0012] Optionally, optimizing the initial auxiliary decision according to the execution effect and the expected execution effect to obtain the final auxiliary decision of the work ticket includes: judging whether the initial auxiliary decision meets the operation requirements by comparing the expected execution effect with the execution effect; when the initial auxiliary decision meets the operation requirements, taking the initial auxiliary decision as the final auxiliary decision; when the initial auxiliary decision does not meet the operation requirements, optimizing the initial auxiliary decision according to the comparison result of the expected execution effect and the execution effect, and taking the optimized initial auxiliary decision as the final auxiliary decision.

[0013] Optionally, judging whether the initial auxiliary decision meets the operation requirements by comparing the expected execution effect with the execution effect includes: judging whether the safety index, the efficiency index, and the specification index all meet the corresponding expected execution effects; if the safety index, the efficiency index, and the specification index all meet the corresponding expected execution effects, determining that the initial auxiliary decision meets the operation requirements; if any one of the safety index, the efficiency index, and the specification index does not meet the corresponding expected execution effect, determining that the initial auxiliary decision does not meet the operation requirements.

[0014] Optionally, optimizing the initial auxiliary decision according to the comparison result between the expected execution effect and the execution effect includes: when the initial auxiliary decision does not meet the operation requirements, determining the index gap between the index that does not meet the expected execution effect in the execution effect and the expected execution effect corresponding to the index according to the execution effect, and using the index gap as the comparison result; modifying the initial auxiliary decision according to the index gap in combination with a preset adjustment rule to obtain the final auxiliary decision.

[0015] In a second aspect, the present invention provides a work ticket auxiliary decision-making system, which is applied to a distribution network. The distribution network includes a plurality of distribution facilities, and each distribution facility corresponds to a different power operation system. The system includes: a data processing unit for collecting and fusing multi-source data of the power operation system of the distribution facility when an operation needs to be performed on the distribution facility to obtain multi-modal data of the power operation system; a feature extraction unit for obtaining deep semantic features of the multi-modal data according to the multi-modal data through a large-scale language model, a data feature extraction model, and an image feature extraction model; a matching unit for matching according to the deep semantic features through a preset knowledge graph to obtain knowledge graph nodes and relationships related to the multi-modal data; an initial decision-making unit for determining an initial auxiliary decision of a work ticket for power operation according to the knowledge graph nodes and relationships in combination with a preset deep reinforcement learning network; a simulation unit for inputting the initial auxiliary decision into a power operation system model for simulation to obtain an execution effect corresponding to the initial auxiliary decision; an optimization unit for optimizing the initial auxiliary decision according to the execution effect and the expected execution effect to obtain the final auxiliary decision of the work ticket.

[0016] In the work ticket auxiliary decision-making method and system of the present invention, since the distribution network includes a plurality of distribution facilities, and each facility has a corresponding power operation system, by collecting and fusing multi-source data of these different systems, it can be ensured that all relevant facility operation states and operation requirements are comprehensively considered when formulating a work ticket. For example, for different facilities such as substations, switchgear, distribution lines, and distribution transformers, data such as their electrical parameters, equipment status, and operating environment are respectively obtained to avoid missing key information.

[0017] Collect multi-source data of the power operation system. Through fusion processing, multi-modal data of the power operation system can be obtained. The multi-modal data can present the on-site situation more completely, providing comprehensive and rich basic information for subsequent analysis and decision-making, and solving the problem that traditional work ticket decisions rely on manual experience to consult limited data and it is difficult to deeply mine and accurately utilize historical data. Moreover, the multi-modal data (such as text, images, time series data, etc.) provided by different power operation systems can complement and verify each other, thereby improving the accurate assessment of the actual state of distribution facilities. For example, by combining image recognition technology to detect abnormalities in the appearance of equipment (such as damaged insulators), and at the same time combining the analysis of operating parameters (such as current, voltage), it can be more accurately judged whether the equipment needs to be repaired or replaced.

[0018] At the same time, different distribution facilities in the distribution network (such as substations, switchgear stations, overhead lines, cable lines, etc.) have different functions and operation requirements. By collecting and analyzing data through the power operation system for each facility, the personalized operation needs of different types of facilities can be met. For example, the operation of a substation may involve complex relay protection and automation systems, while the operation of a distribution line pays more attention to line inspection and fault location. The power operation systems of different distribution facilities can adapt to the complex and changeable distribution network environment. For example, in the core urban area, industrial area and rural area, the configuration and operating conditions of distribution facilities vary greatly. By configuring an independent power operation system for each facility, it is possible to better cope with the operation challenges in different environments and improve the applicability and effectiveness of work tickets.

[0019] By leveraging large language models, data feature extraction models, and image feature extraction models, the deep semantic features of multimodal data are synthesized, enabling the system to deeply understand the meaning behind the data rather than merely staying on the surface as in traditional methods, thereby providing deeper feature support for accurate decision-making. Then, by presupposing a knowledge graph and matching it according to the deep semantic features, the knowledge graph nodes and relationships related to the multimodal data are found. Matching the deep semantic features with the knowledge graph can link the current situation at the job site with existing professional knowledge, providing a knowledge background for subsequent decision-making. The knowledge graph provides a unified and standardized knowledge system, solving the problem of unstable decision-making caused by relying on manual experience and different understandings and judgments among different personnel in traditional methods. Based on the knowledge graph nodes and relationships, combined with a presupposed deep reinforcement learning network, the initial auxiliary decision for the work ticket of the power operation is determined. The deep reinforcement learning network can learn the optimal decision-making strategy according to the current state (i.e., the matched knowledge graph nodes and relationships) and the presupposed reward mechanism. It can select a relatively reasonable initial auxiliary decision from numerous possible decision-making schemes. Compared with the traditional method of manually making decisions, it can generate a preliminary scheme more quickly and intelligently. Further, the initial auxiliary decision is input into the power operation system model for simulation to obtain the execution effect corresponding to the initial auxiliary decision. Through simulation operations, before actual operations, the operation of this decision in the virtual environment can be observed, potential problems can be discovered in advance, and the feasibility of the decision can be evaluated; it helps to avoid trial and error directly in actual operations and reduce risks. Finally, according to the execution effect and the expected execution effect, the initial auxiliary decision is optimized to obtain the final auxiliary decision for the work ticket. This process of the present invention uses a feedback mechanism to adjust the initial decision to make it more in line with the actual needs and expected goals, thereby improving the accuracy and consistency of decision-making and solving the problem of unstable decision-making results in traditional methods, achieving an improvement in the intelligent level of work ticket auxiliary decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of the work ticket auxiliary decision-making method according to an embodiment of the present invention; Figure 2 It is a schematic structural diagram of a distribution network according to an embodiment of the present invention; Figure 3 It is a block diagram of the structure of the work ticket auxiliary decision-making system according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of specific embodiments of the present invention in conjunction with the accompanying drawings. Although certain embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Instead, these embodiments are provided to more thoroughly and completely understand the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not used to limit the protection scope of the present invention.

[0022] Combined with Figure 1 and Figure 2 As shown, a work ticket auxiliary decision-making method provided by an embodiment of the present invention is applied to a distribution network, which includes multiple distribution facilities, and each distribution facility corresponds to a different power operation system.

[0023] Specifically, in a distribution network, distribution facilities mainly include various equipment such as substations, switchgear stations, distribution lines (such as overhead lines and cable lines), distribution transformers, circuit breakers, disconnect switches, etc. These facilities constitute a complete path for power transmission from the power generation end to the user end. For example, a distribution network in an urban area may have multiple 35 kV substations. These substations receive electrical energy from higher voltage level transmission lines (such as 110 kV or 220 kV transmission lines), then reduce the voltage to about 10 kV through transformers, and then transmit it to each user area through 10 kV distribution lines. In the user area, there will be switchgear stations to further distribute electrical energy, and distribution transformers convert 10 kV voltage into 380V / 220V voltage for ordinary users.

[0024] In a preferred embodiment of the present invention, combined with Figure 2 As shown, the power operation systems corresponding to different distribution facilities, such as the substation power operation system, which is a complex integrated system. It includes an electrical equipment monitoring system for real-time monitoring of the operating parameters of equipment such as transformers, circuit breakers, and busbars in the substation, such as voltage, current, active power, reactive power, temperature, etc. For example, through various sensors installed on the transformer, the monitoring system can obtain data such as transformer winding temperature and oil temperature, which are crucial for judging whether the transformer is operating normally. A relay protection system, when a fault (such as a short circuit) occurs in the substation, the relay protection device can quickly act to cut off the faulty equipment and prevent the fault from expanding. Its configuration varies according to factors such as the wiring method and equipment type of the substation. For example, in a substation with a double-busbar connection, the relay protection system needs to consider the situation of busbar segmentation and set corresponding bus differential protection and line protection. An automation system is used for remote control and automated operation of substation equipment, such as remotely controlling the opening and closing of circuit breakers.

