A method for performing operation and maintenance of power grid assets based on a reasoning large model

Abstract

The application belongs to the field of electric power, and provides a method for performing operation and maintenance of power grid assets based on a reasoning large model, collects multi-source operation data of power grid assets; constructs a power grid topology structure to form a power grid operation state diagram; performs targeted training and optimization on a pre-trained reasoning model to ensure that an operation state prediction output by the reasoning model meets an electrical feasible region; causes the reasoning model to automatically avoid a maintenance scheme that is inconsistent with a load curve, backup capacity and dispatch approval in a reasoning process; optimizes the reasoning model through a chain reasoning loss function to enable the reasoning model to predict a cascading effect of equipment failure; and generates a maintenance action sequence that meets a feasible region of a power grid maintenance period and can minimize risk conduction based on an output result of the reasoning model. The method improves the efficiency of operation and maintenance of power grid assets.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of electric power, and in particular relates to a method for operation and maintenance of power grid assets based on a reasoning large model. BACKGROUND

[0002] With the continuous expansion of the power grid scale and the increasing penetration of new energy, the operating state is jointly influenced by the coupling of device health, external environment and network topology. Operation and maintenance not only needs to identify single device hidden dangers, but also needs to consider physical constraints such as power flow balance, power conservation and short circuit safety, and is influenced by time period constraints such as load curve, standby capacity and dispatching approval window. With the rapid growth of multi-source data such as online monitoring, weather and historical work orders, how to uniformly express heterogeneous data and power grid topology, and how to make the reasoning results directly converted into executable maintenance sequences become key requirements.

[0003] In the prior art, common means include periodic maintenance and state-based maintenance, relying on SCADA and online monitoring for threshold alarm and trend analysis, using offline calculation of power flow and short circuit to check the safety boundary before and after maintenance, using experience rules or expert systems to generate power-off and power-on operation tickets and operation sequences, and implementing maintenance through manual coordination or simple scheduling to select low-load windows. Some data-driven methods can perform life estimation and fault prediction, and some methods can structure work order texts to assist decision-making. Some systems introduce hierarchical approval processes and resource constraint management for parallel jobs.

[0004] The prior art still has deficiencies, mainly manifested in that multi-source data and power grid topology are difficult to be uniformly modeled, single device prediction ignores chain transmission leading to underestimation of systemic risk, data-driven results lack physical constraints and appear unexecutable operating states, maintenance plans fail to form a closed-loop linkage with time period feasible region, standby capacity and dispatching approval, and generated operation sequences cannot meet safety boundaries while suppressing risk propagation along the path, making it difficult to be stably landed in engineering practice. SUMMARY

[0005] In order to solve the problems in the prior art, the present application provides a method for operation and maintenance of power grid assets based on a reasoning large model, comprising the following steps:

[0006] Collecting multi-source operating data of power grid assets, the data including device state monitoring data, operating environment parameters and historical maintenance records;

[0007] Constructing a power grid topology structure, mapping the multi-source operating data to the attribute space of power grid topology nodes and edges to form a power grid operating state graph;

[0008] The pre-trained inference model is trained and optimized, and in the training process, power physical constraints are introduced, and power flow balance, power conservation and short-circuit current electrical equations are taken as constraint conditions to ensure that the operation state prediction output by the inference model meets the electrical feasible region; the grid maintenance period constraint is converted into a feasible region embedding vector and input to the inference model as a condition, so that the inference model automatically avoids the maintenance scheme that does not match the load curve, standby capacity and dispatch approval in the inference process; a risk transmission chain model is established to propagate the impact of potential faults along the grid topology path, and the inference model is optimized through a chain inference loss function, so that it can predict the cascading effect of equipment failure;

[0009] Based on the output result of the inference model, a maintenance action sequence that meets the grid maintenance period feasible region and can minimize risk transmission is generated to guide the operation and maintenance of the grid assets.

[0010] Further, the historical maintenance work order and maintenance record are digitally processed, and the converted data is indexed with the device number as the index key, and the device operation monitoring data and environmental data are established in a unified index relationship and stored in the central maintenance database.

[0011] Further, the grid topology structure refers to a graph structure formed by taking grid equipment as nodes and electrical connection relationship between devices as edges;

[0012] The nodes include transformers, busbars, circuit breakers, disconnectors, cables and transmission line endpoints, and the edges include line, busbar connection and switch access relationship;

[0013] The attribute space refers to a set of numerical or categorical features attached to the nodes and edges, used to store operation data related to the nodes or edges.

[0014] Further, the device monitoring data is stored in a real-time database, and the device number is used as a unique index identifier. By comparing the node number in the topology file with the device number in the database, data binding is achieved.

[0015] Further, a node list and an edge list are generated according to the topology file, and an index relationship is established; the data in the monitoring database and the environmental database are filled into the attribute set of the corresponding nodes and edges according to the index number; the filled node attributes and edge attributes are stored as node attribute matrix and edge attribute matrix respectively.

[0016] Further, the power flow balance refers to the balance relationship that should be met between the power injected at each node and the branch power in the power system, to ensure that the system operation state and power distribution are consistent;

[0017] The power conservation means that the sum of the injected power and the outflowing power at any node is balanced, and no power is added or disappeared without reason;

[0018] The short-circuit current means the maximum branch current of the power system when a fault occurs, which is determined by the grid topology and impedance, and is an important parameter for testing the safety of equipment operation;

[0019] The electrical feasible region means a state set jointly limited by electrical constraint equations, and only the operating state falling into the set is considered as an effective state conforming to the electrical law.

[0020] Further, the grid maintenance period constraint means a maintenance allowed window determined by system operating conditions and dispatching rules, including load curve constraints, reserve capacity constraints, and dispatching approval constraints;

[0021] The load curve constraint means the condition that the grid limits the shutdown of part of the equipment during the high load period, the reserve capacity constraint means that sufficient reserve power capacity must be maintained in the system during equipment maintenance, and the dispatching approval constraint means that the maintenance scheme needs to be approved by the dispatching department before implementation under the premise of conforming to the operating safety boundary;

[0022] The feasible region embedding vector means the condition channel that converts the above constraint conditions into numerical form through a feature method and inputs the reasoning model in the form of a vector.

