A self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation

Abstract

This application provides a distribution network fault self-healing system based on large-scale model knowledge reasoning and multi-agent collaborative operation, relating to the field of distribution network fault self-healing technology. It collects multi-source heterogeneous data from the distribution network in real time and generates composite feature vectors. Alarm prompts guide a pre-trained power large-scale model to perform logical reasoning on the composite feature vectors, outputting a natural language reasoning chain. A fault knowledge graph is constructed, and based on this graph, fault root cause identification and causal reasoning are performed to obtain the root cause determination conclusion and causal reasoning evidence chain for the distribution network fault. The natural language reasoning chain is decoded into logical constraints for the action spaces of the isolation layer agents and the recovery layer agents, and security constraints for the action strategies of the isolation layer agents and the recovery layer agents are determined. Finally, fault self-healing control commands for the distribution network are generated and issued to the distribution equipment. This application can improve the continuous diagnosis and collaborative recovery capabilities of distribution network fault self-healing under complex operating conditions.

Description

Technical Field

[0001] This application relates to the field of distribution network fault self-healing technology, and more specifically, to a distribution network fault self-healing system based on large model knowledge reasoning and multi-agent collaborative system. Background Technology

[0002] Distribution network fault self-healing technology is an important technology in the fields of smart grids and power system automation. It is mainly used to complete the handling process such as fault detection, section isolation, power transfer restoration and operation verification after abnormalities occur in distribution lines, switching equipment, distributed power sources and load nodes. With the advancement of the construction of new power systems, the distribution network, as a key link to ensure power supply to end users, has more complex operating conditions. Fault self-healing technology plays an important role in shortening power outage time, reducing the handling pressure on dispatchers, and improving power supply restoration efficiency.

[0003] In existing technologies, distribution network fault handling still relies heavily on human experience, fixed criteria, and pre-set contingency plans. When the distribution network topology changes frequently, the proportion of distributed power sources increases, and environmentally induced faults increase, single electrical quantity judgment methods are insufficient to accurately reconstruct the fault occurrence process, easily leading to problems such as fragmented alarm information, insufficient fault identification, and interruption of the handling chain. At the same time, existing distribution automation systems do not fully integrate and utilize text logs, equipment ledgers, real-time waveforms, protection actions, and environmental data, making it difficult to provide clear root cause evidence under complex faults. Tasks such as switch opening and closing, load regulation, energy storage access, and distributed power source control are often executed in isolation, lacking a unified collaborative handling logic, making it difficult to continuously advance the fault isolation and power restoration process. These problems make it difficult for distribution networks to complete interpretable, collaborative, and recoverable fault self-healing handling under complex operating scenarios. Therefore, how to improve the continuous diagnosis and collaborative recovery capabilities of distribution network fault self-healing handling under complex operating conditions has become a challenge for the industry. Summary of the Invention

[0004] This application provides a distribution network fault self-healing system based on large model knowledge reasoning and multi-agent collaborative operation, which can improve the continuous diagnosis and collaborative recovery capabilities of distribution network fault self-healing under complex operating conditions.

[0005] In a first aspect, this application provides a self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation, comprising the following steps: The feature fusion module is used to collect multi-source heterogeneous data from the power distribution network in real time and perform feature fusion to generate a composite feature vector that integrates electrical quantities and semantic information. The reasoning and diagnosis module is used to guide the pre-trained power big data model to perform logical reasoning on the composite feature vector through alarm prompts and output a natural language reasoning chain. The root cause tracing module is used to construct a fault knowledge graph and perform fault root cause identification and causal reasoning based on the fault knowledge graph to obtain the root cause determination conclusion and causal reasoning evidence chain of the distribution network fault. The agent collaboration module is used to decode the natural language inference chain into logical constraints of the action space of the isolation layer agent and the recovery layer agent, and to determine the security constraints of the action strategies of the isolation layer agent and the recovery layer agent based on the root cause determination conclusion and the causal inference evidence chain. The instruction generation module is used to generate fault self-healing control instructions for the power distribution network based on all logical constraints and all security constraints, and then issue them to the power distribution equipment.

[0006] In this embodiment, the real-time acquisition of multi-source heterogeneous data from the power distribution network and the fusion of features to generate a composite feature vector that integrates electrical quantities and semantic information specifically include: The collected multi-source heterogeneous data is timestamped and missing values ​​are imputed to generate a time-synchronized structured data sequence. The time-domain statistical features of current, voltage, and power are extracted from the structured data sequence to form a subset of electrical quantity features; Based on the electrical quantity feature subset, semantic features are encoded for non-electrical quantity text data in the structured data sequence. The encoded semantic features are then fused with the electrical quantity feature subset to generate a composite feature vector that integrates electrical quantity and semantic information.

[0007] In this embodiment, the pre-trained power big data model is guided by alarm prompts to perform logical reasoning on the composite feature vector, and the output natural language reasoning chain specifically includes: Alarm prompts are generated based on alarm signals and operating information captured when a fault occurs; Retrieve matching fault modes from the typical case library, embed the fault modes into the alarm prompt words, and generate a sequence of guiding alarm prompt words; The sequence of guiding alarm prompts and the composite feature vector are input into a pre-trained power big data model. Under the constraint of power grid topology connectivity, word-by-word probability distribution decoding is performed to generate a preliminary diagnostic reasoning chain that includes fault type and suspected section distribution. The multi-head self-attention mechanism built into the power model is used to perform a global analysis of the timing logic of the switching sequence actions and the causal relationship of electrical quantity mutations in the preliminary diagnostic inference chain, thereby generating a natural language inference chain.

[0008] In this embodiment, constructing the fault knowledge graph specifically includes: Based on a pre-built power ontology model, fault type entities, relay protection entities, and related power distribution equipment entities are identified and their attributes are filled in through equipment asset coding, establishing an initial association between fault events and equipment nodes. Relationships are extracted from the entity nodes in the initial association relationship to construct electrical connection relationships and fault propagation logic edges between devices. The protection action sequences and handling schemes of historical fault cases are injected into the corresponding entity nodes to generate a fault knowledge graph.

[0009] In this embodiment, the root cause determination and causal reasoning based on the fault knowledge graph to obtain the root cause determination conclusion and causal reasoning evidence chain of the distribution network fault specifically include: The evolution probability between nodes in the fault knowledge graph is predicted to generate an association matrix that reflects the direction of fault propagation. Cross-domain aggregation is performed on the node features in the correlation matrix to obtain multi-dimensional fusion features that include equipment status, protection actions, and environmental factors; The fault correlation degree, propagation potential, spatiotemporal consistency and protection matching degree parameters of each node are extracted from the multidimensional fusion features to form an inference feature vector; The inference feature vector is input into the root cause discrimination function, and the root cause type is nonlinearly mapped and classified in combination with the environmental correction term. The root cause determination conclusion is output, and a causal inference evidence chain is generated based on the fault propagation path in the correlation matrix.

[0010] In this embodiment, predicting the evolution probability between nodes in the fault knowledge graph and generating an association matrix reflecting the direction of fault propagation specifically includes: The probability of fault propagation is calculated based on the operating status characteristics, protection action timing, and topological connection relationship of each node in the fault knowledge graph, so as to obtain the evolution probability between each node in the fault knowledge graph. The fault state transition matrix is ​​determined based on the evolution probabilities between nodes in the fault knowledge graph. Topological physical structure constraints are applied to the fault state transition matrix to generate an correlation matrix that reflects the direction of fault propagation.

[0011] In this embodiment, the logical constraints for decoding the natural language inference chain into the action spaces of the isolation layer agent and the recovery layer agent specifically include: The natural language inference chain is structured and parsed to extract fault segment location information, switch action sequence information and load transfer path information, and action constraint description information is generated. The action permission boundaries and collaboration sequence of the isolation layer agent and the recovery layer agent are determined based on the action constraint description information. The action permission boundaries and collaboration timing of the isolation layer agent and the recovery layer agent are integrated into the logical constraints of the action space of the isolation layer agent and the recovery layer agent.

[0012] In this embodiment, the security constraints for determining the action strategies of the isolation layer agent and the recovery layer agent based on the root cause determination conclusion and the causal reasoning evidence chain specifically include: Based on the fault type and fault segment attributes in the root cause determination conclusion, determine the set of prohibited actions and the set of mandatory actions for the isolation layer agent and the recovery layer agent. Based on the protective action sequence and device status information in the causal reasoning evidence chain, determine the action priority, action time window and mutual exclusion operation rules of the isolation layer agent and the recovery layer agent; Determine the safety boundaries of the isolation layer agent and the recovery layer agent after their actions are executed based on the fault propagation path in the causal reasoning evidence chain. The prohibited action set, the required action set, the action priority, the action time window, the mutual exclusion operation rules, and the security boundary are integrated into the security constraints of the action policies of the isolation layer agent and the recovery layer agent.