[0025] For another example, in a switching station power operation system, the switching station is mainly used for power distribution. The power operation system mainly includes a switchgear control system, which is used to control the opening and closing operations of each incoming and outgoing line switch to achieve power distribution and line switching. For example, when a 10kV distribution line needs to be repaired, the switching station disconnects the repaired line by controlling the corresponding switch, and at the same time adjusts the power supply mode through other switches to ensure the normal power supply of other users. At the same time, it also includes a corresponding monitoring system, which is used to monitor the status of the switchgear in the switching station, such as the opening and closing position of the switch, the energy storage status, etc., as well as basic electrical parameters such as bus voltage, to ensure the normal operation of the switching station.

[0026] Or a distribution line power operation system. Taking an overhead line as an example, the power operation system of the overhead line includes a line inspection system, which inspects equipment such as conductors, insulators, and poles of the line through manual inspection or drone inspection. For example, check whether the conductor is damaged, whether the insulator is damaged or has a pollution discharge phenomenon, and whether the pole is tilted. It also includes a line fault location system, which can quickly locate the fault point when a line fault (such as a ground fault or a short circuit fault) occurs. This can be achieved by installing fault indicators on the line. The fault indicator will send a signal after detecting the fault current, and the staff can find the fault point according to the signal indication for repair.

[0027] It also includes, for example, a distribution transformer power operation system. The power operation system of the distribution transformer includes a transformer monitoring system, which mainly monitors the operating parameters of the transformer, such as oil temperature, winding temperature, load current, etc. For example, for an oil-immersed distribution transformer, the oil temperature is a key parameter. If the oil temperature is too high, it may cause problems such as transformer insulation aging. It also includes a protection system for overload protection and temperature protection. When the transformer load exceeds a certain proportion of the rated capacity or the temperature is too high, the protection device will send an alarm signal or automatically cut off the power supply to protect the transformer.

[0028] The method includes: When it is necessary to operate on the distribution facilities, collect and fuse the multi-source data of the power operation system of the distribution facilities to obtain the multi-modal data of the power operation system.

[0029] Specifically, in the power operation system, there are various types of multi-source data, including but not limited to the operating parameters of equipment (such as voltage, current, power, etc.), the behavior data of operators (such as operation actions, movement trajectories, etc.), the meteorological data on site (such as temperature, humidity, wind speed, etc.), and the image or video data of equipment, etc. These multi-source data are collected through sensors, monitoring devices, and personnel records. Then, the collected data in different formats and types are fused and processed to integrate them into a unified and coherent multi-modal dataset. For example, the operating data of equipment is combined with the corresponding image data to more comprehensively reflect the actual state of the equipment on site. By collecting and fusing the multi-source data of the power operation system from multiple channels, multi-modal data is formed. Then, through fusion technology, the data is integrated into a single and consistent multi-modal data format, providing comprehensive basic information for subsequent analysis and decision-making. It solves the problem that traditional work ticket decision-making relies on manual experience to consult limited data, and provides comprehensive and rich basic information for subsequent analysis.

[0030] Through a large-scale language model, a data feature extraction model, and an image feature extraction model, according to the multi-modal data, the deep semantic features of the multi-modal data are obtained.

[0031] Specifically, first, for text-related data (such as equipment fault descriptions, work requirements, etc.), a large-scale language model is used to analyze it. Through learning a large amount of text corpora, the large-scale language model can understand the semantic information in the text, such as being able to identify semantic features such as the equipment fault type described in the text and the urgency of work requirements. Second, for structured data (such as equipment operating parameters, meteorological data, etc.), a data feature extraction model is used. This model can extract key data features from it, for example, by statistical analysis to extract features such as the mean and variance of equipment operating parameters, or to identify abnormal change features in meteorological data. Third, for image or video data, an image feature extraction model is used. This model can identify visual features such as objects in the image (such as equipment components, tools, etc.), scenes (such as indoor substations, outdoor transmission lines, etc.), and the state of objects (such as whether the equipment is damaged, etc.). By integrating the feature extraction of each part of the multi-modal data by the above three models, the deep semantic features of the multi-modal data are finally obtained. These deep semantic features can more deeply and accurately describe the actual situation of the power operation system, providing an information-rich basis for subsequent decision-making.

[0032] Process multimodal data using a large - scale language model, a data feature extraction model, and an image feature extraction model. Among them, the large - scale language model can understand the semantic information in the text description, the data feature extraction model can obtain key features from structured data, and the image feature extraction model can analyze the visual features in on - site images or videos. By integrating these three types of models, deep semantic features of the multimodal data are extracted, enabling the system to deeply understand the meaning behind the data rather than just staying on the surface of the data, thus providing deeper feature support for accurate decision - making.

[0033] Match according to the preset knowledge graph based on the deep semantic features to obtain knowledge graph nodes and relationships related to the multimodal data.

[0034] Specifically, the preset knowledge graph contains a large amount of knowledge in the field of power operations. Among them, the nodes represent various entities (such as equipment types, failure modes, safety measures, etc.), and the relationships represent the associations between these entities (for example, a certain equipment type is prone to a certain failure mode, and a certain failure mode requires specific safety measures, etc.). Match the extracted deep semantic features with the knowledge graph.

[0035] For example, if the deep semantic features contain a specific failure description of a certain piece of equipment, then in the knowledge graph, corresponding equipment failure nodes, as well as nodes related to failure handling methods, associated equipment, etc., can be found, and at the same time, the relationships between these nodes (such as the failure handling process, the electrical connection relationship between equipment, etc.) can be obtained.

[0036] Match according to the preset knowledge graph based on the extracted deep semantic features to find knowledge graph nodes and relationships related to the multimodal data. The knowledge graph integrates rich knowledge and experience related to power operations, such as the normal operating parameter range of equipment, safety measures corresponding to different operation types, etc. Matching the deep semantic features with the knowledge graph can link the current situation at the operation site with existing professional knowledge, provide a knowledge background for subsequent decision - making, overcome the problem of unstable decision - making caused by relying on manual experience and different understandings and judgments of different personnel in traditional methods, and assist decision - making with a unified and standardized knowledge system.

[0037] Determine the initial auxiliary decision of the work permit for power operations according to the knowledge graph nodes and relationships, in combination with a preset deep reinforcement learning network.

[0038] Specifically, the deep reinforcement learning network selects the optimal action strategy according to the current state (i.e., the on - site situation reflected by the nodes and relationships obtained from the knowledge graph). In this scenario, the deep reinforcement learning network generates a reasonable initial auxiliary decision for the work permit based on the knowledge provided by the knowledge graph and the actual situation on - site.

[0039] For example, the nodes and relationships in the knowledge graph indicate that there is a certain equipment failure on site and maintenance operations are required. The deep reinforcement learning network will, according to the preset reward mechanism (such as rewards for ensuring safety, rewards for improving efficiency, etc.), consider various possible operation methods (such as different power outage scopes, safety measure layout methods, etc.), and select a relatively optimal initial auxiliary decision-making plan. This plan includes content such as recommended work content, work time arrangement, personnel division of labor, and safety measures.

[0040] Based on the nodes and relationships of the knowledge graph and combined with the preset deep reinforcement learning network, determine the initial auxiliary decision for the work ticket of the power operation; the deep reinforcement learning network can learn the optimal decision-making strategy according to the current state (i.e., the matched nodes and relationships of the knowledge graph) and the preset reward mechanism. Furthermore, from numerous possible decision-making plans, select a relatively reasonable initial auxiliary decision, which can generate a preliminary plan more quickly and intelligently compared with the traditional manual decision-making method.

[0041] Input the initial auxiliary decision into the power operation system model for simulation to obtain the execution effect corresponding to the initial auxiliary decision.

[0042] Specifically, construct a power operation system model that can simulate the entire process of power operation, including aspects such as changes in the operating state of equipment, operation behaviors of operators, and implementation of safety measures. Input the initial auxiliary decision into this model, and the model will perform simulation operations according to the operation content, process, and safety measures specified in the decision. Through simulation, the execution effect of this initial auxiliary decision in the virtual environment can be obtained.

[0043] For example, it can be observed whether the equipment failure can be successfully eliminated in the simulation, whether the operation is completed within the specified time, whether the safety measures effectively prevent accidents and other execution effect information. This information can help evaluate the feasibility and effectiveness of the initial auxiliary decision.

[0044] Input the initial auxiliary decision into the power operation system model for simulation to obtain the execution effect corresponding to the initial auxiliary decision. Through simulation, before the actual operation, the operation situation of this decision in the virtual environment can be observed, potential problems can be discovered in advance, and the feasibility of the decision can be evaluated. This helps to avoid trial and error directly in the actual operation and reduce risks.