[0023] Further, the conduction coefficient of each edge is calculated, a monotone decreasing mapping of impedance modulus value, a monotone increasing mapping of power relative to rated value, and a monotone increasing mapping of environmental stress factor are weighted combined, and the result is limited in the interval of zero to one through compression mapping to obtain a weighted topology matrix; then a finite step propagation calculation is performed to accumulate the influence from upstream nodes layer by layer to obtain the risk intensity distribution of the nodes in the whole network, and finally a chain reasoning loss function is constructed with reference to the theoretical conduction result.

[0024] Further, the chain reasoning loss function is:

[0025] Suppose that the power grid contains N nodes, W is an edge weight matrix of size N×N, and the element represents the conduction coefficient from node i to node j, with a value range of 0 to 1, s is a seed vector of size N, and the element represents the initial influence value at node i, 0<θ<1 is a propagation attenuation factor, K is an upper bound of the number of propagation steps, and T is a reference conduction matrix calculated from electrical equations and historical samples;

[0026] The propagation prediction vector of the model is defined as:

[0027] ,

[0028] Definition of theoretical propagation vector:

[0029] ,

[0030] The chain reasoning loss function is:

[0031] ,

[0032] wherein, and are non-negative weight coefficients for balancing node-level consistency and multi-step path consistency, represents the k-th power of the matrix W, which measures the cumulative effect of path propagation with length k, represents the two-norm, represents the k-th power of the theoretical conduction matrix, reflecting the propagation result of the path with length k under the theoretical model.

[0033] Further, the maintenance action sequence refers to a set composed of a series of ordered maintenance operation steps, which include one or more combinations of load transfer, switch operation, device de-commissioning, on-site maintenance, test verification and device restoration.

[0034] The present application can ensure that the output result meets the engineering implementability and effectively controls the cascading risk in the maintenance scheme generation link by combining the power grid operation data, physical constraint conditions and reasoning large model. In this process, the inference ability of the reasoning model is no longer limited to the single-point state of the device, but realizes the system-level optimization through topological propagation and time period feasible region constraints. The beneficial effects of the method are reflected in the following aspects:

[0035] Firstly, the present application can significantly improve the scientificity and rationality of the maintenance scheme. Since the constraint conditions of electrical physical equations and the time period feasible region of the power grid are introduced in the reasoning process, the maintenance action sequence output by the model always remains consistent with the power flow balance, power conservation and dispatching approval window, thereby avoiding the unimplementable scheme caused by ignoring physical laws or external limitations in traditional methods.

[0036] Secondly, the present application can effectively reduce the systemic risk. By establishing a risk conduction chain model and propagating the potential fault impact along the power grid topology path, the reasoning model can not only predict the operation risk of the device itself, but also identify the cascading effect, preferentially cut off the high-risk conduction path, and finally generate a maintenance sequence that can minimize the risk diffusion. This mechanism makes the maintenance process not only solve the device hidden danger, but also ensure the stability of the overall power grid.

[0037] Overall, the implementation of the present application can realize the transformation of power grid asset operation and maintenance from experience-driven to intelligent decision-making, not only improving the efficiency and accuracy of scheme generation, but also ensuring the safety and economy of maintenance execution, providing strong support for building a highly reliable modern power grid. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0039] Figure 1 is the main flowchart of the present application;

[0040] Figure 2 is a schematic diagram of power grid operation state;

[0041] Figure 3 is a reasoning model avoidance schematic diagram;

[0042] Figure 4 is a risk transmission schematic diagram. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and beneficial effects of the present application more clear and obvious, the present application will be further described in detail in combination with embodiments. It should be understood that the following embodiments are only used to illustrate the present application, and are not used to limit the protection scope of the present application.

[0044] As shown in Figure 1 , the present embodiment provides a method for operation and maintenance of power grid assets based on reasoning large model, which can combine the operation characteristics of power grid assets and the electrical physical constraint conditions to train and optimize the pre-trained reasoning model, so as to generate intelligent maintenance scheme conforming to the operation law and maintenance conditions of power grid. The method as a whole includes the following stages:

[0045] In the process of power grid asset operation and maintenance, the device state is influenced by multiple factors, including real-time monitored electrical quantity parameters, dynamic conditions of external operating environment, and historical maintenance and fault conditions. If only relying on data from a single source, it is often difficult to accurately reflect the overall operating state of the device, thereby leading to a lack of understanding of global information in the subsequent training and application process of the reasoning model. In order to improve the understanding ability of the reasoning large model to the operation law of the power grid asset, it is necessary to establish a comprehensive data input basis to ensure that the model can simultaneously perceive the internal state of the device, the external operating conditions and the historical evolution trajectory. Therefore, multi-source operation data of the power grid asset is collected, including device state monitoring data, operating environment parameters and historical maintenance records.

[0046] The device state monitoring data refers to real-time parameters obtained by sensors, protection devices and monitoring systems installed on power grid devices. Typical parameters include voltage, current, power, temperature, oil pressure, partial discharge intensity, etc.

[0047] The operating environment parameters refer to external conditions in the external environment of the power grid device that may affect its operation. Typical parameters include air temperature, humidity, rainfall, wind speed, air salt content and geographic altitude.

[0048] The historical maintenance records refer to work order data, fault logs and maintenance reports formed by the power grid operation and maintenance department in the past maintenance and fault handling process. Such records can reflect the deterioration trend, hidden danger location and common fault mode of the device.

[0049] In the specific implementation process, first, a multi-type sensor network is deployed at the power grid device end. The sensors include voltage sensors, current transformers, temperature sensors, oil pressure sensors, partial discharge sensors and vibration sensors, etc. The sensors are connected to the remote terminal unit through the local acquisition module. The remote terminal unit is used for preprocessing the original signals collected by the sensors, including noise filtering, preliminary outlier removal and data compression, and transmitting the processed monitoring data to the central data platform in real time through communication links such as optical fiber private line, microwave communication or wireless public network private network. In this way, the device operating parameters can be preliminarily cleaned and standardized at the acquisition end, reducing the subsequent processing burden.

[0050] Then, an external meteorological database and an environmental monitoring system are called. The meteorological database is a national or regional meteorological information platform accessed by the power grid dispatching department, which can provide meteorological data such as temperature, humidity, wind speed, rainfall, and ice and snow load. The environmental monitoring system is an environmental sensing device deployed in the power grid operation area, which can collect external parameters related to electrical equipment operation such as air salt spray concentration, pollutant concentration, altitude, and soil humidity. The meteorological data and environmental monitoring data are written into the central data platform according to a fixed sampling period, and a spatial mapping relationship is established with the power grid equipment in the corresponding area.