[0013] In this embodiment, determining the security boundaries of the isolation layer agent and the recovery layer agent after their actions are executed based on the fault propagation path in the causal reasoning evidence chain specifically includes: The fault propagation path in the causal reasoning evidence chain is analyzed, and the device node sequence, switch action sequence and network topology connection relationship involved in the fault propagation path are extracted to generate path description information; Extract the voltage limit parameters, branch power flow limit parameters, and equipment operation safety margin parameters corresponding to each node in the device node sequence from the causal reasoning evidence chain, and generate a set of safety constraint parameters; Based on the network topology connections in the path description information and the set of security constraint parameters, and in conjunction with the real-time operating status of the power grid, the security boundaries after the actions of the isolation layer agent and the recovery layer agent are determined.

[0014] In this embodiment, generating fault self-healing control commands for the power distribution network based on all logical constraints and all security constraints, and issuing them to the power distribution equipment, specifically includes: All logical constraints are classified according to the fault section isolation task, load transfer task, and safety verification task. The classified logical constraints are then mapped to the corresponding control objectives according to the operation sequence of each type of task, generating an initial control scheme sequence. Based on all security constraints, each candidate control scheme in the initial control scheme sequence is subjected to security verification, and a security-verified control instruction sequence is generated. The action sequence and execution time window of each control instruction in the control instruction sequence that has been verified for safety are sorted and conflict resolution is performed to generate fault self-healing control instructions for the distribution network. The fault self-healing control command is then sent to the power distribution equipment.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The system comprises the following modules: a feature fusion module for real-time acquisition of multi-source heterogeneous data from the distribution network and feature fusion to generate a composite feature vector integrating electrical quantities and semantic information; a reasoning and diagnosis module for guiding a pre-trained power large-scale model to perform logical reasoning on the composite feature vector using alarm prompts and outputting a natural language reasoning chain; a root cause tracing module for constructing a fault knowledge graph and performing fault root cause identification and causal reasoning based on the fault knowledge graph to obtain the root cause determination conclusion and causal reasoning evidence chain for the distribution network fault; an agent collaboration module for decoding the natural language reasoning chain into logical constraints for the action spaces of the isolation layer agent and the recovery layer agent, and determining the safety constraints for the action strategies of the isolation layer agent and the recovery layer agent based on the root cause determination conclusion and the causal reasoning evidence chain; and an instruction generation module for generating fault self-healing control instructions for the distribution network based on all logical constraints and all safety constraints and issuing them to the distribution equipment.

[0016] Therefore, this application demonstrates that fault self-healing control commands for the distribution network can be generated and issued to the distribution equipment through all logical and security constraints. Firstly, by using a feature fusion module to collect multi-source heterogeneous data from the distribution network in real time and perform feature fusion, electrical quantities, alarm semantics, and operating status can be unified into a composite feature vector, providing a more complete data foundation for fault identification under complex operating conditions and reducing misjudgments and omissions caused by judging a single electrical quantity. Secondly, by using alarm prompts to guide a pre-trained power big data model to perform logical reasoning on the composite feature vector and output a natural language reasoning chain, a clear reasoning relationship can be established between alarm phenomena, equipment status, and operating semantics, making the fault diagnosis process readable, verifiable, and traceable, reducing the impact of fragmented alarm information on the handling process. Furthermore, by constructing… By constructing a fault knowledge graph and performing fault root cause identification and causal reasoning, fault phenomena, protection actions, equipment status, and environmental factors can be linked to root cause determination conclusions and causal reasoning evidence chains. Furthermore, by decoding the natural language reasoning chain into logical constraints for each agent's action space, and combining the root cause determination conclusions and causal reasoning evidence chains to determine the safety constraints of action strategies, actions such as switch opening and closing, load regulation, energy storage access, and distributed power control can be coordinated in an orderly manner under unified logic and safety boundaries, reducing operational conflicts caused by fragmented execution by multiple control subjects. Finally, by generating fault self-healing control commands through logical and safety constraints and issuing them to power distribution equipment, diagnostic basis, root cause conclusions, and action constraints can be transformed into executable control commands, enabling the power distribution network to complete fault isolation and power restoration in more evidence-based and orderly fault scenarios.

[0017] In summary, the technical solution adopted in this application can improve the continuous diagnosis and collaborative recovery capabilities of power distribution network fault self-healing under complex operating conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a module structure diagram of a self-healing system for distribution network faults based on large model knowledge reasoning and multi-agent collaborative operation, provided in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application provides a self-healing system for power distribution networks based on large-scale model knowledge reasoning and multi-agent collaboration. Its core components are: a feature fusion module, used to collect multi-source heterogeneous data from the power distribution network in real time and perform feature fusion to generate a composite feature vector integrating electrical quantities and semantic information; a reasoning and diagnosis module, used to guide a pre-trained large-scale power model to perform logical reasoning on the composite feature vector through alarm prompts, outputting a natural language reasoning chain; a root cause tracing module, used to construct a fault knowledge graph and perform fault root cause identification and causal reasoning based on the fault knowledge graph, obtaining the root cause determination conclusion and causal reasoning evidence chain for the power distribution network fault; an agent collaboration module, used to decode the natural language reasoning chain into logical constraints for the action spaces of isolation layer agents and recovery layer agents, and determine the safety constraints of the action strategies of isolation layer agents and recovery layer agents based on the root cause determination conclusion and the causal reasoning evidence chain; and an instruction generation module, used to generate self-healing control instructions for the power distribution network based on all logical constraints and all safety constraints and issue them to the power distribution equipment.

[0022] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 As shown in the figure, this is a module structure diagram of a distribution network fault self-healing system based on large model knowledge reasoning and multi-agent collaboration according to this embodiment of the application. The distribution network fault self-healing system includes: a feature fusion module 100, a reasoning diagnosis module 200, a root cause tracing module 300, an agent collaboration module 400, and an instruction generation module 500, which are described below: The feature fusion module 100 is used to collect multi-source heterogeneous data of the power distribution network in real time and perform feature fusion to generate a composite feature vector that integrates electrical quantities and semantic information.

[0023] In this embodiment, the real-time acquisition of multi-source heterogeneous data from the power distribution network and the fusion of features to generate a composite feature vector that integrates electrical quantities and semantic information can be achieved through the following steps: The collected multi-source heterogeneous data is timestamped and missing values ​​are imputed to generate a time-synchronized structured data sequence. The time-domain statistical features of current, voltage, and power are extracted from the structured data sequence to form a subset of electrical quantity features; Based on the electrical quantity feature subset, the non-electrical quantity text data in the structured data sequence is semantically encoded, and the encoded semantic features are fused with the electrical quantity feature subset to generate a composite feature vector that integrates electrical quantity and semantic information.

[0024] It should be noted that, in this application, multi-source heterogeneous data refers to different types of data from electrical quantity acquisition, equipment status records, and non-electrical quantity text records during the operation of the distribution network; time-synchronized structured data sequences refer to data sequences used for feature extraction after being organized according to the same time base; electrical quantity feature subsets refer to a set of electrical quantity features reflecting the changes in current, voltage, and power of the distribution network; non-electrical quantity text data refers to text data such as inspection logs, dispatch texts, and alarm descriptions related to the operating status of the distribution network; semantic feature encoding refers to converting non-electrical quantity text data into a vector representation that can be fused with the electrical quantity feature subset; and composite feature vectors refer to feature vectors that simultaneously fuse electrical quantity and semantic information.

[0025] In specific implementation, firstly, the distribution automation master station and edge IoT agent receive transient waveform current, steady-state operating voltage, power measurement values, and non-electrical quantity text data according to a unified acquisition cycle. Using the master station's time synchronization signal as a unified time reference, data from the same device within the same acquisition cycle are written into the same data record. For late data, it is assigned to the corresponding acquisition cycle according to the most recent valid timestamp. For short-term missing current, voltage, and power measurement values, linear interpolation using adjacent sampling points is used to fill in the gaps. Missing non-electrical quantity text data retains an empty field marker. The data record after timestamp merging and data completion is used as a time-synchronized structured data sequence. Secondly, the current, voltage, and power measurement values ​​in the time-synchronized structured data sequence are read according to a fixed sliding window. The window length is pre-configured based on the distribution network sampling cycle and the duration of the fault transient. Within each window, the mean, maximum, minimum, and maximum values ​​of current, voltage, and power are read respectively. Peak-to-peak value, rate of change, and standard deviation are used as time-domain statistical features of current, voltage, and power. The time-domain statistical features of current, voltage, and power obtained within the same window are combined as a subset of electrical quantity features. Next, non-electrical quantity text data from the time-synchronized structured data sequence is read. Inspection logs, scheduling texts, and alarm descriptions are segmented, stop words are removed, and equipment names are normalized. The normalized text content is then aligned with the electrical quantity feature subset according to the equipment objects within the same time window. A text encoder is used to convert the normalized text content into a semantic vector, which is then used as the encoded semantic feature. Finally, the electrical quantity feature subset is normalized and framed. A one-dimensional convolutional neural network is used to extract the local representation corresponding to changes in electrical quantities within the window. This representation is then compressed using pooling operations to obtain an electrical quantity representation with the same dimension as the encoded semantic feature. The electrical quantity representation and the encoded semantic feature are then fused using the following formula: in, This represents a composite feature vector that integrates electrical quantities and semantic information. Electrical quantity data in a structured data sequence representing time synchronization; Represents non-electrical quantity text data; This indicates that a local representation of electrical quantities is extracted using a one-dimensional convolutional neural network; This indicates that the local representation of electrical quantities is compressed using pooling. This indicates that semantic features are encoded using a text encoder; This represents a vector concatenation operation, where the concatenated vector is used as a composite feature vector that integrates electrical quantities and semantic information.

