A power grid fault processing knowledge graph completion method based on multi-task learning

Abstract

The application relates to a power grid fault processing knowledge graph completion method based on multi-task learning, which comprises the following steps: step 1, inputting a power grid regulation history data set; step 2, constructing a power grid fault processing knowledge graph; step 3, obtaining initialization Embedding of a current device node state and a fault treatment action; step 4, taking the triple obtained by performing Embedding initialization in step 3 as the input of two downstream tasks, and obtaining the calculation loss functions of the two downstream tasks; step 5, defining a learnable weight parameter alpha to dynamically superimpose the calculation loss functions of the two downstream tasks, and then used for updating a node and edge self-encoding model; step 6, generating a power grid regulation cognitive graph with more accurate expression; and step 7, realizing power grid fault processing knowledge graph completion. The application can better predict the potential association edges between nodes, and then complete the knowledge graph of power grid regulation.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph completion technology for guiding power grid regulation and control, and relates to a method for completing a knowledge graph for power grid fault handling, particularly a method for completing a knowledge graph for power grid fault handling based on multi-task learning. Background Technology

[0002] With the continuous expansion of the power grid and its increasingly flexible operation, power grid control operations are becoming more complex, and the workload of control personnel is also increasing. This places higher demands on the automation and intelligence of control operations, and power grid dispatch relies on the guidance of power grid control knowledge graphs. Existing power grid control guidance models based on reinforcement learning combined with knowledge graphs can, to some extent, suggest control actions based on the status of power equipment and generate interpretable instruction sets by sequentially representing a series of operations. However, these are all based on the assumption that the knowledge graph itself is relatively accurate. However, the edge connections in the knowledge graph are not complete; there are some unreasonable edges and some potential edges to be explored. Therefore, intelligent completion of the power grid control knowledge graph is a very necessary task.

[0003] A search revealed no publicly available literature of the same or similar prior art as this invention. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and propose a knowledge graph completion method for power grid fault handling based on multi-task learning, which can better predict the potential correlation edges between nodes and thus complete the knowledge graph of power grid regulation.

[0005] The present invention solves its practical problem by adopting the following technical solution:

[0006] A knowledge graph completion method for power grid fault handling based on multi-task learning includes the following steps:

[0007] Step 1: Input the historical dataset of power grid control;

[0008] Step 2: Construct a knowledge graph for power grid fault handling and define expert meta-paths for power grid regulation;

[0009] Step 3: Select the triplet of power grid equipment node status and fault handling from Step 2 to initialize the embedding of the equipment node and the handling action respectively, and finally obtain the initial embedding of the current equipment node status and fault handling action, as well as the autoencoder model used to initialize the node and action embedding.

[0010] Step 4: Use the triples obtained in Step 3, which have been initialized with embedding, as input to the two downstream tasks, and obtain the loss functions for the two downstream tasks.

[0011] Step 5: Based on the loss functions of the two downstream tasks obtained in Step 4, define a learnable weight parameter α to dynamically superimpose the loss functions of the two downstream tasks, and then use it to update the autoencoder model of nodes and edges.

[0012] Step 6: Obtain the autoencoder model of the node and action embedding obtained in Step 5, and generate a more accurate cognitive map of power grid regulation;

[0013] Step 7: Based on the new power grid regulation cognitive graph obtained in Step 6, the power grid fault handling knowledge graph is finally completed.

[0014] Moreover, the power grid control historical dataset in step 1 includes equipment node information and fault handling action set in the power grid;

[0015] Furthermore, the specific method for step 2 is as follows:

[0016] By combining prior knowledge of fault handling with the entity status of power grid equipment and corresponding execution actions in the power grid dataset, a corresponding knowledge graph structure is constructed. At the same time, expert meta-paths for power grid regulation are defined by relevant experts in power grid regulation.

[0017] Furthermore, the specific method for step 3 is as follows:

[0018] Using the knowledge graph constructed in step 2 and the relationship between the state transitions of various equipment entities in the power grid dataset, the equipment nodes and fault handling actions in the power grid dataset in step 1 are extracted in a format based on (node, action, node) triples. The node equipment state is input into the node autoencoder network for embedding initialization, and the fault handling action is input into the relation autoencoder network for embedding initialization. Finally, the initial embedding of the current equipment node state and fault handling action, as well as the autoencoder model used to initialize the node and action embeddings are obtained.