[0051] Further, historical maintenance work orders and repair records are digitized. The work orders and records are usually stored in paper archives or unstructured electronic document form. To ensure consistency with real-time data in subsequent processing, first, the optical character recognition method or text structured annotation method is used to convert such materials into a standardized data format. The converted data is indexed by device number, and a unified index relationship is established with the operation monitoring data and environmental data of the device, and stored in the central maintenance database. This database can support multi-dimensional queries, including quick retrieval by device, time, fault category, or maintenance measures.

[0052] After the above multi-source data is collected into the central data platform, it is standardized by a unified data access module. The data access module first converts the data format from different sources, converting data with multiple sampling frequencies and unit systems into a unified standard system. For example, meteorological data collected at minute intervals and current monitoring data collected at second-level are normalized in time scale. Then, based on a unified timestamp mechanism, all types of data are aligned to eliminate time deviations caused by inconsistent sampling intervals. Finally, a mapping table is established using a unified naming specification to address differences in data naming conventions between different monitoring systems and maintenance systems, ensuring that data from different sources can be consistently stored and called.

[0053] Specifically, in the operation scenario of a certain 220kV main transformer, the real-time collected oil temperature data is 85℃, the winding current data is 620A, the external environment temperature is 35℃, and the humidity is 78%. At the same time, the historical maintenance record shows that the main transformer has experienced two sleeve discharge faults in the past three years. Through the above data collection and unified processing, the operation status of the main transformer can be completely described in the power grid operation status diagram, providing effective support for the subsequent training and optimization of the reasoning model.

[0054] In the process of power grid operation and maintenance, power grid assets are not operated in isolation, and the operation state of a single device is often affected by its position in the topology, the connection mode with adjacent devices, and the overall power flow distribution of the system. If only single-device data is used as input, it is easy to ignore the systemic risks and cascading effects. Therefore, it is necessary to build a topology structure that can reflect the relationship between devices and overall operation constraints, map multi-source data of different sources to the topology nodes and edges, and form a power grid operation state graph that can fully depict the operation characteristics of the power grid. To this end, a power grid topology structure is constructed, and the multi-source operation data is mapped to the attribute space of the nodes and edges of the power grid topology to form a power grid operation state graph.

[0055] In the present application, the power grid topology structure refers to a graph structure formed by taking power grid devices as nodes and the electrical connection relationship between devices as edges. This topology structure can reflect the coupling relationship between devices in the power grid.

[0056] The nodes include transformers, buses, circuit breakers, disconnectors, cables, and transmission line endpoints, and the edges include line, bus tie, and switch access relationships.

[0057] The attribute space refers to a set of numerical or categorical features attached to the nodes and edges, used to store operation data related to the nodes or edges.

[0058] The power grid operation state graph refers to a graph data model with state information formed by combining the topology structure and the attribute space, used to describe the overall operation of the power grid at a specific time.

[0059] In the specific implementation process, first, the topology connection relationship of the power grid devices is obtained, which can reflect the physical access mode and electrical connection logic of various devices in the power grid. This relationship is usually derived from the main wiring diagram or power network model file maintained by the power grid dispatching control system. The main wiring diagram represents the connection mode between transformers, buses, switches, and lines in a graphical manner, and the network model file stores the connection attributes and operation parameters between devices in a structured data manner. By analyzing the above files, the set of device nodes and the connection relationship between nodes can be obtained, thereby forming the initial topology framework of the power grid.

[0060] Then, the device node number is corresponded with the device number of the multi-source operation data, ensuring that each node in the topology can establish a mapping relationship with its monitoring data. Specifically, the device monitoring data is usually stored in a real-time database, and the device number is used as a unique index identifier. By comparing the node number in the topology file with the device number in the database, data binding can be achieved. For example, when the topology file contains a main transformer node numbered T1, the device monitoring data numbered T1, such as oil temperature, current, and partial discharge signal, can be found in the operation database, and the data is attached to the attribute set of node T1.

[0061] Further, the attributes of the line side are set as transmission power, current, impedance and the like. The transmission power and current are collected in real time by the line monitoring device or the online monitoring system, and the impedance value is provided by the design archives or historical test data. When establishing the attributes of the side, external condition attributes are also attached to the side in combination with meteorological and environmental data to reflect the influence of the external environment on the line operation capacity. For example, when the line is in a high-temperature, strong-wind or high-humidity environment, its current-carrying capacity will change, and therefore a correction factor needs to be introduced in the attribute set of the side to adjust the rated operation capacity of the line.

[0062] The mapping process is realized by a topology modeling module, which is a functional component for processing power grid multi-source heterogeneous data. The module can convert data from different sources into a unified data structure. The implementation of the module is as follows: first, generate a node list and an edge list according to the topology file, and establish an index relationship; second, fill the data in the monitoring database and the environmental database into the attribute sets of the corresponding nodes and edges according to the index number; finally, store the filled node attributes and edge attributes as a node attribute matrix and an edge attribute matrix respectively, wherein the node attribute matrix is used to describe the operation state of each device, and the edge attribute matrix is used to describe the transmission characteristics and external influences of the lines and switches. The node attribute matrix and the edge attribute matrix are stored together with the topology connection relationship as a graph data structure for subsequent reasoning model calling and processing.

[0063] In the above manner, the organic combination of the power grid topology structure and the multi-source operation data can be realized, so that the operation data not only exists in the form of isolated numerical values, but also is embedded in the topology structure as the attributes of nodes or edges, thereby forming a power grid operation state graph that can reflect the overall operation characteristics of the power grid.

[0064] Specifically, as shown in Figure 2 in a regional power grid, there are the following devices:

[0065] The main transformer T1 has a rated voltage of 220 kV. The real-time monitored parameters are an oil temperature of 85 °C and a winding current of 620 A. The historical maintenance records show that there have been two sleeve discharge faults in the past three years.

[0066] The bus B1 has a real-time monitored voltage of 219 kV and a current load power of 150 MW.

[0067] The transmission line L1 connects the main transformer T1 and the bus B1. The monitored current is 550 A, and the impedance is 0.23 Ω. The environmental monitoring system records an air temperature of 35 °C and a wind speed of 2.8 m / s, which will correct the current-carrying capacity of the line.