[0026] The reasoning and diagnosis module 200 is used to guide the pre-trained power big data model to perform logical reasoning on the composite feature vector through alarm prompts and output a natural language reasoning chain.

[0027] In this embodiment, the pre-trained large-scale power model adopts a Transformer-based autoregressive language model structure, including a text lexical embedding layer, a composite feature adaptation layer, a position encoding layer, multiple sequentially connected Transformer decoding layers, and a lexical probability output layer. Each Transformer decoding layer includes a multi-head self-attention sublayer, a feedforward neural network sublayer, and corresponding residual connections and normalization structures. The text lexical embedding layer is used to convert alarm prompts, fault case texts, and diagnostic task texts into text lexical embeddings. The composite feature adaptation layer includes a linear projection layer and a feature segmentation layer, used to convert composite feature vectors that fuse electrical quantities and semantic information into feature lexical embeddings with the same vector dimension as the text lexical embeddings. The position encoding layer is used to mark the arrangement positions of text lexical units and feature lexical units in the model input sequence. The Transformer decoding layer is used to establish the correlation between alarm information, fault cases, electrical quantity changes, equipment status, and distribution network topology. The lexical probability output layer is used to output the generation probability of each candidate lexical unit in the natural language inference chain.

[0028] The training data for the power system large-scale model includes textual training data from the power sector and fault sample data from the distribution network. The textual training data includes power operation procedures, relay protection procedures, distribution equipment operation and maintenance manuals, dispatching operation procedures, and distribution network fault handling specifications, used to enable the power system large-scale model to learn distribution equipment names, fault types, protection action terminology, and dispatching operation logic. The distribution network fault sample data includes alarm signals corresponding to historical faults, switch change records, protection action records, current data, voltage data, power data, equipment operating status, distribution network topology snapshots at the time of the fault, dispatching logs, and fault diagnosis reports reviewed by power dispatching personnel. The fault diagnosis reports include at least the fault phenomenon, key alarms, fault type, suspected fault section, protection action logic, topology judgment basis, and diagnostic conclusion.

[0029] In the specific model training process, firstly, the equipment names, line names, protection device names, switch action names, and fault type names in the power sector text training data are standardized, mapping different expressions of the same equipment or the same fault type to unified power terminology. The text is then segmented according to the semantic integrity of the paragraphs, and the resulting standardized text is used as the power sector corpus. Secondly, based on this power sector corpus, a basic language model with general language understanding capabilities is further trained in the power sector. The model parameters of the text word embedding layer, Transformer decoding layer, and word probability output layer are adjusted by predicting subsequent words in each training text. The model that can recognize power sector terminology and its contextual relationships after training is used as the basic power sector model. Next, according to the fault occurrence time and equipment asset code, alarm signals, switch change records, protection action records, electrical quantity data, equipment operating status, and distribution network topology snapshots corresponding to the same historical fault are time-aligned. The aligned data is then converted according to the processing method of the feature fusion module. Historical composite feature vectors are generated, and historical alarm prompt word sequences are constructed based on the alarm signals and operating condition information corresponding to the historical faults. The historical alarm prompt word sequences and the historical composite feature vectors are used together as input to the instruction training samples. The fault diagnosis report reviewed by power dispatchers is converted into a natural language inference chain as the target output of the instruction training samples. Finally, the instruction training samples are used to perform supervised instruction training on the basic model of the power field. During the training process, the historical composite feature vectors are converted into feature word embeddings through the composite feature adaptation layer. The feature word embeddings are then inserted after the operating condition information field corresponding to the historical alarm prompt word sequence, so that the text words corresponding to the historical alarm prompt words and the feature words corresponding to the historical composite feature vectors participate in multi-head self-attention calculation. Then, the model parameters of the composite feature adaptation layer, the Transformer decoding layer, and the word probability output layer are updated in reverse according to the word difference between the natural language inference chain generated by the model and the target output. The model that has completed supervised instruction training is used as a pre-trained large power model.

[0030] In this embodiment, the pre-trained power big data model is guided by alarm prompts to perform logical reasoning on the composite feature vectors and output a natural language reasoning chain, which can be achieved through the following steps: Alarm prompts are generated based on alarm signals and operating information captured when a fault occurs; Retrieve matching fault modes from the typical case library, embed the fault modes into the alarm prompt words, and generate a sequence of guiding alarm prompt words; The sequence of guiding alarm prompts and the composite feature vector are input into a pre-trained power big data model. Under the constraint of power grid topology connectivity, word-by-word probability distribution decoding is performed to generate a preliminary diagnostic reasoning chain that includes fault type and suspected section distribution. The multi-head self-attention mechanism built into the power model is used to perform a global analysis of the timing logic of the switching sequence actions and the causal relationship of electrical quantity mutations in the preliminary diagnostic inference chain, thereby generating a natural language inference chain.

[0031] It should be noted that, in this application, alarm signals represent abnormal prompts triggered by changes in the status of protection devices, switches, and monitoring alarm records when a distribution network fault occurs; operating condition information represents operational information reflecting the operating environment, equipment status, and topology status of the distribution network when a fault occurs; alarm prompt words represent prompt text used to guide the pre-trained power big data model into the distribution network fault reasoning task; typical case library represents a dataset used to store historical fault modes and handling experience; fault mode represents a case pattern that can reflect the fault type, fault phenomenon, and handling logic; guiding alarm prompt word sequence represents a sequence of prompt words used to constrain the reasoning direction of the pre-trained power big data model after embedding the fault mode; and the pre-trained power big data model represents a large-scale power big data model trained on power domain corpus and used for distribution network fault semantic reasoning. The model includes: a power grid topology connectivity constraint, which represents the physical constraints used to limit the inference results to conform to the distribution network topology; a word-by-word probability distribution decoding, which represents the decoding process of gradually outputting diagnostic text according to the word generation probability; a preliminary diagnostic inference chain, which represents the initial inference text containing the fault type and suspected segment distribution; a multi-head self-attention mechanism, which represents the computational mechanism used to associate the inference text and the electrical quantity change relationship in multiple attention subspaces; a switching sequence action timing logic, which represents the constraint relationship between the order of switching actions on the fault isolation and power restoration process; an electrical quantity mutation causal association, which represents the causal relationship between abnormal changes in current, voltage, and power and fault evolution; and a natural language inference chain, which represents the fault inference process output in natural language form and used for subsequent root cause tracing and agent collaborative decision-making.

[0032] In specific implementation, firstly, when a fault occurs, the tripping alarm, overcurrent alarm, grounding alarm, and switch change records uploaded by the protection device are read. Simultaneously, the current line load, voltage over-limit status, switch status, and network topology status are read. Fault type candidate words, abnormal equipment names, fault occurrence time, and current operating scenario are written into the Role-Context-Task prompt word template. The Role field contains the distribution network fault diagnosis role, the Context field contains the alarm signal and operating condition information, and the Task field contains the fault type judgment and suspected section reasoning tasks. The prompt text after field filling is used as the alarm prompt word. Secondly, a search enhancement generation method is used to retrieve historical cases matching the alarm signal and operating condition information from the typical case library. During the search, historical cases matching the alarm signal and operating condition information are retrieved. Using alarm type, protection action name, switch change sequence, and line topology location as search keys, the system reads the fault type, abnormal evolution process, and handling logic from the matching cases. The read case patterns are then used as fault modes, which are written into the case reference field of the alarm prompt words. The prompt words after case embedding are arranged in the order of fault phenomenon, topology location, protection action, and reasoning task to form a guiding alarm prompt word sequence. Next, the guiding alarm prompt word sequence is converted into a text word sequence and input together with the composite feature vector into a pre-trained power grid big data model. During the decoding process, the power grid topology connectivity constraints are read, and fault segment words that do not conform to the topology connection relationship are probabilistically suppressed. Then, a predictive sampling algorithm is used to output the diagnostic text word by word. The word-by-word probability distribution decoding process is represented by the following formula: in, This represents the composite feature vector of the input pre-trained power large model and the sequence of guiding alarm prompts; Indicates the target diagnostic conclusion; Indicates the first One output word; Indicates the first The sequence of words that have been generated before the output word; Represents the connectivity constraints of the power grid topology; This represents the total number of output tokens; the decoded diagnostic text containing fault types and suspected segment distributions is used as the preliminary diagnostic inference chain; finally, a multi-head self-attention mechanism is used to establish attention associations for fault type tokens, suspected segment tokens, switch action tokens, and electrical quantity mutation tokens in the preliminary diagnostic inference chain, projecting the composite feature vector and the preliminary diagnostic inference chain into a query matrix, a key matrix, and a value matrix, and calculating the attention output using the following formula: in, This represents the query matrix obtained by projecting the composite feature vector and the preliminary diagnostic inference chain. This represents the key matrix obtained by projecting the composite feature vector and the preliminary diagnostic inference chain. This represents the value matrix obtained by projecting the composite feature vector and the preliminary diagnostic inference chain. The dimension normalization factor represents the key matrix; The transpose of the key matrix is ​​represented; the reasoning segments that are consistent with the timing logic of the switching sequence and the causal relationship of electrical quantity mutations are selected by attention weight, and the segment descriptions that are inconsistent with the grid topology connectivity constraints are eliminated. The retained reasoning segments are organized into natural language reasoning chains according to the fault phenomenon, fault type, suspected segment distribution and reasoning basis.