[0019] Furthermore, the specific method for step 4 is as follows:

[0020] Select the (node, action, node) triplet that has already undergone initial embedding in step 3; then define two downstream tasks. Task 1 is to use Graph Embedding to calculate the loss of the (node, action, node) triplet, denoted as loss1. Task 2 is to transform the edge representation into weights representing the degree of association between adjacent nodes based on the cognitive knowledge graph constructed from the historical power grid control dataset. Then, based on the cognitive knowledge graph with association weights, the Dijkstra method is used to obtain the set of all shortest instance paths in the historical power grid control dataset. Next, the instances are abstracted into entities, and then the set of all shortest instance paths is abstracted into empirical shortest meta-paths based on entities. Finally, the empirical shortest meta-path set obtained from the historical power grid control dataset is used to calculate the adversarial loss with the expert meta-path defined in step 2, denoted as loss2.

[0021] Furthermore, the specific method for step 5 is as follows:

[0022] Obtain the losses loss1 and loss2 based on the two tasks in step 4, and then define a learnable weight parameter α. The total loss function loss is composed of dynamic parameters and the two losses, which are used to initialize the autoencoder model update of node and action embedding.

[0023] Furthermore, the specific method for step 6 is as follows:

[0024] Obtain the autoencode model of node and action embedding obtained in step 5. Based on the node autoencode model and the action autoencode model, update the representation of nodes and edges in the power grid control knowledge graph to generate a more accurate power grid control cognition graph.

[0025] Furthermore, the specific method for step 7 is as follows:

[0026] Based on the new power grid control knowledge graph obtained in step 6, the association prediction is performed on two non-adjacent nodes according to the representation of nodes and edges. Based on the prediction results with strong association, the edges of the corresponding nodes are completed, and finally the knowledge graph of power grid fault handling is completed.

[0027] Advantages and beneficial effects of the present invention:

[0028] 1. This invention proposes a knowledge graph completion method for power grid fault handling based on multi-task learning. It employs multi-task constraints to optimize the representation of entities and relationships in the knowledge graph and to more accurately complete the edge completion of potential relationships between entities, thereby updating the power grid fault handling knowledge graph. This invention can be applied to the upstream task results of multiple objective tasks, enabling the use of multi-objective task constraints to update the parameters of the upstream representation learning task. This results in higher accuracy and robustness of the representation learning model, leading to more accurate embedding representations of nodes and actions, better prediction of potential association edges between nodes, and ultimately, completion of the power grid regulation knowledge graph.

[0029] 2. This invention uses ReLU to convert the representation of edges between nodes into weight values ​​that can represent the strength of the association between nodes. This addresses the problem that embedding edges between nodes cannot intuitively express the degree of association between two nodes. For example... Figure 1 As shown, in this invention, the sum of the current node's connected edges and the Euclidean distance difference between the current node and its corresponding neighboring nodes are processed using the ReLU function. This transforms the edge embedding representation into an intuitive numerical value with weighted properties, replacing the edge embedding in the cognitive graph with numerical values. This generates a cognitive knowledge graph that can intuitively display the strength of the connections between nodes. This effectively lays the groundwork for subsequently generating the shortest instance state transition matrix using Dijkstra's algorithm.

[0030] 3. The present invention as follows Figure 2 As shown, by using multi-task constraints, namely the dual-task adversarial constraint function of Graph Embedding and meta-path, the parameters of the autoencoder network module of edges and nodes are updated, thereby improving the accuracy and robustness of the autoencoder model, which is helpful for the later correction of node and edge embedding in cognitive knowledge graphs.

[0031] 4. This invention proposes a comprehensive framework for completing the knowledge graph of power grid fault handling based on multi-task learning. This framework first uses multi-task learning to effectively update the parameters of the representation learning model. Then, it uses the representation learning model to re-encode historical data in power grid control. By utilizing the relationship between node embeddings and edge embeddings, it predicts and adds potential edges to the knowledge graph, thereby completing the knowledge graph of power grid fault handling. This invention trains the model parameters based on multi-task learning. On the one hand, this increases loss constraints, making it easier for the model to converge in an effective direction; on the other hand, by applying loss constraints specific to specific scenarios while updating the model parameters, the trained model is more applicable to those scenarios. In this context, the resulting autoencoder model can be better suited for completing the knowledge graph of power grid fault handling.