[0045] Optimize the initial auxiliary decision according to the execution effect and the expected execution effect to obtain the final auxiliary decision for the work ticket.

[0046] Specifically, the execution effect obtained by simulation is compared with the preset expected execution effect. The expected execution effect includes multiple aspects such as the job completion time meeting the planned requirements, the job process being safe without accidents, and the equipment returning to normal operation.

[0047] If there is a deviation between the execution effect of the initial auxiliary decision-making and the expected effect, for example, it is found in the simulation that the job time is too long or there are potential safety hazards, etc., then the initial auxiliary decision-making needs to be optimized and adjusted. By combining a feedback mechanism, such as adjusting relevant parameters in the decision-making according to the degree of deviation (such as increasing the number of personnel, changing the order of job steps, etc.), and conducting simulation evaluation again until the execution effect reaches or approaches the expected effect. The finally obtained work ticket auxiliary decision-making will be more reasonable and reliable, which helps to ensure the safety and efficiency of power operations.

[0048] For the work ticket auxiliary decision-making method and system of the present invention, since the distribution network contains multiple distribution facilities, and each facility has a corresponding power operation system, by collecting and integrating multi-source data of these different systems, it can be ensured that all relevant facility operation states and job requirements are comprehensively considered when formulating work tickets. For example, for different facilities such as substations, switchgear stations, distribution lines, and distribution transformers, data such as their electrical parameters, equipment status, and operating environment are obtained respectively to avoid missing key information.

[0049] Collect multi-source data of the power operation system. Through fusion processing, multi-modal data of the power operation system can be obtained. The multi-modal data can more completely present the on-site situation, providing comprehensive and rich basic information for subsequent analysis and decision-making, and solving the problem that traditional work ticket decision-making relies on manual experience to consult limited data and is difficult to deeply mine and accurately utilize historical data. Moreover, the multi-modal data provided by different power operation systems (such as text, images, time series data, etc.) can complement and verify each other, thereby improving the accurate assessment of the actual state of distribution facilities. For example, by combining image recognition technology to detect abnormalities in the appearance of equipment (such as damaged insulators), and at the same time combining the analysis of operating parameters (such as current, voltage), it can be more accurately judged whether the equipment needs repair or replacement.

[0050] Meanwhile, different power distribution facilities in the power distribution network (such as substations, switchgear stations, overhead lines, cable lines, etc.) have different functions and operation requirements. By collecting and analyzing data through the power operation system for each facility, the personalized operation needs of different types of facilities can be met. For example, the operation of a substation may involve complex relay protection and automation systems, while the operation of a power distribution line focuses more on line inspection and fault location. The power operation systems of different power distribution facilities can adapt to the complex and changeable power distribution network environment. For example, in the core urban area, industrial area, and rural area, there are significant differences in the configuration and operation conditions of power distribution facilities. By configuring an independent power operation system for each facility, the operation challenges in different environments can be better addressed, and the applicability and effectiveness of work tickets can be improved.

[0051] Then, by using a large-scale language model, a data feature extraction model, and an image feature extraction model to synthesize the three, the deep semantic features of multi-modal data are obtained, enabling the system to deeply understand the meaning behind the data, rather than just staying on the surface of the data as in traditional methods, thus providing deeper feature support for accurate decision-making. Then, by presetting a knowledge graph to match according to the deep semantic features, the knowledge graph nodes and relationships related to the multi-modal data are found. Matching the deep semantic features with the knowledge graph can link the situation of the current operation site with the existing professional knowledge, providing a knowledge background for subsequent decision-making. The knowledge graph provides a unified and standardized knowledge system, solving the problem of unstable decision-making caused by relying on manual experience and different understandings and judgments of different personnel in traditional methods. Based on the knowledge graph nodes and relationships, combined with a preset deep reinforcement learning network, the initial auxiliary decision of the work ticket for power operation is determined. The deep reinforcement learning network can learn the optimal decision-making strategy according to the current state (i.e., the matched knowledge graph nodes and relationships) and the preset reward mechanism. It can select a relatively reasonable initial auxiliary decision from numerous possible decision-making schemes. Compared with the traditional method of manually making decisions, it can generate a preliminary scheme more quickly and intelligently. Further, the initial auxiliary decision is input into the power operation system model for simulation to obtain the execution effect corresponding to the initial auxiliary decision. Through simulation operations, before actual operation, the operation situation of this decision in the virtual environment can be observed, potential problems can be discovered in advance, and the feasibility of the decision can be evaluated; it helps to avoid trial and error directly in actual operation and reduce risks. Finally, according to the execution effect and the expected execution effect, the initial auxiliary decision is optimized to obtain the final auxiliary decision of the work ticket. This process of the present invention uses a feedback mechanism to adjust the initial decision to make it more in line with the actual needs and expected goals, thereby improving the accuracy and consistency of decision-making, solving the problem of unstable decision-making results in traditional methods, and realizing the improvement of the intelligent level of work ticket auxiliary decision-making.

[0052] Optionally, when it is necessary to perform operations on the power distribution facilities, multi-source data of the power operation system of the power distribution facilities is collected and fused to obtain multi-modal data of the power operation system, including: when it is necessary to perform operations on the power distribution facilities, collecting text data, sensing data, image data, and equipment operation data in the power operation system of the power distribution facilities, and using the text data, the sensing data, the image data, and the equipment operation data as the multi-source data; converting the multi-source data into a preset format through data fusion technology, and fusing the multi-source data in the unified format to obtain the multi-modal data of the power operation system.

[0053] Specifically, in this optional embodiment, the text data usually comes from documents such as work logs, equipment maintenance records, and accident reports at the work site. These documents can be extracted from paper documents by manual entry or using optical character recognition (OCR) technology. For example, the work log records text description information such as the daily operation status of the equipment, minor faults that occur, and corresponding handling measures. The sensing data is collected in real time by various sensors installed on the power equipment. For example, current transformers and voltage transformers are used to monitor the current and voltage changes in the circuit respectively; temperature sensors monitor the temperature of key parts of the equipment; pressure sensors monitor oil pressure or air pressure, etc. These sensors convert physical quantities into electrical signals and then convert them into digital data through an analog-to-digital converter. The image data mainly comes from camera monitoring at the work site and appearance inspection images of the equipment. The camera can be a fixed-installed device for real-time monitoring of the work area or a portable shooting device carried by the staff to record the appearance state of the equipment, the operation behavior of the operators, etc. The equipment operation data includes the start / stop time, operation duration, load condition, fault shutdown time, etc. of the equipment. These data are automatically recorded by the control system or monitoring system of the equipment and stored in the form of structured data, such as table records in a database.

[0054] Convert the collected multi-source data into a unified preset format, which involves standardizing different types of data. Specifically, for text data, convert it into a unified text encoding format (such as UTF-8) and organize it into a structured text file or text fields in a database. For example, store information such as dates, event descriptions, and handling measures in work logs in different fields of the database for subsequent processing. For sensor data, which is usually time-series data, convert it into a unified timestamp format and standardize the dimensions and units of the data. For example, convert all temperature data to degrees Celsius as the unit and unify the timestamp in seconds. For image data, perform preprocessing operations such as image size adjustment and color space conversion to make it meet the requirements of a unified image format. For example, adjust all images to the RGB color space and unify the image resolution. For device operation data, convert it from the original recording format (which may be a device-specific binary file or proprietary data format) to a common structured data format, such as CSV (comma-separated values) or JSON (JavaScript Object Notation). After converting the multi-source data into a unified format, integrate these data through data fusion technology to generate multi-modal data.

[0055] In one embodiment, data fusion can be aligned based on the time dimension. For example, taking time as the reference, correlate events described in text data, numerical values of physical quantities monitored by sensor data, on-site images captured by image data, and operating states recorded by device operation data at the same time point or time period. In this way, a multi-modal data set that comprehensively reflects the situation at the work site can be obtained, where the data at each time point contains complementary information from different data sources. In another embodiment, data fusion can be aligned based on the space dimension. For example, combine sensor data at different locations with image data at that location to determine whether the operating state and appearance of a specific part of the device are consistent. If the temperature sensor at a certain device part shows an abnormally high temperature, and at the same time the image data of that part shows smoke or color change, then through spatial dimension fusion, it is possible to more accurately judge the possible faults of the device.