[0033] The root cause tracing module 300 is used to construct a fault knowledge graph and perform fault root cause identification and causal reasoning based on the fault knowledge graph to obtain the root cause determination conclusion and causal reasoning evidence chain of the distribution network fault.

[0034] In this embodiment, the construction of the fault knowledge graph can be achieved through the following steps: Based on a pre-built power ontology model, fault type entities, relay protection entities, and related power distribution equipment entities are identified and their attributes are filled in through equipment asset coding, establishing an initial association between fault events and equipment nodes. Relationships are extracted from the entity nodes in the initial association relationship to construct electrical connection relationships and fault propagation logic edges between devices. The protection action sequences and handling schemes of historical fault cases are injected into the corresponding entity nodes to generate a fault knowledge graph.

[0035] It should be noted that the pre-built power ontology model in this application represents a knowledge structure used to define the category relationships and attribute ranges between fault type entities, relay protection entities, and related power distribution equipment entities; the equipment asset code represents a code used to uniquely identify power distribution equipment and associate it with equipment operation records; the fault type entity represents an entity node used to describe the fault categories of the power distribution network; the relay protection entity represents an entity node used to describe protection devices, protection actions, and protection settings; the related power distribution equipment entities represent equipment entity nodes such as lines, switches, transformers, and busbars related to the fault event; attribute filling represents the attribute completion process of writing equipment type, operating condition, health score, and spatial coordinates into the entity node; and the initial association relationship represents the initial relationship between the fault event and the equipment node. The fault knowledge graph consists of several parts: a step-by-step connection relationship; an entity node representing a node object in the fault knowledge graph that carries entity attributes and relational connections; a relation extraction representing the process of extracting connection relationships, action relationships, and propagation relationships from entity nodes; an electrical connection relationship representing the relationship between related power distribution equipment entities that conforms to the power distribution network topology; a fault propagation logic edge representing the relational edge used to describe the direction of fault phenomenon propagation along equipment nodes; a protection action sequence representing the action sequence obtained by arranging relay protection actions in the order of occurrence in historical fault cases; a handling plan representing the handling plan used to isolate fault sections and restore power supply in historical fault cases; and a fault knowledge graph representing the graph structure used to organize fault type entities, relay protection entities, related power distribution equipment entities, and their fault propagation relationships.

[0036] In specific implementation, firstly, the category, attribute, and relationship fields of fault type entities, relay protection entities, and related power distribution equipment entities in the pre-set power ontology model are read. Using the equipment asset code as the primary key for entity matching, the equipment names, protection device names, and fault category names in fault records, protection action records, and equipment ledgers are standardized for matching. For text-based fault records, power named entity recognition is used to identify fault type entities, relay protection entities, and related power distribution equipment entities from the text. Then, the equipment type, operating condition, health score, and spatial coordinates corresponding to the entity nodes are completed using the equipment asset code. The fault type entities, relay protection entities, and related power distribution equipment entities with completed attribute filling are taken as entity nodes, and the records connecting the fault events to the corresponding entity nodes are taken as the initial association between the fault events and the equipment nodes. The vectorized record of the entity node is represented by the following formula: in, This represents a sequence of knowledge vectors obtained by transforming entity nodes; This represents the device vector corresponding to the relevant power distribution equipment entity; This represents the protection vector corresponding to the relay protection entity; This represents the environmental vector corresponding to the operating condition information; The first step is to use a topology vector representing the network topology connection relationship between relevant power distribution equipment entities. This knowledge vector sequence is then used as a vectorized record of the entity nodes in the initial association relationship. Next, the network topology connection relationship between entity nodes in the initial association relationship is read according to equipment terminal connections, upstream and downstream line connections, and switch opening and closing positions. The topology connection record between adjacent equipment nodes is used as the electrical connection relationship. Then, according to the occurrence order of protection action sequences in historical fault cases, the action triggering relationship between fault type entities, relay protection entities, and relevant power distribution equipment entities is read. The transmission relationship between fault phenomena, protection actions, and equipment status is used as the fault propagation logic edge. Subsequently, the protection action sequence of historical fault cases is written into the corresponding relay protection entity, and the handling plan is written into the corresponding fault type entity and relevant power distribution equipment entity. This is stored in a triplet format of "entity node—relationship—entity node". The fault knowledge graph is represented by the following formula: in, Represents a fault knowledge graph; Represents the head entity node; Represents the tail entity node; This indicates the relationship between the head entity node and the tail entity node; It represents a set of entities; it represents a set of relationships. The set of triples containing electrical connection relationships, fault propagation logic edges, protection action sequences, and disposal schemes is used as the fault knowledge graph.

[0037] In this embodiment, the root cause determination and causal reasoning based on the fault knowledge graph to obtain the root cause determination conclusion and causal reasoning evidence chain of the distribution network fault can be achieved by the following steps: The evolution probability between nodes in the fault knowledge graph is predicted to generate an association matrix that reflects the direction of fault propagation. Cross-domain aggregation is performed on the node features in the correlation matrix to obtain multi-dimensional fusion features that include equipment status, protection actions, and environmental factors; The fault correlation degree, propagation potential, spatiotemporal consistency and protection matching degree parameters of each node are extracted from the multidimensional fusion features to form an inference feature vector; The inference feature vector is input into the root cause discrimination function, and the root cause type is nonlinearly mapped and classified in combination with the environmental correction term. The root cause determination conclusion is output, and a causal inference evidence chain is generated based on the fault propagation path in the correlation matrix.

[0038] It should be noted that, in this application, node features represent the feature representations of equipment status, protection actions, and environmental factors corresponding to entity nodes in the fault knowledge graph; cross-domain aggregation represents the feature update process of neighborhood fusion of node features from different sources according to the connection relationship between entity nodes; multi-dimensional fusion features represent fusion features including equipment status, protection actions, and environmental factors; fault correlation represents a parameter reflecting the strength of the correlation between entity nodes and distribution network faults; propagation potential represents a parameter reflecting the trend of fault propagation along the fault propagation path; spatiotemporal consistency represents a parameter reflecting the degree of coordinated occurrence of multiple abnormal nodes in terms of occurrence time and topological location; protection matching degree parameter represents a parameter reflecting the degree of deviation between the protection action setpoint triggering value and the measured value; and inference feature vector represents a feature vector composed of fault correlation, propagation potential, spatiotemporal consistency, and protection matching degree parameters. The root cause discrimination function represents a classification discrimination function that outputs the matching confidence of each root cause type based on the inference feature vector, the root cause type dictionary, and the environmental correction term; the root cause type dictionary represents a set of categories consisting of root cause types archived in historical fault cases; the root cause type matching confidence represents a numerical result of the degree of matching between the inference feature vector and each root cause type; the environmental correction term represents a correction term used to correct the influence of environmental factors on root cause type discrimination; the root cause type represents the category to which the cause of the distribution network fault belongs; the nonlinear mapping classification represents the mapping process that converts the inference feature vector into the root cause type classification result; the root cause determination conclusion represents the determination result of the root cause type of the distribution network fault; the fault propagation path represents the path of fault transmission between entity nodes; the causal inference evidence chain represents the chain record of the fault propagation path and related evidence units used to support the root cause determination conclusion.