[0032] 5. The difference between this invention and previous power grid control methods lies in the use of a power grid fault handling knowledge graph completion based on multi-task learning. This invention focuses on completing the power grid control cognitive knowledge graph to guide fault handling actions, rather than learning downstream decision-making strategies. This improves the quality of the cognitive knowledge graph to better guide the selection of fault handling actions.

[0033] 6. The network used to transform edge embeddings into edge weights in this invention differs from previous methods. This invention utilizes a ReLU network to numerically weight the Euclidean distance between (node-edge) and adjacent nodes, thereby transforming the cognitive knowledge graph containing edge embeddings into a cognitive knowledge graph containing edge weights. This invention considers that in power grid fault handling tasks, starting from the initial power grid state, multiple fault handling action sequences can be used to reach the target power grid state. How to select the optimal fault handling action sequence from these sequences—based on the sequence length or other factors? Therefore, this paper proposes transforming edge embeddings into edge weight values. These edge weight values ​​can then reflect the value of each action within the entire fault handling action sequence, making it easier to select the optimal fault handling action sequence. Attached Figure Description

[0034] Figure 1 This is a flowchart of the ReLU function-based relation edge to weight conversion module of the present invention;

[0035] Figure 2 This is a flowchart of the parameter update process of the Embedding module based on multi-task constraints of the present invention;

[0036] Figure 3 This is a schematic diagram of the overall framework for knowledge graph completion of power grid fault handling based on multi-task learning in this invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the accompanying drawings:

[0038] A knowledge graph completion method for power grid fault handling based on multi-task learning, such as... Figures 1 to 3 As shown, it includes the following steps:

[0039] Step 1: Input the historical dataset of power grid control;

[0040] The power grid control historical dataset in step 1 includes equipment node information and fault handling action set in the power grid.

[0041] Step 2: Construct a knowledge graph for power grid fault handling and define expert meta-paths for power grid regulation;

[0042] The specific method for step 2 is as follows:

[0043] By combining prior knowledge of fault handling with the entity status of power grid equipment and corresponding execution actions in the power grid dataset, a corresponding knowledge graph structure is constructed. At the same time, expert meta-paths for power grid regulation are defined by relevant experts in power grid regulation.

[0044] In this embodiment, the specific method for constructing a knowledge graph of inter-entity relationships based on the entity states of power grid equipment and corresponding fault handling actions contained in the power grid dataset in step 2 is as follows:

[0045] (1) Obtain the fault handling records of each power grid equipment node in its initial state;

[0046] (2) The state of each power grid equipment node is regarded as an entity node in the knowledge graph, and the fault handling actions taken for the state of each power grid equipment node are regarded as the association edges between entity nodes.

[0047] (3) The status of power grid equipment nodes in the entire power grid dataset is associated with the edges corresponding to the fault handling actions, ultimately forming a knowledge graph containing the status of power grid equipment nodes and fault handling in the dataset.

[0048] Step 3: Select the triplet of power grid equipment node status and fault handling from Step 2 to initialize the embedding of the equipment node and the handling action respectively, and finally obtain the initial embedding of the current equipment node status and fault handling action, as well as the autoencoder model used to initialize the node and action embedding.

[0049] The specific method for step 3 is as follows:

[0050] Using the knowledge graph constructed in step 2 and the relationship between the state transitions of various equipment entities in the power grid dataset, the equipment nodes and fault handling actions in the power grid dataset in step 1 are extracted in a format based on (node, action, node) triples. The node equipment state is input into the node autoencoder network for embedding initialization, and the fault handling action is input into the relation autoencoder network for embedding initialization. Finally, the initial embedding of the current equipment node state and fault handling action, as well as the autoencoder model used to initialize the node and action embeddings are obtained.

[0051] In this embodiment, step 3 specifically includes the following steps:

[0052] (1) Based on the state of the power grid equipment node, define the entity class corresponding to each state of the power grid equipment node, and define the number of entity classes as n; at the same time, define the dimension size of each state input in reinforcement learning as embed_size.