[0056] In this optional embodiment, by collecting multi-source data such as text data, sensing data, image data, and equipment operation data of the power operation system, and using data fusion technology to convert them into a unified format for fusion, multi-modal data is generated, significantly improving the integrity and accuracy of the data. This comprehensive data collection method ensures that subsequent decisions can be made based on rich and multi-dimensional information, avoiding one-sided decisions caused by data loss. At the same time, the unified format conversion eliminates the format differences between different data sources, facilitating subsequent data processing and analysis, and improving the efficiency and accuracy of data processing. More importantly, the multi-modal data fusion enhances the correlation between different data, enabling the events described in the text, the physical quantity changes monitored by the sensing data, the on-site pictures captured by the image data, and the operation status recorded by the equipment operation data to corroborate each other, more accurately reflecting the real situation of the power operation system, and providing a solid data foundation for the auxiliary decision-making of work tickets.

[0057] The multi-source data collection and fusion process of this embodiment provides strong support for the data-driven decision-making of work tickets, significantly improving the efficiency and quality of work ticket decision-making. In traditional methods, staff need to spend a lot of time searching and sorting information in different formats and scattered data sources, and are easily affected by incomplete or inaccurate data, resulting in low decision-making efficiency and uneven quality. However, the multi-modal data after collection and fusion can be directly used by the decision-making system, reducing the workload of manual data searching and sorting. At the same time, due to the improvement of data integrity and accuracy, the decision-making system can more accurately identify potential risks and patterns, thus generating more reasonable and reliable work ticket auxiliary decision-making schemes, effectively ensuring the safety and efficiency of power operations.

[0058] Optionally, obtaining the deep semantic features of the multi-modal data through a large language model, a data feature extraction model, and an image feature extraction model according to the multi-modal data includes: Inputting the multi-modal data into the large language model, and extracting the semantics of the text data in the multi-modal data through the large language model to obtain the semantic information of the multi-modal data; inputting the multi-modal data into the data feature extraction model, and extracting the features of the sensing data and the equipment operation data in the multi-modal data through the data feature extraction model to obtain the data features of the multi-modal data; inputting the multi-modal data into the image feature extraction model, and extracting the features of the image data in the multi-modal data through the image feature extraction model to obtain the image features of the multi-modal data; fusing the semantic information, the data features, and the image features to obtain the deep semantic features.

[0059] Specifically, first, the collected text data is preprocessed, including removing noise (such as meaningless symbols, white space characters, etc.), word segmentation (dividing continuous text into a sequence of words or phrases), and part-of-speech tagging (tagging the part of speech of each word, such as noun, verb, etc.).

[0060] Exemplarily, for the text description in the work log "The main transformer had an overload alarm at 08:30, and the duration was 15 minutes", after preprocessing, it is segmented into words such as "main transformer", "at", "08:30", "had", "overload alarm", "duration", "was", "15 minutes", etc., and the part of speech is tagged. The preprocessed text data is input into a large language model. Among them, the large language model in this embodiment is based on the Transformer architecture and includes multiple layers of self-attention mechanisms and feed-forward neural networks. The self-attention mechanism can capture the correlation between different words in the text, regardless of their distance in the text. For example, in the above text, there is an important semantic association between the "main transformer" and the "overload alarm", and the self-attention mechanism can calculate the correlation weight between them, enabling the model to understand that the "main transformer" is the object that had the "overload alarm".

[0061] By performing a pooling operation (such as taking the output vector of the first token in the text sequence, or performing average pooling on the output vectors of all tokens) on the output layer of the large language model, a fixed-dimensional vector representation is obtained, which is the semantic information vector of the text data. This vector can capture the key semantic features in the text data, such as the type of event described (equipment failure, maintenance operation, etc.), the equipment involved, and the severity of the event, etc. For the sensing data and device operation data, first, time alignment processing is performed to ensure that they have a unified timestamp format. Then, the data is normalized, and the values are scaled to a specific range (such as 0 to 1) to eliminate the influence of different dimensions and value ranges on feature extraction. For example, for temperature sensor data (unit: degree Celsius, range may be from -50°C to 100°C) and current sensor data (unit: ampere, range may be from 0A to 1000A), they are converted to the range of 0 to 1 through a normalization formula (such as min-max normalization).

[0062] The data feature extraction model can adopt a recurrent neural network architecture such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU) to process time series data. In this alternative embodiment, taking LSTM as an example, LSTM includes three structures: an input gate, a forget gate, and an output gate. The input gate controls how much of the input information at the current moment is written into the cell state; the forget gate determines how much of the cell state at the previous moment is retained; the output gate determines how much of the cell state at the current moment is output. In this way, LSTM can learn long-term dependencies in time series data. For sensing data and device operation data, LSTM can extract features such as the change trend, periodic pattern, and outliers of the data. For example, for the current monitoring data of a device, LSTM can capture abnormal features such as the rising trend of the current when the device starts up, the stable current value during normal operation, and possible sudden increases in the current.

[0063] Preprocess the collected image data, including resizing the image, such as uniformly scaling it to 224×224 pixels, performing color space conversion, such as converting from RGB to grayscale, or performing other color enhancement processes to highlight specific features, as well as data augmentation, such as randomly rotating and flipping the image to increase the generalization ability of the model. The image feature extraction model adopts a convolutional neural network (CNN) architecture. First, perform a convolution operation on the image through a convolutional layer, sliding multiple convolutional kernels on the image to extract local features of the image (such as edges, textures, etc.). For example, a 3×3 convolutional kernel sliding on the image can detect edge features in the image. Then introduce non-linearity through an activation function (such as ReLU) so that the model can learn more complex feature representations. Next, downsample the feature map through a pooling layer (such as max pooling) to reduce the amount of data and extract the main features. After multiple convolutional and pooling operations, the resulting feature map can represent the high-level semantic features of the image, such as the appearance shape of the device, whether there are signs of damage (such as cracks, deformations, etc.), and the relative positional relationship of the device components.

[0064] Fuse the semantic information vectors extracted by the large language model, the data feature vectors obtained by the data feature extraction model, and the image feature vectors obtained by the image feature extraction model. The concatenation fusion method can be used to connect the three vectors end to end to form a longer fusion vector. For example, assume that the dimension of the semantic information vector is 512, the dimension of the data feature vector is 256, and the dimension of the image feature vector is 1024. The dimension of the fused vector after concatenation is 512 + 256 + 1024 = 1792. Further, in order to make the fused features better applicable to subsequent tasks such as knowledge graph matching and decision-making, the fused vector can be input into a fully connected neural network layer to further integrate and reduce the dimension of the features. At the same time, normalization processing (such as L2 normalization) can be used to make the fused feature vector have a unit length, facilitating subsequent operations such as similarity calculation. After the above processing, a deep semantic feature vector of multimodal data is obtained, which can comprehensively represent various information of the power operation system, including the semantics of text descriptions, the features of device operation and environmental monitoring data, and the visual features of images, providing strong feature support for subsequent work ticket auxiliary decision-making.

[0065] In this alternative embodiment, deep semantic features of multimodal data are extracted through the large language model, the data feature extraction model, and the image feature extraction model, realizing targeted processing of different types of data. The large language model can deeply mine the semantic information in text data and accurately capture key semantic points such as events, objects, and their associations described in the text; the data feature extraction model focuses on extracting data features from sensor data and device operation data and grasping the laws such as data change trends, periodicity, and anomalies; the image feature extraction model can accurately extract visual features in image data and identify important information such as objects, scenes, and states in the images. On this basis, fusing the three types of features can integrate the semantics of the text, the laws of the data, and the visual content of the images, forming a comprehensive, in-depth, and unified feature description of the power operation system, that is, deep semantic features, providing a high-quality and information-rich feature basis for subsequent knowledge graph matching and auxiliary decision-making.

[0066] This alternative embodiment significantly improves the understanding and utilization efficiency of multimodal data. On the one hand, professional models are used for feature extraction for text, data, and images respectively, ensuring the accuracy and professionalism of feature extraction for each type of data and fully leveraging the information value contained in each type of data; on the other hand, the process of feature fusion breaks down the barriers between data types and integrates the scattered features into unified deep semantic features, enabling subsequent processing to take into account the comprehensive information conveyed by text, data, and images simultaneously, avoiding information fragmentation and incompleteness caused by processing each type of data separately, and thus providing a more accurate and comprehensive basis for work ticket auxiliary decision-making, which helps to improve the scientificity and reliability of decision-making.

[0067] Optionally, the matching of the preset knowledge graph according to the deep semantic features to obtain the knowledge graph nodes and relationships related to the multimodal data includes: performing vectorization processing on the deep semantic features to obtain a feature vector of the deep semantic features; embedding the entities and relationships in the preset knowledge graph into a vector space through graph embedding technology to obtain knowledge graph embedding vectors; calculating the cosine similarity between the knowledge graph embedding vectors and the feature vectors; and determining the knowledge graph nodes and relationships related to the multimodal data according to the cosine similarity.