[0039] In specific implementation, firstly, the evolution probability between nodes in the fault knowledge graph is predicted to generate an association matrix reflecting the fault propagation direction. Secondly, using the association matrix as the adjacency input of the graph reasoning network, the equipment status field, protection action field, and environmental factor field of each entity node in the fault knowledge graph are read. The equipment status field includes switch open / closed status, voltage over-limit status, branch power flow status, and equipment operating status. The protection action field includes protection action name, action time, protection setting value, and measured trigger value. The environmental factor field includes lightning location signal, temperature and humidity record, and load rate record. The voltage value, branch power flow value, load rate, and temperature and humidity record are normalized according to the corresponding operating limits. The process involves converting switch open / closed states, protection action names, and equipment operating states into encoded values ​​according to a preset state dictionary. Then, the normalized results and encoded values ​​of the same entity node are concatenated in a fixed order of equipment status, protection action, and environmental factors. The concatenated result is used as the node feature of that entity node. In the graph convolutional layer, the node features of adjacent entity nodes are read according to the directed connection relationships recorded in the correlation matrix. Each entity node retains its own node features and incorporates the equipment status, protection action, and environmental factors transmitted from upstream adjacent entity nodes. The node features transmitted from adjacent entity nodes are compressed using mean aggregation. The compressed result is then concatenated with the entity node's own node features and written to that entity node. The node representation after input is a multi-dimensional fusion feature including equipment status, protection actions, and environmental factors. Then, within the multi-dimensional fusion feature of each entity node, the connection strength between the entity node and the distribution network fault, the transmission direction of upstream nodes along the fault propagation path, the action time difference and topological distance of multiple anomalies, and the deviation between protection settings and measured trigger values ​​are read. The connection strength between the entity node and the distribution network fault is used as the fault correlation degree; the transmission trend from upstream nodes to the current entity node along the fault propagation path is used as the propagation potential energy; the coordination situation where the action time difference falls within the same fault time window and the topological distance is within the adjacent range is used as the spatiotemporal consistency; and the deviation between protection settings and measured trigger values ​​is used as the spatiotemporal consistency. Deviation conditions are used as protection matching degree parameters, and are concatenated in a fixed order of fault correlation, propagation potential energy, spatiotemporal consistency, and protection matching degree parameters. The concatenated vector is used as the inference feature vector. Finally, the inference feature vector is written into the input of the root cause discrimination function in a fixed order of fault correlation, propagation potential energy, spatiotemporal consistency, and protection matching degree parameters. The root cause type dictionary archived in the historical fault cases is read. The root cause type dictionary is classified according to environmental causes, operating conditions, equipment status, and protection actions, including at least lightning flashover, external force damage, tree discharge, foreign object bridging, overload, insulation degradation, equipment aging, wire breakage, poor joint contact, protection maloperation, and protection failure to operate.The root cause discrimination function is implemented using a feedforward classifier trained on historical fault cases. During training, the fault correlation, propagation potential, spatiotemporal consistency, and protection matching parameters from historical fault cases are used as training sample features, and the corresponding archived root cause types are used as sample labels. The classifier parameters are updated through error backpropagation to obtain the discrimination parameters used for root cause type classification. During runtime, the current inference feature vector is input into the trained feedforward classifier, which outputs the root cause type matching confidence score for each class. When there is a lightning location signal in the environmental factors that coincides with the fault time window, the root cause type matching confidence score corresponding to the lightning flashover is increased through an environmental correction term. When the load rate is close to or exceeds the operating limit... When the environmental correction term is used, the confidence level of the root cause type matching corresponding to the overload is increased. When there are records of overdue service or insulation deterioration in the equipment status record, the confidence level of the root cause type matching corresponding to the equipment deterioration is increased by the environmental correction term. The root cause type with the highest confidence level of root cause type matching after environmental correction term is taken as the root cause determination conclusion. At the same time, the corresponding fault type entity, relay protection entity, related power distribution equipment entity, protection action sequence, fault correlation degree, propagation potential energy, spatiotemporal consistency and protection matching degree parameters are read along the fault propagation path in the correlation matrix. They are then connected in the order of fault occurrence, protection action, equipment status change and root cause determination conclusion, and the connection result is used as the causal reasoning evidence chain.

[0040] In this embodiment, predicting the evolution probabilities between nodes in the fault knowledge graph and generating an association matrix reflecting the direction of fault propagation can be achieved through the following steps: The probability of fault propagation is calculated based on the operating status characteristics, protection action timing, and topological connection relationship of each node in the fault knowledge graph, so as to obtain the evolution probability between each node in the fault knowledge graph. The fault state transition matrix is ​​determined based on the evolution probabilities between nodes in the fault knowledge graph. Topological physical structure constraints are applied to the fault state transition matrix to generate an correlation matrix that reflects the direction of fault propagation.

[0041] It should be noted that, in this application, the operating state characteristics represent the characteristics of voltage, current, power, switch status, and protection status of equipment nodes; the protection action sequence represents the timing relationship of protection device startup, circuit breaker tripping, and switch change arranged according to the action time; the topology connection relationship represents the electrical connection relationship, connection direction, and switch on / off relationship between each equipment node; the evolution probability represents the probability value of a fault state propagating from one equipment node to an adjacent equipment node; the fault state transition matrix represents the matrix of evolution probabilities written between each node according to the arrangement order of equipment nodes; the topology physical structure constraint represents the restriction rules for filtering out invalid fault propagation relationships based on the actual electrical connection relationship of the distribution network, switch on / off status, and power supply direction; and the correlation matrix represents the matrix recording the effective fault propagation direction and corresponding evolution probability between each equipment node.

[0042] In specific implementation, firstly, the voltage, current, power, switch status, and protection status of each node are read from the fault knowledge graph according to the equipment asset code. The voltage, current, and power are normalized using the rated voltage, rated current, and rated capacity of the corresponding equipment. The normalized voltage, current, and power are then arranged in a fixed field order with the switch status and protection status, and the arrangement result is used as the operating status feature of each node. The records of protection device start-up, circuit breaker tripping, and switch change are arranged in sequence according to the action time. The action type, action node, and time interval between adjacent actions are numerically encoded, and the encoding result is used as the protection action sequence. An adjacency matrix is ​​established based on the fault propagation logic edges in the fault knowledge graph. Nodes with electrical connections and switches on the connection path in the on state are written with a value of 1 for their corresponding elements, and other nodes are written with a value of 0 for their corresponding elements. The adjacency matrix is ​​used as the topological connection relationship. The operating status feature of each node and the corresponding protection action sequence are vector-concatenated, and the concatenation result is used as the node input feature. The node input feature and the adjacency matrix are then input into the graph attention network. Graph attention networks map the input features of each node to the same feature space by sharing a linear transformation matrix. For each pair of adjacent nodes with an element of 1 in the adjacency matrix, the mapping features of the starting node and the target node are concatenated. Then, the attention score of the node pair is calculated through the attention vector, and the attention score is nonlinearly transformed using a linear rectified function with leakage. The attention scores of all adjacent nodes from the same starting node are normalized using a normalized exponential function so that the sum of the attention weights corresponding to the same starting node is 1. The normalized attention weights are used as the evolution probability between nodes in the fault knowledge graph.Before use, the graph attention network is trained under supervision using historical fault cases. Pairs of adjacent nodes with a confirmed fault propagation relationship, as confirmed by fault recordings, protection action records, and fault segment location results, are labeled as positive samples. Pairs of adjacent nodes with topological connections but no fault propagation are labeled as negative samples. The operating state characteristics at the time of the historical fault, the timing of protection actions, and the topological connections are used as training inputs. A binary cross-entropy loss is used to calculate the difference between the attention weights and the sample labels. The shared linear transformation matrix and attention vector are updated using backpropagation and adaptive moment estimation optimization algorithms. Training stops when the loss value of the verification sample no longer decreases after several consecutive training rounds. The graph attention network with fixed parameters is then used for evolution probability calculation. Next, the row and column indices of the matrix are set according to the arrangement order of the equipment asset codes. The starting node of fault propagation is used as the node corresponding to the row index, and the target node of fault propagation is used as the node corresponding to the column index. The evolution probability between each node is calculated. Write the values ​​at the corresponding matrix positions. For node pairs without topological connections and positions where a node points to itself, write zero values. Use the matrix after writing the evolution probabilities as the fault state transition matrix. Finally, establish a topology mask matrix with the same dimension as the fault state transition matrix based on the topological connections at the time of the fault. For node pairs with actual electrical connections, switches on the connection path being in the ON state, and propagation direction conforming to the current power supply direction, write the value 1 at the corresponding position. For node pairs without actual electrical connections, switches on the connection path being in the OFF state, or propagation direction not conforming to the current power supply direction, write the value 0 at the corresponding position. Multiply the matrix elements at the same position in the topology mask matrix and the fault state transition matrix item by item to filter out evolution probabilities that do not conform to the topological physical structure constraints. Then normalize the non-zero matrix elements retained in each row so that the sum of the evolution probabilities of the same starting node pointing to adjacent nodes is 1. Use the normalized matrix as the correlation matrix reflecting the fault propagation direction.

[0043] The agent collaboration module 400 is used to decode the natural language inference chain into logical constraints of the action space of the isolation layer agent and the recovery layer agent, and to determine the security constraints of the action strategies of the isolation layer agent and the recovery layer agent based on the root cause determination conclusion and the causal inference evidence chain.

[0044] In this embodiment, decoding the natural language inference chain into logical constraints for the action spaces of the isolation layer agent and the recovery layer agent can be achieved through the following steps: The natural language inference chain is structured and parsed to extract fault segment location information, switch action sequence information and load transfer path information, and action constraint description information is generated. The action permission boundaries and collaboration sequence of the isolation layer agent and the recovery layer agent are determined based on the action constraint description information. The action permission boundaries and collaboration timing of the isolation layer agent and the recovery layer agent are integrated into the logical constraints of the action space of the isolation layer agent and the recovery layer agent.