[0068] The transmission line L2 extends from the bus B1 to the external substation node B2. The monitored current is 430 A, and the impedance is 0.31 Ω. The external environmental humidity is 78%, which will affect the insulation safety of the line.

[0069] The external substation B2 has a voltage of 220 kV and a load power of 180 MW. There are no major fault events in the historical operation records.

[0070] In constructing the power grid operation state graph, the main transformer T1, the bus B1, and the external substation B2 are taken as nodes, and the transmission lines L1 and L2 are taken as edges. The nodes and edges are attached with corresponding attribute information to form a graph structure with attributes.

[0071] The pre-trained inference model refers to a deep learning model that has completed basic training on a large-scale dataset and has general pattern recognition and logical reasoning capabilities. This type of model learns universal laws through multi-modal data or large-scale time series data in the pre-training phase, thereby showing strong generalization ability in different tasks. The optional implementations of the pre-trained inference model include but are not limited to the general large model DeepSeek model, the GPT model, the LLaMA model, as well as the graph neural network model for graph structure analysis and the Transformer model for time series prediction.

[0072] Since the above pre-trained inference model has mastered basic capabilities such as time series patterns, spatial topological relationships, and semantic understanding in the training process, and the power grid asset operation and maintenance data naturally have time series, topological, and semantic properties, the power grid specific operation data can be mapped to the input space of this type of model, and the power grid exclusive physical constraints and operation rules can be introduced through retraining or fine-tuning. This can quickly realize the migration and adaptation to the power grid operation and maintenance scene based on the existing reasoning ability, avoiding the high data and computing costs required for training the model from scratch.

[0073] If the general pre-trained model is used directly without targeted training and optimization, the output result may not conform to the electrical physical law, may not meet the power grid maintenance period constraint, and may lack the judgment of the risk conduction effect between devices, resulting in the generated maintenance scheme lacking implementability and professionalism. Therefore, targeted training and optimization are needed on the basis of the pre-trained model, so that the model has the understanding ability of the operation characteristics of the power grid assets while maintaining the general inference ability, and can generate a maintenance scheme that conforms to the electrical law, the dispatching requirements and the system risk constraints.

[0074] In power grid operation and maintenance, the device operating state not only depends on the monitoring data, but also is strictly constrained by the physical law of the power system. If the inference model only relies on data-driven in the training process without considering the electrical law, it may produce prediction results that violate the principles of power flow distribution, power conservation or short-circuit current calculation, thereby leading to the infeasibility of the maintenance recommendations in actual implementation, and even introducing new system risks. Therefore, it is necessary to introduce power physical constraints in model training, so that the inference model learns data features while always following the basic rules of power grid operation. To this end, in the training process, the power physical constraints are introduced, and the power flow balance, power conservation and short-circuit current electrical equations are used as constraint conditions to ensure that the operating state prediction output by the inference model meets the electrical feasible region.

[0075] The power flow balance refers to the balance relationship that should be met between the power injected at each node and the branch power in the power system, which is used to ensure that the system operating state and power distribution are consistent.

[0076] The power conservation refers to the fact that at any node, the sum of the injected power and the outgoing power remains balanced, and there is no case of power increasing or disappearing without reason.

[0077] The short-circuit current refers to the maximum value of the branch current when the power system fails, which is determined by the topology and impedance of the power grid, and is an important parameter for testing the safety of device operation.

[0078] The electrical feasible region refers to the state set defined by the above electrical constraint equations. Only the operating state falling into this set is considered to be an effective state that conforms to the electrical law.

[0079] In the implementation process, first, a loss function structure with power physical constraints is established for the inference model in the training data processing stage. The loss function not only includes a conventional prediction error term for measuring the deviation between the model output and the actual monitoring data, but also further introduces the residuals of power flow balance, power conservation and short-circuit current and other electrical equations as additional constraint factors to participate in optimization together. In this way, the inference model is driven by two factors during the learning process. On the one hand, it approximates the actual observation data, and on the other hand, it strictly follows the physical laws of the power system, thereby ensuring that the model obtained during training converges to a unified state between prediction ability and physical rationality.

[0080] For the power flow balance constraint, at each node of the power system, the active power and reactive power injected by the node should be equal to the algebraic sum of the power flowing in or out through adjacent branches. During training, first, the node voltage amplitude and voltage phase angle parameters predicted by the model are used to solve the power flow of each branch according to the power flow calculation formula, and then the calculated branch power is aggregated to the corresponding node to construct the node power balance equation. By comparing the predicted node power with the theoretical balance condition, the residual obtained is the power flow balance residual, which is introduced as a constraint term into the loss function. During the iterative optimization process, the model parameters are constantly adjusted under the action of the constraint term, so that the prediction result gradually approaches the power flow balance condition, avoiding states that do not conform to the power distribution rules of the power grid.

[0081] For the power conservation constraint, at any node of the power system, the injected power and the outgoing power must strictly maintain the conservation relationship. To achieve this, the injected power of the node predicted by the inference model and the outgoing power of the adjacent lines are differentiated, and the power conservation residual is obtained after squaring or weighting the difference. This residual is set as a penalty term in the loss function, prompting the model to avoid outputting results that do not satisfy the power conservation law during training. In this way, the energy exchange process of the node can be effectively guaranteed to comply with the electrical physical laws, thereby improving the feasibility of the overall prediction.

[0082] For the short-circuit current constraint, when a fault occurs in the power system, the size of the short-circuit current is determined by the network topology and branch impedance. In the implementation, first, a node impedance matrix is constructed based on the power grid topology, and the theoretical short-circuit current under the condition of short-circuit at a specific node is calculated according to the electrical theory method, which is taken as the reference result. Then, the predicted short-circuit current value obtained by the inference model is compared with the reference value, and the deviation between the two is calculated to form the short-circuit current constraint residual. This residual is written into the loss function as an additional constraint term and repeatedly participates in parameter updating during the training iteration process, forcing the model to correct the prediction result of the short-circuit current during the learning process, so that the prediction gradually converges to a range consistent with the electrical theory.

[0083] During the training iteration process, the model is repeatedly executed through forward propagation and backward update, while minimizing the prediction error term and the electrical equation constraint residual. Finally, in the state of convergence, the model can not only give the prediction results close to the monitoring data, but also ensure that these prediction results are located within the electrical feasible region. Through this training method, the inference model not only accurately captures the trend of the device running state, but also always maintains consistency with the electrical physical law, thereby providing reliable data support for power grid operation and maintenance decision-making.