[0068] Specifically, the fused deep semantic feature vector is normalized to have a length of 1 to facilitate subsequent similarity calculation. For example, L2 normalization is used. Calculate the L2 norm of the vector (i.e., the square root of the sum of the squares of the vector elements), and then divide each element by this norm to obtain the normalized feature vector.

[0069] Suppose the deep semantic feature vector is , and its L2 norm is , and the normalized vector is ; After obtaining the normalized deep semantic feature vector, this vector contains the key information of the multimodal data of the power operation system, and each dimension corresponds to a specific semantic or feature meaning. For example, a certain dimension may reflect the severity of equipment failure, and another dimension may reflect the stability of the equipment operation state, etc. The preset knowledge graph includes various entities in the power operation system (such as equipment types, failure modes, safety measures, operation processes, etc.) and the relationships between entities (such as "causes", "requires", "belongs to", etc.). For example, there is a "causes" relationship between the entity "circuit breaker" and "overload fault", and there is a "requires" relationship between the entity "overload fault" and "increase cooling". The TransE algorithm is used for graph embedding. Initialize each entity and relationship as a random vector, and then optimize the objective function through training so that the embedding vectors of the correct triples (head entity, relationship, tail entity) in the knowledge graph satisfy a certain distance constraint. For example, for a triple such as "circuit breaker", "causes", "overload fault", after training, it satisfies that the head entity vector plus the relationship vector is approximately equal to the tail entity vector, that is .

[0070] In this way, the entities and relationships in the knowledge graph are embedded into a low-dimensional vector space to obtain a set of knowledge graph embedding vectors. Calculate the cosine similarity between the embedding vectors of each entity and relationship in the knowledge graph and the deep semantic feature vector.

[0071] Among them, the calculation formula of the cosine similarity is: , where a is the knowledge graph embedding vector and b is the deep semantic feature vector.

[0072] Since both are normalized and the denominator is 1, the calculation can be simplified to a vector dot product. For example, if the dot product of the embedding vector of the "overload fault" entity in the knowledge graph and the deep semantic feature vector is 0.85, it indicates a high similarity between the two. A similarity matrix is constructed, where the rows represent the entities and relationships in the knowledge graph, the columns represent the deep semantic features to be matched, and the matrix elements are the corresponding cosine similarity values, providing a quantitative basis for determining relevant knowledge graph nodes and relationships subsequently. Then, according to actual requirements and experience, a cosine similarity threshold is set, such as 0.7. The entities and relationships in the similarity matrix that are higher than this threshold are determined as the knowledge graph nodes and relationships related to the multimodal data. The nodes and relationships with similarity higher than the threshold are sorted from high to low according to similarity, and the top K (such as the top 5) are selected as the final matching results. These results are the information in the knowledge graph that is most relevant to the current power operation system's multimodal data, including specific entities and the relationships between them, and will provide professional knowledge support for subsequent work order auxiliary decision-making.

[0073] In this optional embodiment, the preset knowledge graph specifically includes various types of equipment and their parameters, operating states, and location information in power operations, such as the rated voltage, current, power, and operating temperature range of equipment such as transformers and circuit breakers; includes possible fault types, phenomena, causes, and impacts, such as short circuits, ground faults, overload faults, etc. and the resulting equipment damage and power supply impacts; covers the safety measures of operators, such as safety tools, operating specifications, personal protective equipment, and safety distances; clarifies specific operation procedures, such as the steps, sequences, and related systems of equipment inspection, fault location, repair operation, and test verification; and also involves environmental parameters, such as meteorological conditions and work site conditions.

[0074] Exemplarily, the transformer causes an overload fault due to overload, and the circuit breaker causes a ground fault due to insulation damage. In case of a short circuit fault, emergency power outage operations and setting multiple grounding wires are required. In case of a ground fault, warning signs need to be hung and insulated boots need to be worn for inspection. Overload fault prevention measures include adding cooling equipment and regularly checking the equipment load. And the equipment inspection operation includes wearing a safety helmet and arc-proof clothing, and the repair operation includes setting up a safety fence and hanging warning signs. The operation sequence of setting the grounding wire is after the power outage operation, and the removal sequence of the grounding wire is after the work is completed and accepted. The objects of the equipment inspection operation are faulty equipment such as transformers and circuit breakers, and the objects of the test verification operation are the repaired faulty equipment. The repair operation can change the equipment state from "faulty" to "normal", and the power outage operation changes the equipment state from "running" to "power off". Regarding the impact of the environment on the equipment, for example, a high-temperature environment affects the occurrence probability of equipment overheating faults, and a humid environment increases the risk of equipment leakage.

[0075] In this alternative embodiment, through such a knowledge graph, accurate matching of the deep semantic features of multimodal data can be achieved, providing rich background knowledge and reasoning basis for the work ticket auxiliary decision-making system, thereby effectively solving the problem of intelligent work ticket formulation in power operations and improving the scientificity and safety of decision-making. Specifically, by presetting the knowledge graph to perform matching according to deep semantic features, the knowledge graph nodes and relationships related to multimodal data are obtained, realizing the effective docking of deep semantic features and the knowledge graph. By performing vectorization processing on the deep semantic features, they can be transformed into a mathematical representation form isomorphic to the knowledge graph embedding vectors, making the two comparable. At the same time, the graph embedding technology is used to embed the entities and relationships in the knowledge graph into the vector space, fully retaining the topological structure and semantic information of the knowledge graph, laying a foundation for subsequent similarity calculation. The method of calculating the cosine similarity can accurately measure the semantic similarity between the feature vector and the knowledge graph embedding vector, and its result is intuitive and easy to interpret. By setting a reasonable similarity threshold and sorting and screening mechanism, the knowledge graph nodes and relationships related to multimodal data can be accurately determined, effectively filtering out irrelevant information, and improving the accuracy and pertinence of the matching.

[0076] Optionally, determining the initial auxiliary decision of the work ticket for power operations according to the knowledge graph nodes and relationships, in combination with a preset deep reinforcement learning network, includes: taking the risk assessment level, safety measure selection, and work process arrangement of the power operation as the actions of the deep reinforcement learning network, and taking the knowledge graph nodes and relationships as the states of the deep reinforcement learning network; splicing the embedding vectors of the knowledge graph nodes and relationships with the feature vectors of the work site environment parameters to form a state vector and inputting the state vector into the deep reinforcement learning network; selecting to update the actions according to the state vector, and judging whether the update of the actions is completed according to a preset reward function; taking the updated actions as the initial auxiliary decision of the work ticket.

[0077] Specifically, in the deep reinforcement learning network, the risk assessment level of the power operation is taken as part of the action. The risk assessment level can be divided into multiple levels, such as low risk, medium risk, high risk, etc. Each risk level corresponds to a different decision-making strategy. In specific implementation, the risk assessment level can be transformed into a discrete action space. For example, integer values are used to represent different risk levels (such as 0 for low risk, 1 for medium risk, 2 for high risk). The selection of safety measures is taken as another part of the action. Safety measures can include various specific safety operations, such as wearing safety equipment, setting warning signs, power outage operations, etc. These safety measures can form a discrete set of actions. For example, an action space containing multiple safety measures is defined, and each safety measure corresponds to a unique identifier (such as 0 for wearing a safety helmet, 1 for setting up a fence, etc.). Also, the work process arrangement is taken as part of the action. The work process can include multiple steps, such as equipment inspection, fault handling, equipment repair, etc. These steps can form a serialized action space. In specific implementation, the work process can be decomposed into multiple sub-steps, and an action is defined for each sub-step. For example, an integer sequence is used to represent different stages of the work process (such as 0 for start, 1 for equipment inspection, 2 for fault handling, etc.). Entities and relationships in the knowledge graph are embedded into the vector space through graph embedding technology (such as the TransE algorithm). The TransE algorithm makes the embedding vectors of correct triples (head entity, relationship, tail entity) in the knowledge graph satisfy certain distance constraints through training. Among them, the work site environment parameters can include data collected by sensors (such as temperature, humidity, voltage, etc.) and equipment operation data (such as equipment status, operation time, etc.). These parameters can be transformed into feature vectors after preprocessing (such as normalization). For example, the values of temperature, humidity, voltage, etc. are normalized to the range of 0 to 1, and these values are concatenated into a vector. The embedding vectors of the knowledge graph nodes and relationships are concatenated with the feature vectors of the work site environment parameters to form a unified state vector.

[0078] Exemplarily, in specific implementation, the two vectors can be concatenated end to end. For example, the knowledge graph embedding vector is , and the feature vector of the work site environment parameters is .