[0045] It should be noted that, in this application, the structured parsing refers to the semantic parsing process that converts natural language reasoning chains into field-based task information that can be recognized by intelligent agents; the fault segment location information refers to the location information used to limit the fault isolation range and the location of its boundary equipment; the switch action sequence information refers to the timing information used to limit the sequential relationship between the opening and closing actions of related switches; the load transfer path information refers to the path information used to limit the equipment nodes and interconnecting switches through which power is restored after fault isolation; the action constraint description information refers to the field-based description information that records the action object, action type, and sequential relationship of actions; and the isolation layer intelligent agent refers to... The decision-making entity responsible for controlling the boundary switches of the faulty section to perform isolation actions; the recovery layer agent represents the decision-making entity responsible for controlling the activation of backup power and the reconfiguration of the power grid; the action permission boundary represents the permission range that the agent is allowed to control and the range of action types that can be selected; the cooperation sequence represents the execution order that limits the order of action triggering between the isolation layer agent and the recovery layer agent; the action space represents the set of control actions that the agent can select according to its corresponding task responsibilities; the logical constraints of the action space represent the set of rules used to limit the optional actions, prohibited actions, and action triggering order of the isolation layer agent and the recovery layer agent.

[0046] In specific implementation, firstly, the power named entity recognition method is used to identify feeder names, faulty equipment names, sectionalizing switch names, tie switch names, and backup power supply names from the natural language inference chain, and the identification results are mapped to the corresponding equipment nodes in the distribution network topology through equipment asset coding; based on positional relation words such as "located in," "adjacent," "upstream," and "downstream" in the inference chain, the faulty equipment and its adjacent boundary switches are extracted, and the topological location of the obtained equipment nodes and boundary switches is used as the fault section location information; based on action sequence words such as "open first," "close later," "isolation completed," and "closing prohibited," the relevant switches are extracted. The action types and their sequence relationships are used to determine the order of actions, which is then used as the switching action sequence information. A topology path is matched along the power supply, tie switches, and load nodes to be restored recorded in the natural language inference chain. The sequence of equipment nodes that can bypass the fault section and connect to the load to be restored is used as the load transfer path information. This information is organized according to the action object, action type, action order, and path node fields, and the resulting fielded information is used as the action constraint description information. Next, based on the fault section location information in the action constraint description information, the boundary switches on both sides of the fault section and their tripping actions are included in the selectable range of the isolation layer intelligent agent. The system defines the scope of the faulty section and includes the closing actions of the corresponding equipment within the faulty section within the prohibited scope of the isolation layer intelligent agent. The resulting equipment scope and action type serve as the action permission boundaries of the isolation layer intelligent agent. Based on the load transfer path information, the system includes the standby power supply activation action, tie switch closing action, and grid reconfiguration action within the selectable scope of the recovery layer intelligent agent. Equipment control actions not belonging to the load transfer path are included within the prohibited scope of the recovery layer intelligent agent, and the resulting equipment scope and action type serve as the action permission boundaries of the recovery layer intelligent agent. Simultaneously, based on the switch action sequence information, the isolation layer intelligent agent completes the opening and closing of the faulty section boundary switch. The moment of obtaining the action receipt serves as the trigger condition for the recovery layer agent to initiate the corresponding recovery action. The obtained trigger sequence is used as the cooperation timing between the isolation layer agent and the recovery layer agent. Finally, the optional actions, prohibited actions, and action permission boundaries of the isolation layer agent and the recovery layer agent are written into the corresponding action space, and the action trigger conditions are set according to the cooperation timing. This ensures that the recovery layer agent only selects the backup power supply, tie switch closing, or grid reconfiguration action after the fault section is isolated. The obtained action selection rules and action trigger rules are integrated into the logical constraints of the action space of the isolation layer agent and the recovery layer agent.

[0047] In this embodiment, the security constraints for determining the action strategies of the isolation layer agent and the recovery layer agent based on the root cause determination conclusion and the causal reasoning evidence chain can be achieved through the following steps: Based on the fault type and fault segment attributes in the root cause determination conclusion, determine the set of prohibited actions and the set of mandatory actions for the isolation layer agent and the recovery layer agent. Based on the protective action sequence and device status information in the causal reasoning evidence chain, determine the action priority, action time window and mutual exclusion operation rules of the isolation layer agent and the recovery layer agent; Determine the safety boundaries of the isolation layer agent and the recovery layer agent after their actions are executed based on the fault propagation path in the causal reasoning evidence chain. The prohibited action set, the required action set, the action priority, the action time window, the mutual exclusion operation rules, and the security boundary are integrated into the security constraints of the action policies of the isolation layer agent and the recovery layer agent.

[0048] It should be noted that, in this application, the fault type refers to the diagnostic results used to distinguish between short-circuit faults, ground faults, and other distribution network fault categories; the fault section attribute refers to the section information reflecting the equipment range, topological location, and isolation status involved in the fault section; the prohibited action set refers to the set of control actions that restrict the agent's selection to avoid fault expansion or power loss in non-faulty areas; the mandatory action set refers to the set of control actions required to complete fault section isolation or power restoration; the protection action sequence refers to the sequence of actions recorded by relay protection devices and switching equipment according to the fault evolution process; the equipment status information refers to the information reflecting the current operating status of circuit breakers, sectionalizing switches, tie switches, and backup power supplies; the action priority refers to the execution level of the isolation layer agent and the recovery layer agent in selecting different control actions; the action time window refers to the time range within which the corresponding control action is allowed to start and complete; the mutual exclusion operation rule refers to the rule used to prevent control operations with conflicting actions from being selected simultaneously or executed continuously; and the safety constraint of the action strategy refers to the set of safety rules used to limit the action selection range, execution order, and execution conditions of the isolation layer agent and the recovery layer agent.

[0049] In specific implementation, firstly, the fault type, equipment nodes involved in the fault section, and isolation range of the fault section recorded in the root cause determination conclusion are read. The tripping action of the fault section boundary switch and the isolation action of the faulty equipment are matched with the preset fault handling rules. The actions required to cut off the fault current path are taken as the set of mandatory actions for the isolation layer intelligence, and the actions that may reconnect the fault section or expand the power outage range are taken as the set of prohibited actions for the isolation layer intelligence. At the same time, the actions of connecting the backup power supply and reconfiguring the grid after the fault section is isolated are taken as the set of mandatory actions for the recovery layer intelligence, and the actions that may reverse power supply to the fault section that has not been isolated are taken as the set of prohibited actions for the recovery layer intelligence. Secondly, the protection action sequence is arranged according to the occurrence time of each protection action in the causal reasoning evidence chain, and the current equipment status of the relevant circuit breakers, sectionalizing switches, tie switches, and backup power supplies is read. The tripping action of the fault section boundary switch and the isolation action of the faulty equipment are matched with the preset fault handling rules. The tripping action of the switch is set to have a higher priority than the power restoration action. The allowable operation period between the occurrence of the protection action and the confirmation of the corresponding equipment status is used as the action time window. The tripping and closing actions of the same switch, as well as the restoration actions before the isolation is completed, are marked as mutually exclusive. The resulting action conflict relationship is used as the mutual exclusion operation rule. Next, the safety boundary after the action execution of the isolation layer agent and the restoration layer agent is determined according to the fault propagation path in the causal reasoning evidence chain. Finally, the prohibited action set is used to exclude the control actions that the isolation layer agent and the restoration layer agent cannot select. The mandatory action set is used as the necessary actions for the corresponding agent to complete fault isolation and power restoration. The action priority, action time window, and mutual exclusion operation rule are used to limit the selection order and execution conditions of each control action. Combined with the safety boundary, they are integrated in a unified manner. The integrated safety rule is used as the safety constraint for the action strategy of the isolation layer agent and the restoration layer agent.

[0050] In this embodiment, determining the safety boundaries of the isolation layer agent and the recovery layer agent after their actions are executed based on the fault propagation path in the causal reasoning evidence chain can be achieved through the following steps: The fault propagation path in the causal reasoning evidence chain is analyzed, and the device node sequence, switch action sequence and network topology connection relationship involved in the fault propagation path are extracted to generate path description information; Extract the voltage limit parameters, branch power flow limit parameters, and equipment operation safety margin parameters corresponding to each node in the device node sequence from the causal reasoning evidence chain, and generate a set of safety constraint parameters; Based on the network topology connections in the path description information and the set of security constraint parameters, and in conjunction with the real-time operating status of the power grid, the security boundaries after the actions of the isolation layer agent and the recovery layer agent are determined.

[0051] It should be noted that, in this application, the fault propagation path refers to the equipment path through which the fault impact extends in the distribution network according to the equipment connection relationship; the equipment node sequence refers to the set of distribution equipment nodes arranged in the order of fault propagation; the switch action sequence refers to the action record sequence of circuit breakers, sectionalizing switches, and tie switches related to the fault propagation path; the network topology connection relationship refers to the electrical connection relationship and switch on / off relationship between equipment nodes; the path description information refers to the field information used to describe the equipment, switch actions, and their topological associations involved in fault propagation; the voltage limit parameter refers to the upper and lower voltage limits that the equipment nodes are allowed to operate at; and the branch tide... The current limit parameter represents the maximum active power, reactive power, or current allowed to be carried by the distribution branch; the equipment operation safety margin parameter represents the equipment carrying capacity reserve reserved for operational fluctuations during switch operation and power restoration; the safety constraint parameter set represents the set of voltage limit parameters, branch power flow limit parameters, and equipment operation safety margin parameters organized according to the equipment node sequence; the real-time operation status of the power grid represents the voltage of each equipment node, the power flow of each branch, the switch on / off status, and the power output status before the action is executed; the safety boundary represents the operating range that the equipment node voltage, branch power flow, and equipment carrying capacity should meet after the isolation layer agent and the recovery layer agent execute the corresponding actions.