[0084] In a certain regional power grid, there is a 220kV main transformer T1, a section of bus B1, two transmission lines L1 and L2, and an external substation B2. The specific situation is as follows:

[0085] Main transformer T1: the predicted winding current is 620A, the oil temperature is 85℃, and the historical maintenance record shows that there have been two sleeve discharge faults in the past three years.

[0086] Bus B1: the predicted voltage is 219kV, and the predicted power injection is 150MW.

[0087] Transmission line L1: connecting the main transformer T1 and the bus B1, the predicted transmission current is 550A, and the impedance is 0.23Ω. When the external air temperature is 35℃ and the wind speed is 2.8m / s, the current-carrying capacity is corrected to 530A.

[0088] Transmission line L2: connecting bus B1 and external substation B2, the predicted transmission current is 430A, and the impedance is 0.31Ω. When the external humidity is 78%, the insulation safety margin decreases by 5%.

[0089] External substation B2: the predicted voltage is 220kV, and the load power is 180MW.

[0090] In this operating state, the prediction results of the inference model will be subjected to the following constraint tests:

[0091] Power flow balance constraint: the power injection of bus B1 node 150MW should be equal to the sum of the power flow of L1 and L2 lines. If the difference is not zero, a power flow residual is formed, which is constrained by the loss function.

[0092] Power conservation constraint: the predicted injection power of T1 node must be equal to the power outflow of L1, and the injection power of B2 node must be equal to the power outflow of L2. If they are not equal, a conservation residual is formed.

[0093] Short-circuit current constraint: assuming that a single-phase ground fault occurs at B1 node, the theoretical short-circuit current is 9.8kA based on impedance matrix calculation, while the model prediction result is 10.3kA. The deviation between the two is recorded as a short-circuit residual, which participates in the loss function.

[0094] The model adjusts parameters to make the prediction results meet the electrical feasible region by minimizing the weighted sum of the prediction error, the power flow residual error, the power conservation residual error and the short-circuit current residual error in the process of training iteration.

[0095] In the process of power grid operation and maintenance, the maintenance behavior is not only affected by the equipment state, but also affected by the system load change, the spare capacity configuration and the maintenance period approved by the dispatching department and other external constraints. If the reasoning model is only based on the equipment operating state and historical data for reasoning without considering the above constraints, the maintenance scheme may be generated which does not conform to the actual operating conditions, for example, proposing the suggestion of shutting down the key equipment in the peak load period or arranging a large range of maintenance in the case of insufficient spare capacity. Although such scheme is feasible at the data level, it cannot be executed in actual operation, and even may cause system operation risk. In order to ensure that the output result of the reasoning model has engineering implementability, it is necessary to convert the power grid maintenance period constraint into a form that can be processed by the model and embedded in the reasoning process, so as to fundamentally avoid the scheme that does not conform to the conditions. Therefore, the power grid maintenance period constraint is converted into a feasible region embedding vector and input as a condition to the reasoning model, so that the reasoning model automatically avoids the maintenance scheme that does not conform to the load curve, the spare capacity and the dispatching approval in the reasoning process.

[0096] The power grid maintenance period constraint refers to the maintenance allowed window determined by the system operating conditions and the dispatching rules, including the load curve constraint, the spare capacity constraint and the dispatching approval constraint.

[0097] The load curve constraint refers to the condition that the power grid limits part of the equipment shutdown in the high load period, the spare capacity constraint refers to the condition that sufficient spare power capacity must be maintained in the system during equipment maintenance, and the dispatching approval constraint refers to the condition that the maintenance scheme needs to be approved by the dispatching department before implementation on the premise of conforming to the operating safety boundary.

[0098] The feasible region embedding vector refers to the condition channel that converts the above constraint conditions into numerical form through the feature method and inputs the reasoning model in the form of a vector.

[0099] In the specific implementation process, first, the load curve of the power grid operation is time series segmented. The load curve refers to the function of the total power of the system changing with time within a certain time range, which usually has obvious periodicity and regularity, such as peak in the daytime, valley at night and transition stage of gradually rising or falling load. In the present application, the load curve of a day is divided into several periods, such as peak period, valley period and transition period. For each period, the representative load level parameters are extracted, including the maximum load, the minimum load and the average load, and the parameter values are mapped into interval vectors after being numerized. The interval vector is used as the feature expression of the load curve to identify the load state of different time periods.

[0100] Secondly, the standby capacity is calculated in real time. The standby capacity refers to the additional power capacity that can be provided in the power grid operation after removing the normal load, including rotating standby capacity and cold standby capacity. In the implementation process, the current operating state and output upper limit of each power generation unit or standby device are first obtained, and then the standby capacity value of each time period is calculated. The calculated standby capacity is compared with the preset minimum standby demand value. If the standby capacity is greater than or equal to the minimum demand, it is identified as sufficient capacity, and if the standby capacity is less than the minimum demand, it is identified as insufficient capacity. The identification information is further converted into binary or numerical feature input to reflect the safety margin of the power grid in the time period.

[0101] Further, the dispatch approval constraint is encoded. The power grid maintenance must be approved by the dispatch department and is limited by the approval window and the approval priority. The approval window refers to the specific time range allowed by the dispatch department for maintenance, and the approval priority refers to the sorting principle of different maintenance tasks by the dispatch department when multiple maintenance applications exist simultaneously. In the implementation process, the approval plan provided by the dispatch system is first read, and the approval status is marked as a numerical label, such as approval status marked as 1 and unapproved status marked as 0. The time range corresponding to the approval window is mapped to the time period index. The approval priority is converted into a level code, such as priority 1 representing the highest priority and priority 3 representing a lower priority. After the above steps, the dispatch constraint feature label that can be directly input into the model is formed.

[0102] After uniform processing of the interval vector of the load curve, the sufficiency identification of the standby capacity, and the feature label of the dispatch approval constraint, they are combined into a multi-dimensional embedding vector. The multi-dimensional embedding vector can represent the load condition, standby margin, and dispatch approval situation in the vector space simultaneously. The embedding vector is continuously updated with real-time data of power grid operation in the time dimension to ensure that the model always reasons based on the latest operating environment.