[0079] In a preferred embodiment, the reinforcement learning network can adopt a Deep Q-Network (DQN) architecture. The network input layer receives the state vector s, and the hidden layer contains multiple fully connected layers, with each layer using the ReLU activation function. For example, the first hidden layer has 128 neurons, and the second hidden layer has 64 neurons. The number of neurons in the output layer corresponds to the size of the action space, and each neuron outputs the Q-value corresponding to an action, representing the expected reward for executing that action in the current state. At each decision step, the state vector is input into the DQN to obtain the Q-values of all actions. The action with the highest Q-value is selected as the current decision. According to a preset reward function, it is judged whether the action update is completed. The reward function can be defined as follows: a positive reward (such as +1) is given when the selected action effectively reduces risk and complies with safety regulations; a negative reward (such as -1) is given when the action leads to an increase in risk or violates safety regulations; and a zero reward is given in other cases. During the training process, an experience replay buffer is used to store a number of recent (state, action, reward, next state) quadruples. In each training iteration, a small batch of samples is randomly drawn from the experience replay buffer for updating the parameters of the DQN. The target network is updated regularly to generate stable target Q-values to guide network training. After training is completed, the current state vector is input into the trained DQN to obtain the optimal action, which is the initial auxiliary decision for the work permit of electric power operation. The initial auxiliary decision includes the risk assessment level, the selection of safety measures, and the work process arrangement, guiding the operators to perform tasks according to the decision.

[0080] In this alternative embodiment, by determining the initial auxiliary decision for the work permit of electric power operation through a deep reinforcement learning network, the precise control and optimization of the key elements of electric power operation are achieved; the risk assessment level, the selection of safety measures, and the work process arrangement are defined as the actions of the deep reinforcement learning network, ensuring that the decision-making scheme can comprehensively cover the core links of electric power operation, and each action is closely related to the safety and efficiency of the operation. At the same time, taking the knowledge graph nodes and relationships as the state input fully integrates the professional knowledge in the field of electric power operation and the actual on-site situation, providing rich decision-making basis for the network. By splicing the knowledge graph embedding vector and the feature vector of the on-site environment parameters to form the state vector, the organic integration of multi-source information is realized, enabling the network to make decisions based on comprehensive state information. Through continuous learning and optimization, the deep reinforcement learning network can automatically generate the optimal initial auxiliary decision in different on-site environments, avoiding decision-making errors caused by experience differences or incomplete information in traditional manual decision-making, and improving the consistency and reliability of decision-making.

[0081] Optionally, simulating the initial auxiliary decision in the power operation system model to obtain the execution effect corresponding to the initial auxiliary decision includes: inputting the initial auxiliary decision into the power operation system model, and performing simulation execution according to the risk assessment level, the safety measure selection, and the work process arrangement; after the simulation execution is completed, obtaining the safety index, the efficiency index, and the specification index of the power operation system model after the simulation execution; and using the safety index, the efficiency index, and the specification index as the execution effect.

[0082] Specifically, the power operation system model is a virtual simulation environment based on digital twin technology, including a virtual model of power equipment, a virtual agent of operating personnel, a virtual execution module of the work process, and a virtual simulation module of environmental conditions. For example, for a substation operation scenario, the system model includes virtual models of equipment such as transformers, circuit breakers, and switch cabinets, as well as virtual agents of operating personnel for simulating operation behaviors in actual operations. The virtual model of the equipment is constructed based on the actual parameters and operating characteristics of the equipment, such as parameters of the transformer like power, current, voltage, etc., and its operating state under different load conditions.

[0083] The virtual agent of the operating personnel has the ability to simulate human behaviors, such as moving, operating equipment, and implementing safety measures. These models are implemented through programming languages (such as Python or C++) and simulation engines (such as Unity or Unreal Engine). The virtual execution module of the work process defines each step and process of the operation task, such as equipment inspection, fault handling, repair operations, etc. The virtual simulation module of environmental conditions can simulate the influence of environmental factors such as temperature, humidity, and wind speed on the operation. These modules are set by writing scripts and configuration files to ensure the accuracy and authenticity of the simulation. The initial auxiliary decision (including the risk assessment level, the safety measure selection, and the work process arrangement) is passed as input parameters to the power operation system model. For example, the risk assessment level is "medium risk", the safety measure selection is "wear a safety helmet and set up a grounding wire", and the work process arrangement is "equipment inspection → fault location → repair operation → test verification".

[0084] The system model performs simulation execution according to the input decision parameters. The virtual agent of the operating personnel performs corresponding operations according to the work process arrangement, and the virtual model of the equipment responds according to the operations and environmental conditions. For example, in the simulation, the virtual agent first checks the equipment status, then wears a safety helmet and sets up a grounding wire according to the safety measure requirements, and then performs fault location and repair operations.

[0085] During the simulation execution, the system model monitors and records various index data in real time. Safety indicators include whether equipment operations comply with safety specifications, whether operators avoid dangerous areas, etc.; efficiency indicators include work completion time, equipment downtime, etc.; specification indicators include whether operation steps conform to standard processes, etc. These data are collected and stored through sensor simulation and log recording functions. The evaluation of safety indicators includes two aspects: equipment safety and personnel safety. For example, equipment safety indicators can be evaluated by monitoring whether the equipment is overloaded, whether a failure occurs, etc.; personnel safety indicators can be evaluated by checking whether the virtual agent enters a dangerous area, whether safety equipment is correctly worn, etc. Additionally, during actual application operations, a safety violation counter can be set, and the counter is incremented by 1 whenever a safety violation occurs. After the simulation ends, the number of violations is counted as the quantified value of the safety indicator.

[0086] The evaluation of efficiency indicators includes work completion time, resource utilization rate, etc. For example, the work completion time can be calculated by recording the time from the start to the end of the simulation; the resource utilization rate can be evaluated by monitoring the usage time of equipment and the workload of personnel. Additionally, during actual application operations, timestamps and resource usage logs can be used to record relevant data, and then analysis and calculation are performed. The evaluation of specification indicators mainly checks whether the work process conforms to the preset standards and specifications. For example, it can be evaluated by checking whether the execution order of each step in the work process is correct, whether key steps are omitted, etc. In specific implementation, the actual executed work process can be compared with the preset standard process, and the number of conforming steps and the number of steps with incorrect order are counted as the quantified value of the specification indicator. The safety indicators, efficiency indicators, and specification indicators obtained after the simulation execution are sorted out and analyzed. For example, a report containing the numerical values of each indicator is generated, such as the number of safety violations is 0, the work completion time is 45 minutes, the proportion of steps conforming to the standard process is 90%, etc.

[0087] In this optional embodiment, by simulating the initial auxiliary decision in the power operation system model, the virtual execution and effect evaluation of the power operation plan are realized. During the simulation execution process, the system model operates according to the risk assessment level, safety measure selection, and work process arrangement in the initial auxiliary decision, comprehensively testing the feasibility and effectiveness of the decision in actual operations. After the simulation execution is completed, the safety index, efficiency index, and specification index of the system are obtained, and these indexes can accurately and quantitatively reflect the effect after the decision execution. For example, the safety index can reveal potential risks and safety loopholes during the operation process, the efficiency index can reflect the smoothness of the work process and the resource utilization efficiency, and the specification index can verify whether the operation complies with the established safety and operation specifications. Furthermore, it provides data support and intuitive feedback for the optimization of power operation decisions. Through simulation execution, potential problems and risks in the initial auxiliary decision can be discovered in a timely manner before actual operations, avoiding trial and error directly in actual operations, thereby reducing safety accidents and economic losses that may be caused by decision-making mistakes. In addition, the quantitative evaluation of safety, efficiency, and specification indexes provides a clear direction and basis for the adjustment and optimization of decisions, helps to formulate a more scientific and reasonable work ticket decision, improves the overall safety and efficiency of power operations, and ensures that power operations are carried out on the track of safety, efficiency, and standardization.

[0088] Optionally, optimizing the initial auxiliary decision according to the execution effect and the expected execution effect to obtain the final auxiliary decision of the work ticket includes: comparing the expected execution effect with the execution effect to determine whether the initial auxiliary decision meets the operation requirements; when the initial auxiliary decision meets the operation requirements, taking the initial auxiliary decision as the final auxiliary decision; when the initial auxiliary decision does not meet the operation requirements, optimizing the initial auxiliary decision according to the comparison result of the expected execution effect and the execution effect, and taking the optimized initial auxiliary decision as the final auxiliary decision.

[0089] Specifically, compare the safety, efficiency, and specification indexes in the execution effect with the expected values. The expected value of the safety index, such as the number of safety violations is 0, the closer the actual value is to 0, the better the decision. The expected value of the efficiency index, such as the work completion time ≤ 30 minutes, the smaller the actual value, the better. The expected value of the specification index, such as the process compliance rate ≥ 95%, the higher the actual value, the better. And set thresholds for each index, such as the number of safety violations ≤ 1 time, the work completion time ≤ 45 minutes, and the process compliance rate ≥ 90%. If the actual value is within the threshold range, it is considered that the initial decision meets the operation requirements; otherwise, it is considered not to meet.