[0052] In specific implementation, firstly, the fault propagation path is read according to the timestamps and causal relationships of each fault event in the causal reasoning evidence chain. The faulty equipment, protection devices, circuit breakers, sectionalizing switches, and tie switches in the path are mapped to corresponding nodes in the distribution network topology using equipment asset codes. The equipment nodes are then arranged according to the order of fault event occurrence, and the arrangement result is used as the equipment node sequence. Simultaneously, the action type, action time, and action result of each switch are read, and the resulting records are used as the switch action sequence. Next, the branch connection relationships between adjacent equipment nodes and the switch on / off status are read from the distribution network topology. The equipment node sequence, switch action sequence, and network topology connection relationships are associated according to a unified equipment asset code, and the resulting field-based association result is used as path description information. Secondly, based on the equipment asset codes in the path description information, the rated voltage range and allowable operating voltage range of each equipment node are read from the equipment ledger and operating parameter configuration, and the resulting parameters are used as voltage limit parameters. The allowable current and allowable transmission power of the line and transformer are read from the branch operating parameters, and the resulting parameters are used as... Branch power flow limit parameters are calculated, and preset operating margins are read according to equipment type, current load level, and corresponding switch action type. The obtained margins are used as equipment operation safety margin parameters. Then, an index relationship between each equipment node and its corresponding parameters is established according to the equipment node sequence, and the indexed parameter set is used as the safety constraint parameter set. Finally, the real-time operating status of the power grid is read, and the current distribution network operation section is established according to the network topology connection relationship in the path description information. The fault section switch tripping action corresponding to the isolation layer intelligent agent and the backup power supply input and network reconfiguration action corresponding to the recovery layer intelligent agent are sequentially written into the operation section. The node voltage and branch power flow after the action execution are obtained through distribution network power flow calculation. The node voltage is compared with the corresponding voltage limit parameter range, and the branch power flow is compared with the corresponding branch power flow limit parameter capacity. The allowable action range is reduced according to the equipment operation safety margin parameters. The operating interval that simultaneously meets the requirements of node voltage range, branch power flow range, and equipment operation margin is used as the safety boundary after the action execution of the isolation layer intelligent agent and the recovery layer intelligent agent.

[0053] The instruction generation module 500 is used to generate fault self-healing control instructions for the power distribution network based on all logical constraints and all security constraints, and then issue them to the power distribution equipment.

[0054] In this embodiment, generating fault self-healing control commands for the power distribution network based on all logical constraints and all security constraints, and issuing them to the power distribution equipment, can be achieved through the following steps: All logical constraints are classified according to the fault section isolation task, load transfer task, and safety verification task. The classified logical constraints are then mapped to the corresponding control objectives according to the operation sequence of each type of task, generating an initial control scheme sequence. Based on all security constraints, each candidate control scheme in the initial control scheme sequence is subjected to security verification, and a security-verified control instruction sequence is generated. The action sequence and execution time window of each control instruction in the control instruction sequence that has been verified for safety are sorted and conflict resolution is performed to generate fault self-healing control instructions for the distribution network. The fault self-healing control command is then sent to the power distribution equipment.

[0055] It should be noted that, in this application, the fault section isolation task refers to the control task of isolating the fault section from the normal power supply network; the load transfer task refers to the control task of transferring the affected load to the available power supply path after the fault section is isolated; the safety verification task refers to the task of verifying whether the candidate control scheme meets the requirements of voltage, power flow, switch status, and power supply path; the control objective refers to the equipment control result corresponding to each type of task; the initial control scheme sequence refers to the arrangement result of candidate control schemes obtained by mapping all logical constraints according to task category and operation sequence; the candidate control scheme refers to the single control scheme waiting for safety verification in the initial control scheme sequence; the safety verification refers to the process of verifying the executability of the candidate control scheme according to all safety constraints; and the safety conditions refer to the equipment voltage and branch conditions under the candidate control scheme. The execution conditions that the power flow, switch status, and power supply path should meet; the scheme parameters represent the parameters corresponding to the switch action object, action type, execution sequence, execution time window, and transfer path in the candidate control scheme; the control instruction sequence after safety verification represents the arrangement of control instructions after eliminating invalid candidate control schemes and correcting the scheme parameters; the action sequence represents the execution order relationship between each control instruction; the execution time window represents the time range within which each control instruction is allowed to be issued and executed; conflict resolution represents the process of adjusting the conflict between action sequence, execution time window, and mutually exclusive actions; the fault self-healing control instruction represents the fault isolation, load transfer, and safety verification related control instructions issued to the power distribution equipment; the power distribution equipment represents the switches, lines, transformers, buses, and distributed energy access equipment that receive and execute the fault self-healing control instructions.

[0056] In specific implementation, firstly, the fault section location information, switch action sequence information, load transfer path information, action permission boundaries, and cooperation timing recorded in all logical constraints are read, and then classified into fault section isolation tasks, load transfer tasks, and safety verification tasks according to the task fields. For fault section isolation tasks, the fault section boundary switch opening, fault section blocking, and isolation result verification are written into the isolation control target. For load transfer tasks, the tie switch closing, backup power supply activation, and affected load restoration are written into the transfer control target. For safety verification tasks, voltage over-limit verification, branch power flow verification, topology connectivity verification, and switch status review are written into the verification control target. Then, according to the cooperation... The control objectives are arranged in the sequence of "first isolate the faulty section, then perform load transfer, and finally perform safety verification" in the timing table. The set of control objectives and the corresponding action permission boundaries are used as the initial control scheme sequence. Next, the candidate control schemes in the initial control scheme sequence are read one by one, and all prohibited action sets, mandatory action sets, action priorities, action time windows, mutual exclusion operation rules, and safety boundaries in the safety constraints are called for safety verification. For switching actions, load transfer actions, and equipment adjustment actions in the candidate control schemes that fall into the prohibited action set, they are directly removed from the initial control scheme sequence. For candidate control schemes that lack the necessary actions in the mandatory action set, the following steps are taken: The corresponding boundary switch tripping, interlocking verification, branch power flow verification, or transfer path verification actions are added. For candidate control schemes with mismatched action time windows, the execution time is adjusted to the time range after the protection action is completed and the equipment status meets the operability conditions. For candidate control schemes that are close to the safety boundary but have not yet crossed the boundary, the transfer load ratio is reduced or the closing sequence of the tie switch is adjusted. The corrected scheme parameters are written back to the corresponding candidate control scheme, and the retained candidate control schemes are used as the control instruction sequence that has passed safety verification. Then, the action object, action type, action priority, action sequence, and execution time window of each control instruction in the control instruction sequence that has passed safety verification are read, and the action priority is first determined. The control commands corresponding to fault section isolation, load transfer, and safety verification are arranged in a hierarchical manner, and then sorted according to the action sequence. For control commands that simultaneously open and close the same switch within the same execution time window, the control command that satisfies the fault section isolation task is retained and the other control command is postponed. For control commands that have opposite power flow regulation on the same transfer path, the control command that satisfies the safety boundary is retained and conflicting control commands are deleted. For cases where multiple control commands share the same execution time window and there are mutually exclusive operation rules, the execution time window is reallocated according to the action priority, and the set of control commands after sorting and conflict resolution is used as the fault self-healing control command of the distribution network.Finally, the fault self-healing control commands for the distribution network are encapsulated into sendable messages according to the communication protocols corresponding to the distribution equipment. The messages include the equipment asset code, controlled object, action type, execution time window, verification identifier, and acknowledgment address. For switchgear distribution equipment, opening, closing, or blocking commands are sent to the corresponding switch control terminal. For transformer and line-related distribution equipment, branch power flow verification and voltage over-limit verification commands are sent to the corresponding monitoring terminal. For distributed energy access equipment, input, output withdrawal, or output adjustment commands are sent to the corresponding control terminal. After sending, the execution acknowledgments and status changes returned by the distribution equipment are read. Successful acknowledgments and status changes consistent with the fault self-healing control commands are used as the basis for the distribution equipment to complete command execution.