[0103] When the inference model receives the embedding vector, the model automatically calls the constraint information in the internal reasoning path generation stage. In the candidate maintenance scheme generation process, any scheme that does not meet the feasible region conditions described by the embedding vector will be directly excluded. For example, when the load is in the peak period, the standby capacity is insufficient, and the dispatch is not approved, even if the device state shows a potential fault, the model will not output the maintenance action scheme for that period, but will automatically postpone the maintenance to a period that meets the load curve trough, sufficient standby capacity, and dispatch approval. Finally, the model generates a maintenance action sequence only within the feasible region conditions, thereby ensuring the engineering implementability and operational safety of the output results.

[0104] As Figure 3As shown, in a certain regional power grid, there is a 220 kV main transformer T1, a section of bus B1, and two transmission lines L1 and L2, which are connected to external substations B2 and regional load centers, respectively. The system operation data and constraint conditions are as follows:

[0105] Load curve data: The load curve provided by the dispatching system shows that the peak period is from 9:00 to 21:00 during the day, with a load of 180 MW to 200 MW; the valley period is from 23:00 to 6:00 at night, with a load of 110 MW to 130 MW; the remaining time is the transition period, with the load gradually rising or falling. According to the segmentation rule, the peak, valley, and transition periods are mapped to interval vectors, with the peak period marked as [0.8-1.0], the valley period marked as [0.4-0.6], and the transition period marked as [0.6-0.8].

[0106] Backup capacity data: During the peak period, the backup capacity is only 15 MW, while the minimum backup requirement specified by the dispatching is 20 MW, so it is marked as insufficient capacity. During the night low valley period, the backup capacity is 35 MW, which is greater than the minimum backup requirement, so it is marked as sufficient capacity. This capacity state is converted into a binary feature, with sufficient capacity marked as 1 and insufficient capacity marked as 0.

[0107] Dispatch approval data: The dispatching system approval plan shows that the approval maintenance window is from 23:00 to 5:00 on weekdays, with priority 1 for critical equipment emergency maintenance and priority 3 for regular maintenance tasks. This period is marked as an available maintenance window, with the approval state coded as 1, the time interval index as the night low valley interval, and the priority marked as 3.

[0108] By constructing a multi-dimensional embedding vector [load interval vector, backup capacity identifier, approval state and priority] from the above information, for example, at night 23:00, the embedding vector is [0.4-0.6, 1, (1, 3)]. When the inference model receives this embedding vector, it will automatically identify it as a low-load, sufficient-capacity, and approved period when generating a maintenance plan, allowing the output of a maintenance plan for the main transformer T1. If inference is performed at 10:00 during the day, the embedding vector is [0.8-1.0, 0, (0, -)], at which time the inference model will automatically exclude all maintenance plans, even if the T1 device status shows potential risks, the model will delay the maintenance recommendation to avoid unreasonable maintenance actions in high-load and insufficient-capacity periods.

[0109] In the process of power grid operation and maintenance, the potential failure of a single device is often not limited to itself, but can also be transmitted to adjacent devices and downstream systems through the power grid topology, thereby triggering a series of cascading effects. If the inference model only makes isolated predictions for a single device without considering the risk transmission on the topology path, the generated maintenance scheme may underestimate the overall risk, leading to cascading failures during operation, and even causing large-scale power outages. Therefore, it is necessary to establish a risk transmission chain model to propagate the impact of potential failures along the power grid topology path, and optimize the inference model through a chain inference loss function, so that it can predict the cascading effects of device failures.

[0110] The risk transmission chain model refers to a path propagation model constructed on the power grid topology structure, which can simulate the transmission relationship between potential failures among devices and lines.

[0111] The chain inference loss function refers to, in the training process of the inference model, in addition to considering the prediction error of the device itself, the residual error of the result after the transmission of the fault influence on the topology path is calculated, and the chain constraint is applied to the loss function, so as to force the model to consider both the local device state and the system-level risk transmission characteristics when optimizing parameters.

[0112] In the specific implementation process, first, the power grid topology structure is modeled to form a set of node and edge relationships, the node is used to represent the electrical equipment such as transformer, bus, line end, circuit breaker, and the edge is used to represent the physical connection relationship between devices. Secondly, the initial impact value of the potential failure is set as a seed signal and injected into the fault device node, and propagated to the downstream in turn according to the topology connection. In the propagation process, a transmission coefficient is assigned to each edge, which is determined by the line impedance, the real-time power of the line and the environmental stress factor. Through the iterative propagation process, the risk intensity value of each node downstream under the influence of the fault is obtained. Then, the risk transmission prediction result generated by the inference model is compared with the theoretical transmission result calculated according to the electrical equation and historical samples, the residual error is calculated and written into the chain inference loss function. The chain inference loss function acts simultaneously with the conventional defined prediction error in the training process, so that the model considers both the single-point prediction correctness and the consistency of the cascading effect when updating parameters, thereby realizing the accurate characterization of the system-level risk.

[0113] For the convenience of understanding, the related terms are explained as follows. The initial impact value of potential fault refers to a non-negative quantity used to start the propagation calculation at the fault device node, indicating the fault injection strength of the node. The conduction coefficient refers to a proportional factor for transmitting fault impact along the topology edge, with a value range of zero to one, and the greater the value, the easier the risk is transmitted along the edge. The risk intensity value refers to the quantitative value of the chain risk obtained by each node after the propagation ends. The theoretical conduction result refers to the benchmark propagation result obtained according to the power flow sensitivity, impedance parameter and historical outage sample, which is used as a reference for comparison in training. The chain reasoning loss function refers to a loss term for measuring the deviation between the predicted propagation result and the theoretical propagation result, and can be superimposed with path consistency constraints or stability constraints to improve the physical rationality of training convergence.

[0114] In the specific implementation process, first, the node set and edge set are parsed from the main wiring diagram or network model file derived from the dispatching system, a weighted adjacency matrix is constructed, and the consistency binding of device number and topology number is completed. Then in the fault diagnosis stage, the seed node and its initial impact value are determined according to the device state indicator, and the value is loaded into the propagation initial value vector. Further, the conduction coefficient of each edge is calculated, which is weighted combined by using a monotone decreasing mapping of impedance modulus value, a monotone increasing mapping of power relative rated value and a monotone increasing mapping of environmental stress factor, and the result is limited in the interval of zero to one through compression mapping to obtain a weighted topology matrix. Then, a limited step propagation calculation is performed, the influence from upstream nodes is accumulated layer by layer, and the risk intensity distribution of all nodes in the network is obtained. Finally, taking the theoretical conduction result as a reference, a chain reasoning loss function is constructed, and the prediction error and chain loss are minimized in each training iteration, so that the model simultaneously meets the consistency constraints of data fitting and topology chain.