[0090] When the initial auxiliary decision does not meet the operation requirements, it is optimized according to the comparison result. For example, if the risk assessment is too high, safety measures can be increased or the work process can be adjusted. Specifically, it includes: increasing safety measures, such as adding measures like "setting double grounding wires" and "increasing on-site safety supervisors" in operations with a "high-risk" risk assessment level; adjusting the work process, such as advancing the "equipment inspection" step in the work process and adding a "safety recheck" link. The optimized decision needs to be input into the power operation system model again for simulation until the simulation result meets the operation requirements, forming a closed-loop optimization process to ensure the scientificity and feasibility of the final decision. The optimized auxiliary decision is finally verified, input into the power operation system model again for simulation execution, and it is confirmed whether it meets the operation requirements. When the optimized auxiliary decision meets the operation requirements through simulation verification, it is determined as the final auxiliary decision for the work ticket. It includes content such as the risk assessment level, safety measure selection, and work process arrangement, forming a complete work ticket decision-making plan.

[0091] In this optional embodiment, by quantifying the comparison indicators and setting thresholds, it is possible to objectively and accurately judge whether the initial auxiliary decision meets the operation requirements, avoiding the uncertainty brought by subjective judgment. Secondly, the optimization process has strong pertinence and operability. By adjusting the initial auxiliary decision according to the comparison result, it can quickly and effectively improve the scientificity and rationality of the decision. Moreover, the closed-loop optimization process ensures the reliability of the final decision. Through multiple simulation verifications and continuous optimization and adjustment until the decision meets the operation requirements, it provides a solid guarantee for power operations. This process not only improves the quality of the work ticket decision-making but also lays a foundation for the safe and efficient execution of power operations, having important practical application value.

[0092] Optionally, the method of judging whether the initial auxiliary decision meets the operation requirements by comparing the expected execution effect with the execution effect includes: judging whether the safety index, the efficiency index, and the specification index all meet the corresponding expected execution effects; if the safety index, the efficiency index, and the specification index all meet the corresponding expected execution effects, it is determined that the initial auxiliary decision meets the operation requirements; if any one of the safety index, the efficiency index, and the specification index does not meet the corresponding expected execution effect, it is determined that the initial auxiliary decision does not meet the operation requirements.

[0093] Specifically, the safety indicators are judged by preset thresholds. Based on the safety specifications and historical data of power operations, the thresholds of safety indicators are set. For example, the threshold of the number of safety violations is set to 0 times, the threshold of the equipment stable operation indicator is the equipment operation status code 1 (indicating normal operation), and the threshold of the duration ratio of the operator in the safe area is 100%. The actual number of safety violations, the equipment operation status code, and the duration ratio of the operator in the safe area are obtained from the simulation execution results. For example, after the simulation execution, the number of safety violations is 0 times, the equipment operation status code is 1, and the duration ratio of the operator in the safe area is 100%. If the actual number of safety violations does not exceed the threshold (i.e., 0 times), the equipment operation status code is equal to 1, and the duration ratio of the operator in the safe area is equal to 100%, it is determined that the safety indicators meet the requirements. Moreover, based on the efficiency requirements and historical data of power operations, the efficiency indicator thresholds are set. For example, the set working completion time is within 30 minutes, and the average value of the equipment operation interval time is within 5 minutes. The actual working completion time and the average value of the equipment operation interval time are obtained from the simulation execution results. For example, after the simulation execution, the working completion time is 25 minutes, and the average value of the equipment operation interval time is 4 minutes. If the actual working completion time is less than or equal to 30 minutes and the average value of the equipment operation interval time is less than or equal to 5 minutes, it is determined that the efficiency indicators meet the requirements. Finally, based on the operation specifications and process requirements of power operations, the thresholds of the specification indicators are set. For example, after the simulation execution, the matching degree of the operation step sequence is 100%, and the proportion of operations meeting the specification requirements is 100%. If the actual matching degree of the operation step sequence is equal to 100% and the proportion of operations meeting the specification requirements is equal to 100%, it is determined that the specification indicators meet the requirements.

[0094] Generally speaking, when the safety indicators, efficiency indicators, and specification indicators all meet their respective expected execution effects, it can be determined that the initial auxiliary decision-making meets the operation requirements. This judgment method is comprehensive and rigorous, and can effectively ensure the safety, efficiency, and standardization of power operations. If any one of the indicators does not meet the requirements, it indicates that there are deficiencies in the initial auxiliary decision-making and further optimization and adjustment are needed until all indicators reach the expected effects, so as to ensure the reliability and effectiveness of the final auxiliary decision-making.

[0095] In this alternative embodiment, by quantitatively comparing safety, efficiency, and compliance indicators with the expected values, it is possible to accurately evaluate whether the initial auxiliary decision meets the operation requirements. The evaluation of the safety indicator focuses on risk avoidance to ensure the safety of operators and equipment. The evaluation of the efficiency indicator focuses on whether the operation can be efficiently completed within the specified time. The evaluation of the compliance indicator verifies whether the operation follows the established procedures and specifications. If all indicators meet the expectations, the initial auxiliary decision meets the operation requirements and can be directly used as the final decision. Once an indicator does not meet the expectations, it can be determined that the decision needs to be optimized. This alternative embodiment provides clear evaluation criteria and bases. Through the quantitative comparison of specific indicators, the subjectivity and ambiguity of human judgment are eliminated, making the decision evaluation more objective and accurate. A comprehensive review of the initial auxiliary decision is achieved, covering key aspects such as safety, efficiency, and compliance, ensuring the scientificity and reliability of the final decision. Potential problems can be identified in a timely manner in this embodiment, preventing the application of risky decisions to actual operations, providing a strong guarantee for the safe and efficient execution of power operations, and having important practical application value.

[0096] Optionally, optimizing the initial auxiliary decision according to the comparison result between the expected execution effect and the execution effect includes: when the initial auxiliary decision does not meet the operation requirements, determining the index gap between the index that does not meet the expected execution effect in the execution effect and the expected execution effect corresponding to the index, and using the index gap as the comparison result; modifying the initial auxiliary decision in combination with a preset adjustment rule according to the index gap to obtain the final auxiliary decision.

[0097] Specifically, when the initial auxiliary decision-making does not meet the operation requirements, first determine the indicators in the execution effect that do not meet the expected execution effect. Specifically: If the expected number of safety violations is 0 times, but 1 violation occurs in the actual simulation execution, the gap in the safety indicator is 1 time. If the expected work completion time is 30 minutes, but the actual simulation execution takes 35 minutes, the gap in the efficiency indicator is 5 minutes. If the expected compliance rate of operation steps is 100%, but the compliance rate is only 90% in the actual simulation execution, the gap in the specification indicator is 10%. By comparing the simulation execution results with the preset expected execution effect, the gap is obtained, and the specific value represents the degree of difference between the actual execution and the expected value. Set the preset adjustment rules according to the indicator type and the size of the gap. Exemplarily, for the safety indicator adjustment rule, if the safety indicator gap is 1 violation, additional safety measures can be added to the initial auxiliary decision-making, such as adding safety supervisors, increasing safety training sessions, or adding protective equipment in specific areas. For the efficiency indicator adjustment rule, if the efficiency indicator gap is 5 minutes, the work process can be optimized. For example, adjust the equipment operation sequence, add operating personnel, or improve the equipment operation efficiency. Specifically, an additional operating personnel can be added to share the workload, or the equipment operation sequence can be adjusted to reduce unnecessary waiting time. For the specification indicator adjustment rule, if the specification indicator gap is 10%, the work process can be re-reviewed and revised, the operation specification training can be strengthened, or a confirmation link can be added to key operation steps. For example, add a two-person confirmation link after each key operation step to ensure that the operation meets the specification requirements. The modified initial auxiliary decision-making (i.e., the optimized decision) needs to be input into the power operation system model again for simulation verification: input the optimized decision into the power operation system model, re-perform the simulation execution, and obtain new safety, efficiency, and specification indicators. Compare the new execution effect with the expected execution effect again. If all indicators meet the expected values, determine this optimized decision as the final auxiliary decision. If there are still indicators that do not meet the requirements, repeat the above process, determine the indicator gap again according to the comparison result of the new execution effect and the expected effect, and modify it according to the preset adjustment rules until the optimized decision meets the operation requirements.

[0098] In this alternative embodiment, through quantitative comparison and targeted adjustment, continuous optimization of the initial auxiliary decision-making is achieved, ensuring that the final decision meets the operation requirements. First, by clarifying the index gap, the deficiencies in the decision-making can be accurately located, providing a specific direction for subsequent adjustments. Second, the preset adjustment rules provide a clear operation guide for decision modification, ensuring that the adjustment process is rule-based and avoiding blindness. Finally, through re-simulation verification and iterative optimization, a closed-loop optimization process is formed, which can continuously improve the decision-making scheme until it meets all operation requirements. This optimization method based on data-driven and rule-guided not only improves the scientificity and rationality of decision-making but also provides a strong guarantee for the safe and efficient execution of power operations.