[0057] Therefore, this application demonstrates that fault self-healing control commands for the distribution network can be generated and issued to the distribution equipment through all logical and security constraints. Firstly, by using a feature fusion module to collect multi-source heterogeneous data from the distribution network in real time and perform feature fusion, electrical quantities, alarm semantics, and operating status can be unified into a composite feature vector, providing a more complete data foundation for fault identification under complex operating conditions and reducing misjudgments and omissions caused by judging a single electrical quantity. Secondly, by using alarm prompts to guide a pre-trained power big data model to perform logical reasoning on the composite feature vector and output a natural language reasoning chain, a clear reasoning relationship can be established between alarm phenomena, equipment status, and operating semantics, making the fault diagnosis process readable, verifiable, and traceable, reducing the impact of fragmented alarm information on the handling process. Furthermore, by constructing… By constructing a fault knowledge graph and performing fault root cause identification and causal reasoning, fault phenomena, protection actions, equipment status, and environmental factors can be linked to root cause determination conclusions and causal reasoning evidence chains. Furthermore, by decoding the natural language reasoning chain into logical constraints for each agent's action space, and combining the root cause determination conclusions and causal reasoning evidence chains to determine the safety constraints of action strategies, actions such as switch opening and closing, load regulation, energy storage access, and distributed power control can be coordinated in an orderly manner under unified logic and safety boundaries, reducing operational conflicts caused by fragmented execution by multiple control subjects. Finally, by generating fault self-healing control commands through logical and safety constraints and issuing them to power distribution equipment, diagnostic basis, root cause conclusions, and action constraints can be transformed into executable control commands, enabling the power distribution network to complete fault isolation and power restoration in more evidence-based and orderly fault scenarios.

[0058] In summary, the technical solution adopted in this application can improve the continuous diagnosis and collaborative recovery capabilities of power distribution network fault self-healing under complex operating conditions.

[0059] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0061] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation, characterized in that, include: The feature fusion module is used to collect multi-source heterogeneous data from the power distribution network in real time and perform feature fusion to generate a composite feature vector that integrates electrical quantities and semantic information. The reasoning and diagnosis module is used to guide the pre-trained power big data model to perform logical reasoning on the composite feature vector through alarm prompts and output a natural language reasoning chain. The root cause tracing module is used to construct a fault knowledge graph and perform fault root cause identification and causal reasoning based on the fault knowledge graph to obtain the root cause determination conclusion and causal reasoning evidence chain of the distribution network fault. The agent collaboration module is used to decode the natural language inference chain into logical constraints of the action space of the isolation layer agent and the recovery layer agent, and to determine the security constraints of the action strategies of the isolation layer agent and the recovery layer agent based on the root cause determination conclusion and the causal inference evidence chain. The instruction generation module is used to generate fault self-healing control instructions for the power distribution network based on all logical constraints and all security constraints, and then issue them to the power distribution equipment.

2. The power distribution network fault self-healing system based on large model knowledge reasoning and multi-agent collaboration of claim 1, wherein, Real-time acquisition of multi-source heterogeneous data from the power distribution network and feature fusion to generate a composite feature vector that integrates electrical quantities and semantic information specifically includes: The collected multi-source heterogeneous data is timestamped and missing values ​​are imputed to generate a time-synchronized structured data sequence. The time-domain statistical features of current, voltage, and power are extracted from the structured data sequence to form a subset of electrical quantity features; Based on the electrical quantity feature subset, semantic features are encoded for non-electrical quantity text data in the structured data sequence. The encoded semantic features are then fused with the electrical quantity feature subset to generate a composite feature vector that integrates electrical quantity and semantic information.

3. The self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation as described in claim 1, characterized in that, The pre-trained power system model is guided by alarm prompts to perform logical reasoning on the composite feature vectors, and the output natural language reasoning chain specifically includes: Alarm prompts are generated based on alarm signals and operating information captured when a fault occurs; Retrieve matching fault modes from the typical case library, embed the fault modes into the alarm prompt words, and generate a sequence of guiding alarm prompt words; The sequence of guiding alarm prompts and the composite feature vector are input into a pre-trained power big data model. Under the constraint of power grid topology connectivity, word-by-word probability distribution decoding is performed to generate a preliminary diagnostic reasoning chain that includes fault type and suspected section distribution. The multi-head self-attention mechanism built into the power model is used to perform a global analysis of the timing logic of the switching sequence actions and the causal relationship of electrical quantity mutations in the preliminary diagnostic inference chain, thereby generating a natural language inference chain.

4. The self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation as described in claim 1, characterized in that, Constructing a fault knowledge graph specifically includes: Based on a pre-built power ontology model, fault type entities, relay protection entities, and related power distribution equipment entities are identified and their attributes are filled in through equipment asset coding, establishing an initial association between fault events and equipment nodes. Relationships are extracted from the entity nodes in the initial association relationship to construct electrical connection relationships and fault propagation logic edges between devices. The protection action sequences and handling schemes of historical fault cases are injected into the corresponding entity nodes to generate a fault knowledge graph.

5. The self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation as described in claim 1, characterized in that, Based on the fault knowledge graph, fault root cause identification and causal reasoning are performed to obtain the root cause determination conclusion and causal reasoning evidence chain for distribution network faults, specifically including: The evolution probability between nodes in the fault knowledge graph is predicted to generate an association matrix that reflects the direction of fault propagation. Cross-domain aggregation is performed on the node features in the correlation matrix to obtain multi-dimensional fusion features that include equipment status, protection actions, and environmental factors; The fault correlation degree, propagation potential, spatiotemporal consistency and protection matching degree parameters of each node are extracted from the multidimensional fusion features to form an inference feature vector; The inference feature vector is input into the root cause discrimination function, and the root cause type is nonlinearly mapped and classified in combination with the environmental correction term. The root cause determination conclusion is output, and a causal inference evidence chain is generated based on the fault propagation path in the correlation matrix.

6. The self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation as described in claim 5, characterized in that, Predicting the evolution probabilities between nodes in the fault knowledge graph to generate an association matrix reflecting the direction of fault propagation specifically includes: The probability of fault propagation is calculated based on the operating status characteristics, protection action timing, and topological connection relationship of each node in the fault knowledge graph, so as to obtain the evolution probability between each node in the fault knowledge graph. The fault state transition matrix is ​​determined based on the evolution probabilities between nodes in the fault knowledge graph. Topological physical structure constraints are applied to the fault state transition matrix to generate an correlation matrix that reflects the direction of fault propagation.

7. The self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation as described in claim 1, characterized in that, The logical constraints for decoding the natural language inference chain into the action spaces of the isolation layer agent and the recovery layer agent specifically include: The natural language inference chain is structured and parsed to extract fault segment location information, switch action sequence information and load transfer path information, and action constraint description information is generated. The action permission boundaries and collaboration sequence of the isolation layer agent and the recovery layer agent are determined based on the action constraint description information. The action permission boundaries and collaboration timing of the isolation layer agent and the recovery layer agent are integrated into the logical constraints of the action space of the isolation layer agent and the recovery layer agent.

8. The self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation as described in claim 1, characterized in that, The security constraints for determining the action strategies of the isolation layer agent and the recovery layer agent based on the root cause determination conclusion and the causal reasoning evidence chain specifically include: Based on the fault type and fault segment attributes in the root cause determination conclusion, determine the set of prohibited actions and the set of mandatory actions for the isolation layer agent and the recovery layer agent. Based on the protective action sequence and device status information in the causal reasoning evidence chain, determine the action priority, action time window and mutual exclusion operation rules of the isolation layer agent and the recovery layer agent; Determine the safety boundaries of the isolation layer agent and the recovery layer agent after their actions are executed based on the fault propagation path in the causal reasoning evidence chain. The prohibited action set, the required action set, the action priority, the action time window, the mutual exclusion operation rules, and the security boundary are integrated into the security constraints of the action policies of the isolation layer agent and the recovery layer agent.

9. The self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation as described in claim 8, characterized in that, Determining the safety boundaries of the isolation layer agent and the recovery layer agent after their actions are executed based on the fault propagation path in the causal reasoning evidence chain specifically includes: The fault propagation path in the causal reasoning evidence chain is analyzed, and the device node sequence, switch action sequence and network topology connection relationship involved in the fault propagation path are extracted to generate path description information; Extract the voltage limit parameters, branch power flow limit parameters, and equipment operation safety margin parameters corresponding to each node in the device node sequence from the causal reasoning evidence chain, and generate a set of safety constraint parameters; Based on the network topology connections in the path description information and the set of security constraint parameters, and in conjunction with the real-time operating status of the power grid, the security boundaries after the actions of the isolation layer agent and the recovery layer agent are determined.

10. The self-healing system for distribution network faults based on large-model knowledge reasoning and multi-agent collaborative operation as described in claim 1, characterized in that, The fault self-healing control command for the power distribution network is generated based on all logical constraints and all security constraints and then issued to the power distribution equipment. Specifically, this includes: All logical constraints are classified according to the fault section isolation task, load transfer task, and safety verification task. The classified logical constraints are then mapped to the corresponding control objectives according to the operation sequence of each type of task, generating an initial control scheme sequence. Based on all security constraints, each candidate control scheme in the initial control scheme sequence is subjected to security verification, and a security-verified control instruction sequence is generated. The action sequence and execution time window of each control instruction in the control instruction sequence that has been verified for safety are sorted and conflict resolution is performed to generate fault self-healing control instructions for the distribution network. The fault self-healing control command is then sent to the power distribution equipment.

Images