[0115] To ensure the clarity of implementation, a specific chain reasoning loss function is given. Let the power grid contain N nodes, W be the edge weight matrix of size N x N, and the element represent the conduction coefficient from node i to node j, with a value range of 0 to 1, s be the seed vector of size N, and the element represent the initial impact value at node i, 0 < θ < 1 be the propagation attenuation factor, K be the upper bound of the number of propagation steps, and T be the benchmark conduction matrix calculated from the electrical equation and historical samples. The propagation prediction vector of the model is defined as

[0116]

[0117] The theoretical propagation vector is defined as

[0118]

[0119] The chain reasoning loss function is

[0120] ​​ ,

[0121] wherein, and are non-negative weight coefficients for balancing node-level consistency and multi-step path consistency. The meanings of the above symbols are as follows, denotes the k-th power of the matrix W, measuring the cumulative effect of path propagation with length k, denotes the two-norm, denotes the k-th power of the theoretical propagation matrix, reflecting the propagation result of the path with length k under the theoretical model.

[0122] Through the above scheme, the trained inference model can accurately reflect the cascading propagation relationship of fault influence on the topology while maintaining single-point prediction accuracy, reduce the probability of underestimating systemic risk, and improve the global rationality when generating the maintenance action sequence.

[0123] As shown in Figure 4 , in a certain regional power grid, a simplified topology including a main transformer T1, a bus B1, and an external substation B2 is constructed, and a transmission line L1 connects T1 and B1, and a transmission line L2 connects B1 and B2. The initial impact value of the potential fault is injected into T1, and the seed risk is set to 1.0. The propagation coefficient is calculated according to the impedance, load, and environmental stress, and the propagation coefficient of L1 is set to 0.72, and the propagation coefficient of L2 is set to 0.64. The propagation attenuation factor is set to 0.8, and the propagation step number is set to 2. The amount of the first step propagation to B1 is 0.8x0.72, and the predicted risk of B1 is 0.576. The amount of the second step propagation along the path T1→B1→B2 is 0.8x0.8x0.72x0.64, and the predicted risk of B2 is 0.295. Taking the theoretical contrast estimated by the power flow sensitivity and the historical sample as the benchmark, the theoretical risk of B1 is set to 0.560, and the theoretical risk of B2 is set to 0.270, and the node-level residual error is obtained, the residual error of B1 is 0.016, and the residual error of B2 is 0.025. During training, the above residual error and the path component residual error are jointly counted into the chain inference loss, and the prediction error defined in the conventional manner is minimized synchronously, so that the model approximates the electrically reasonable cascading propagation result while meeting the physical constraints.

[0124] In the process of power grid operation and maintenance, although the inference model can predict potential equipment failures and risk propagation at the topology level, it cannot implement maintenance measures in engineering practice if the results are not converted into executable operation sequences. Maintenance activities must meet the time window approved by the dispatch, maintain sufficient backup capacity, and avoid transmitting unstable factors on paths where risks are easily amplified. If the model only outputs risk scores without providing specific action sequences, operation personnel will have difficulty operating accordingly, and even unreasonable maintenance tasks may be implemented during peak load or critical operation stages, thereby causing more serious cascading risks. Therefore, based on the output results of the inference model, a maintenance action sequence that meets the feasible region of the power grid maintenance period and can minimize risk propagation is generated to guide the operation and maintenance of power grid assets.

[0125] The maintenance action sequence refers to a set composed of a series of ordered maintenance operation steps, including load transfer, switch operation, equipment isolation, on-site maintenance, test verification, and equipment restoration.

[0126] The power grid maintenance period feasible region refers to the allowed time range formed after considering the constraints of load curve, backup capacity, and dispatch approval, ensuring that operations performed within this range will not negatively affect system operation stability.

[0127] The risk propagation minimization refers to prioritizing maintenance operations that can effectively cut off high-risk propagation paths when generating the action sequence to reduce the likelihood of cascading effects.

[0128] In the specific implementation process, first, the inference model outputs a list of potential faulty equipment, risk propagation paths, and risk intensity evaluation values, while extracting the feasible region information of the current operation period. Second, according to the time window of the feasible region, candidate maintenance time periods are divided, and time periods that do not meet the dispatch approval or have insufficient backup capacity are directly excluded. Third, in the remaining candidate time periods, the priority of potential faults is sorted, and equipment with a greater impact on system-level risks is processed first. Then, based on the topology path analysis results, the maintenance action sequence that can quickly weaken risk propagation is arranged, such as load transfer first, then isolation of the faulty node, and finally maintenance in the isolated environment. Finally, these actions are arranged in chronological order to form an action sequence that meets the constraint conditions.

[0129] In a preferred implementation, the maintenance action sequence introduces priority weights during generation, prioritizing equipment that has the greatest impact on core nodes or important loads into the early part of the sequence. This has the beneficial effect of quickly reducing the overall system risk level.

[0130] In another preferred implementation, the generation of the sequence of maintenance actions incorporates resource allocation information, including personnel, tools and concurrent work capacity, to avoid generating actions that cannot be executed due to insufficient resources or unfeasible field conditions, with the beneficial effect of ensuring that the maintenance plan is not only safe and feasible, but also practically operable.

[0131] In another preferred implementation, the sequence of maintenance actions employs a rolling generation mechanism, forming an initial sequence before plan execution and dynamically adjusting it during execution in combination with real-time operation data, with the beneficial effect of enhancing the adaptability and robustness of the maintenance plan, enabling it to cope with load fluctuations or unexpected events.

[0132] The sequence of maintenance actions generated by the above method not only satisfies the constraints of dispatch approval and spare capacity, but also effectively cuts off potential risk transmission paths, enabling the system operation risk to be controlled during maintenance, ultimately forming a maintenance plan that is both practically implementable and enhances overall operational safety.

[0133] In a specific example, the scenario of main transformer T1, bus B1, external substation B2 and lines L1, L2, the inference model predicts that T1 has a potential insulation degradation fault, and the risk is transmitted to B2 along L1 and L2, causing the downstream load risk to rise. In the dispatch-approved night low-load window, the generated sequence of maintenance actions includes the following steps: first, transfer part of the load to the standby power supply path before the start of the window; second, split T1 and perform safety grounding measures; third, field personnel perform casing and oil sample detection on T1; fourth, implement insulation defect treatment and sealing reinforcement; fifth, remove the grounding and gradually restore the switches after completing the maintenance; sixth, restore the original topology and monitor the stability of the power flow. This sequence successfully cuts off the high-risk transmission path under the premise of ensuring sufficient spare capacity and dispatch approval conditions, significantly reducing the system risk metric, demonstrating the implementability and beneficial effects of the method in engineering practice.