[0099] As Figure 3 shown, the present invention also provides a work ticket auxiliary decision-making system. The system is applied to a distribution network, and the distribution network includes a plurality of distribution facilities. Each distribution facility corresponds to a different power operation system. The system includes: a data processing unit for collecting and fusing multi-source data of the power operation system of the distribution facility when an operation needs to be performed on the distribution facility to obtain multi-modal data of the power operation system; a feature extraction unit for obtaining deep semantic features of the multi-modal data according to the multi-modal data through a large-scale language model, a data feature extraction model, and an image feature extraction model; a matching unit for matching according to the deep semantic features through a preset knowledge graph to obtain knowledge graph nodes and relationships related to the multi-modal data; an initial decision-making unit for determining an initial auxiliary decision of the work ticket for the power operation according to the knowledge graph nodes and relationships in combination with a preset deep reinforcement learning network; a simulation unit for inputting the initial auxiliary decision into a power operation system model for simulation to obtain an execution effect corresponding to the initial auxiliary decision; and an optimization unit for optimizing the initial auxiliary decision according to the execution effect and the expected execution effect to obtain a final auxiliary decision of the work ticket.

[0100] The advantages of the work ticket auxiliary decision-making system of the present invention compared with the prior art are the same as those of the above work ticket auxiliary decision-making method compared with the prior art, and will not be elaborated here.

[0101] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will all fall within the protection scope of the present invention.

Claims

1. A work ticket auxiliary decision-making method, characterized in that The method is applied to a distribution network, which includes multiple distribution facilities, and each distribution facility corresponds to a different power operation system. The method includes: When an operation needs to be performed on the distribution facility, multi-source data of the power operation system of the distribution facility is collected and fused to obtain multi-modal data of the power operation system; Through a large-scale language model, a data feature extraction model, and an image feature extraction model, deep semantic features of the multi-modal data are obtained according to the multi-modal data; Through a preset knowledge graph, matching is performed according to the deep semantic features to obtain knowledge graph nodes and relationships related to the multi-modal data; According to the knowledge graph nodes and relationships, combined with a preset deep reinforcement learning network, an initial auxiliary decision for the work ticket of the power operation is determined; The initial auxiliary decision is input into the power operation system model for simulation to obtain the execution effect corresponding to the initial auxiliary decision; According to the execution effect and the expected execution effect, the initial auxiliary decision is optimized to obtain the final auxiliary decision of the work ticket.

2. The work ticket auxiliary decision-making method according to claim 1, wherein The step of when an operation needs to be performed on the distribution facility, collecting and fusing multi-source data of the power operation system of the distribution facility to obtain multi-modal data of the power operation system includes: When an operation needs to be performed on the distribution facility, text data, sensing data, image data, and equipment operation data in the power operation system of the distribution facility are collected, and the text data, the sensing data, the image data, and the equipment operation data are used as the multi-source data; Through data fusion technology, the multi-source data is converted according to a preset format, and the multi-source data after being unified in format is fused to obtain the multi-modal data of the power operation system.

3. The work ticket auxiliary decision-making method according to claim 2, characterized in that The step of obtaining deep semantic features of the multi-modal data through a large-scale language model, a data feature extraction model, and an image feature extraction model according to the multi-modal data includes: The multi-modal data is input into the large-scale language model, and the large-scale language model performs semantic extraction on the text data in the multi-modal data to obtain semantic information of the multi-modal data; The multi-modal data is input into the data feature extraction model, and the data feature extraction model performs feature extraction on the sensing data and the equipment operation data in the multi-modal data to obtain data features of the multi-modal data; The multi-modal data is input into the image feature extraction model, and the image feature extraction model performs feature extraction on the image data in the multi-modal data to obtain image features of the multi-modal data; The semantic information, the data features, and the image features are fused to obtain the deep semantic features.

4. The work ticket auxiliary decision-making method according to claim 1, characterized in that, The step of obtaining knowledge graph nodes and relationships related to the multi-modal data by matching according to the deep semantic features through a preset knowledge graph includes: Performing vectorization processing on the deep semantic features to obtain a feature vector of the deep semantic features; Embed the entities and relationships in the preset knowledge graph into a vector space through graph embedding technology to obtain knowledge graph embedding vectors; Calculate the cosine similarity between the knowledge graph embedding vectors and the feature vectors; Determine the knowledge graph nodes and relationships related to the multimodal data according to the cosine similarity; 5. The work ticket auxiliary decision-making method according to claim 1, characterized in that, The initial auxiliary decision of the work ticket for the power operation is determined according to the knowledge graph nodes and relationships, in combination with a preset deep reinforcement learning network, including: Take the risk assessment level, safety measure selection, and work process arrangement of the power operation as the actions of the deep reinforcement learning network, and take the knowledge graph nodes and relationships as the states of the deep reinforcement learning network; Concatenate the embedding vectors of the knowledge graph nodes and relationships with the feature vectors of the work site environment parameters to form a state vector and input it into the deep reinforcement learning network; Select to update the actions according to the state vector, and judge whether the update of the actions is completed according to a preset reward function; Take the updated actions as the initial auxiliary decision of the work ticket; 6. The work ticket auxiliary decision-making method according to claim 5, characterized in that, The initial auxiliary decision is input into the power operation system model for simulation to obtain the execution effect corresponding to the initial auxiliary decision, including: Input the initial auxiliary decision into the power operation system model and perform simulation execution according to the risk assessment level, the safety measure selection, and the work process arrangement; After the simulation execution is completed, obtain the safety index, efficiency index, and specification index of the power operation system model after the simulation execution; Take the safety index, the efficiency index, and the specification index as the execution effect; 7. The work ticket auxiliary decision-making method according to claim 6, characterized in that The initial auxiliary decision is optimized according to the execution effect and the expected execution effect to obtain the final auxiliary decision of the work ticket, including: Judge whether the initial auxiliary decision meets the operation requirements by comparing the expected execution effect with the execution effect; When the initial auxiliary decision meets the operation requirements, take the initial auxiliary decision as the final auxiliary decision; When the initial auxiliary decision does not meet the operation requirements, optimize the initial auxiliary decision according to the comparison result of the expected execution effect and the execution effect, and take the optimized initial auxiliary decision as the final auxiliary decision; 8. The work ticket auxiliary decision-making method according to claim 7, wherein The judgment of whether the initial auxiliary decision meets the operation requirements by comparing the expected execution effect with the execution effect includes: Judge whether the safety index, the efficiency index, and the specification index all meet the corresponding expected execution effects; If the safety index, the efficiency index, and the specification index all meet the corresponding expected execution effects, it is determined that the initial auxiliary decision meets the operation requirements; If any one of the safety index, the efficiency index, and the specification index does not meet the corresponding expected execution effect, it is determined that the initial auxiliary decision does not meet the operation requirements; 9. The work ticket auxiliary decision-making method according to claim 7, wherein The optimization of the initial auxiliary decision according to the comparison result of the expected execution effect and the execution effect includes: When the initial auxiliary decision does not meet the operation requirements, determine the index gap between the index that does not meet the expected execution effect in the execution effect and the expected execution effect corresponding to the index, and use the index gap as the comparison result; Modify the initial auxiliary decision according to the preset adjustment rule based on the index gap to obtain the final auxiliary decision.

10. A work ticket auxiliary decision-making system, characterized in that, The system is applied to a distribution network, and the distribution network includes a plurality of distribution facilities. Each distribution facility corresponds to a different power operation system. The system includes: A data processing unit, configured to collect and fuse multi-source data of the power operation system of the distribution facility when an operation needs to be performed on the distribution facility, to obtain multi-modal data of the power operation system; A feature extraction unit, configured to obtain deep semantic features of the multi-modal data according to the multi-modal data through a large-scale language model, a data feature extraction model, and an image feature extraction model; A matching unit, configured to perform matching according to the deep semantic features through a preset knowledge graph to obtain knowledge graph nodes and relationships related to the multi-modal data; An initial decision unit, configured to determine an initial auxiliary decision of a work ticket for power operation according to the knowledge graph nodes and relationships, in combination with a preset deep reinforcement learning network; A simulation unit, configured to input the initial auxiliary decision into a power operation system model for simulation to obtain an execution effect corresponding to the initial auxiliary decision; An optimization unit, configured to optimize the initial auxiliary decision according to the execution effect and the expected execution effect to obtain the final auxiliary decision of the work ticket.

Images