[0134] The part of the module structure of the present application that is not specifically clear shall be subject to the content described in the prior art. The prior art mentioned in the foregoing background section and the specific embodiment section of the present application can be used as a part of the present application to understand the meaning of some technical features or parameters.

Claims

1. A method for operation and maintenance of power grid assets based on a large-scale inference model, characterized in that, The method includes the following steps: Collect multi-source operational data of power grid assets, including equipment status monitoring data, operating environment parameters, and historical maintenance records; Construct a power grid topology and map the multi-source operational data to the attribute space of the power grid topology nodes and edges to form a power grid operation state diagram; The pre-trained inference model is trained and optimized in a targeted manner. During the training process, electrical physical constraints are introduced, with power flow balance, power conservation, and short-circuit current electrical equations used as constraints to ensure that the operating state prediction output by the inference model meets the electrical feasible region. The grid maintenance period constraints are transformed into feasible region embedding vectors and used as conditions input to the inference model, enabling the inference model to automatically avoid maintenance plans that do not conform to the load curve, reserve capacity, and dispatch approval during the inference process. A risk transmission chain model is established to propagate the impact of potential faults along the grid topology path, and the inference model is optimized through a chain-like inference loss function so that it can predict the cascading effects of equipment failures. Based on the output of the inference model, a maintenance action sequence that conforms to the feasible domain of the power grid maintenance period and can minimize risk transmission is generated to guide the operation and maintenance of power grid assets. The conduction coefficient of each edge is calculated by weighting a combination of a monotonically decreasing mapping for impedance magnitude, a monotonically increasing mapping for power relative to rated value, and a monotonically increasing mapping for environmental stress factor. The results are then constrained to the interval between zero and one by a compression mapping to obtain a weighted topology matrix. Subsequently, a finite-step propagation calculation is performed to accumulate the influence from upstream nodes layer by layer to obtain the risk intensity distribution of all nodes in the network. Finally, a chain-like inference loss function is constructed with the theoretical conduction results as a reference. The chain inference loss function is: Suppose the power grid contains N nodes, and W is an N×N edge weight matrix with elements... This represents the propagation coefficient from node i to node j, with values ​​ranging from 0 to 1. s is a seed vector of size N, with elements... Let θ represent the initial influence value at node i, 0 < θ < 1 be the propagation attenuation factor, K be the upper bound of the propagation steps, and T be the reference propagation matrix calculated from the electrical equations and historical samples; Define the propagation prediction vector of the model: Define the theoretical propagation vector: ， The loss function for chained inference is: in, and These are non-negative weighting coefficients used to balance node-level consistency with multi-step path consistency. This represents the k-th power of matrix W, measuring the cumulative effect of path propagation over a path of length k. Represents the L2 norm, This represents the k-th power of the theoretical propagation matrix, reflecting the propagation result of a path of length k under the theoretical model.

2. The method for operation and maintenance of power grid assets based on a large inference model according to claim 1, characterized in that, Historical maintenance work orders and maintenance records are digitized. The converted data is indexed by the equipment number and a unified index relationship is established with the equipment operation monitoring data and environmental data. The data is then stored in the central maintenance database.

3. The method for operation and maintenance of power grid assets based on a large inference model according to claim 1, characterized in that, The power grid topology refers to a graph structure formed with power grid equipment as nodes and the electrical connections between equipment as edges; The nodes include transformers, busbars, circuit breakers, disconnect switches, cables, and transmission line endpoints; the edges include line, busbar connections, and switch connection relationships. The attribute space refers to the set of numerical or categorical features attached to nodes and edges, used to store operational data related to nodes or edges.

4. The method for operation and maintenance of power grid assets based on a large inference model according to claim 1, characterized in that, Equipment monitoring data is stored in a real-time database, using the device number as a unique index identifier. Data binding is achieved by comparing the node number in the topology file with the device number in the database.

5. The method for operation and maintenance of power grid assets based on a large inference model according to claim 4, characterized in that, Generate a node list and an edge list based on the topology file, and establish an index relationship; populate the corresponding node and edge attribute sets with data from the monitoring database and the environment database according to the index number; store the populated node attributes and edge attributes as node attribute matrices and edge attribute matrices, respectively.

6. The method for operation and maintenance of power grid assets based on a large inference model according to claim 1, characterized in that, The power flow balance refers to the balance relationship that should be satisfied between the injected power of each node and the power of the branch in the power system, so as to ensure that the system operating state is consistent with the power distribution. The power conservation principle means that at any node, the sum of the injected power and the outflowing power should remain in balance, and there should be no situation where the power increases or disappears without cause. The short-circuit current refers to the maximum value of the branch current when a fault occurs in the power system. Its magnitude is determined by the power grid topology and impedance, and it is an important parameter for verifying the safe operation of equipment. The electrically feasible region refers to the set of states jointly defined by electrical constraint equations. Only operating states that fall within this set are considered valid states that conform to electrical laws.

7. The method for operation and maintenance of power grid assets based on a large inference model according to claim 1, characterized in that, The power grid maintenance period constraint refers to the maintenance allowance window determined by the system operating conditions and dispatch rules, including load curve constraints, reserve capacity constraints, and dispatch approval constraints. Load curve constraints refer to the conditions under which the power grid restricts the shutdown of some equipment during periods of high load. Reserve capacity constraints refer to the requirement that sufficient reserve power capacity must be maintained in the system during equipment maintenance. Dispatch approval constraints refer to the requirement that maintenance plans must be approved by the dispatching department before they can be implemented, provided that they meet the operational safety boundaries. The feasible region embedding vector refers to the condition channel that transforms the above constraints into numerical form through featureization methods and inputs them into the inference model in a vector manner.

8. The method for operation and maintenance of power grid assets based on a large inference model according to claim 1, characterized in that, The maintenance action sequence refers to a set of orderly maintenance operation steps, including one or more combinations of load transfer, switching operation, equipment disconnection, on-site maintenance, testing and verification, and equipment reset.

Images