[0053] (2) The entity class is initialized for representation learning based on the number m of the corresponding power grid equipment node states contained in each entity class. The dimension of the initialization vector is m*embed_size.

[0054] (3) Define the node autoencoder model. Based on the device node information in the power grid dataset, initialize the node autoencoder model. The dimension of the initialization vector is embed_size.

[0055] (4) Define an action self-encoding model and set the dimension of the initial vector of the fault handling action to 1*embed_size;

[0056] (5) Based on the fault handling dataset under the relevant state, obtain the corresponding record, which contains instance records corresponding to n entity classes, as an n-tuple. Based on the n-tuple, generate triples (node ​​i, fault handling action r, node j) with corresponding relationships. The number of such triples is denoted as k. Save these k triples that have been initialized with embedding as input for downstream multi-tasks. Finally, obtain the initial embedding of the current device node state and fault handling action, as well as the autoencoder model used to initialize the node and action embeddings.

[0057] Step 4: Use the triples obtained in Step 3, which have been initialized with embedding, as input to the two downstream tasks, and obtain the loss functions for the two downstream tasks.

[0058] The specific method for step 4 is as follows:

[0059] Select the (node, action, node) triplet that has already undergone initial embedding in step 3; then define two downstream tasks. Task 1 is to use Graph Embedding to calculate the loss of the (node, action, node) triplet, denoted as loss1. Task 2 is to transform the edge representation into weights representing the degree of association between adjacent nodes based on the cognitive knowledge graph constructed from the historical power grid control dataset. Then, based on the cognitive knowledge graph with association weights, the Dijkstra method is used to obtain the set of all shortest instance paths in the historical power grid control dataset. Next, the instances are abstracted into entities, and then the set of all shortest instance paths is abstracted into empirical shortest meta-paths based on entities. Finally, the empirical shortest meta-path set obtained from the historical power grid control dataset is used to calculate the adversarial loss with the expert meta-path defined in step 2, denoted as loss2.

[0060] In this embodiment, in step 4, the (node ​​i, fault handling action r, node j) triple obtained in step 3 after embedding initialization is used as the input to the two downstream tasks, and the computational loss of the two downstream tasks is obtained. The specific steps include:

[0061] (1) First, define downstream task 1, which calculates the representation loss of triples in the knowledge graph based on TransH in the Graph Embedding model.

[0062] (2) Task 1 inputs the triplet (node ​​i, fault handling action r, node j) obtained in step 3, which has been initialized with embedding, into the TransH model. The output is the loss of the embedding between the node and the edge and the representation distance between the node and the adjacent node calculated based on TransH, denoted as loss1.

[0063] (3) Define downstream task 2, which is used to calculate the loss between the meta-path obtained from empirical data and the meta-path defined by experts.

[0064] (4) The input to Task 2 is the triplet (node ​​i, fault handling action r, node j) obtained in Step 3 that has already undergone embedding initialization. First, the edge embedding is defined as an edge weight reset function as follows:

[0065] w ij =Relu(e i +r i -e j )

[0066] Where e i It is the head node embedding in the triple, e j It is the tail node embedding representation in the triple, r i It is the action embedding representation in the triple. By calculating the Euclidean distance, the distance result is fed into ReLU to transform the edge embedding into non-negative weights.

[0067] (5) Replace the original edge embedding with the obtained edge weights and update the original cognitive knowledge graph into a cognitive knowledge graph with weights.

[0068] (6) Using the Dijkstra algorithm, based on a weighted cognitive knowledge graph, the shortest path from each node to the target node is obtained.

[0069] (7) Abstract instances of similar categories into the same entity, and then transform all the shortest instance paths in the historical dataset of power grid regulation into the shortest empirical element paths.

[0070] (8) Calculate the adversarial loss based on the shortest empirical meta-path and the expert-defined meta-path, denoted as loss2.

[0071] Step 5: Based on the loss functions of the two downstream tasks obtained in Step 4, define a learnable weight parameter α to dynamically superimpose the loss functions of the two downstream tasks, and then use it to update the autoencoder model of nodes and edges.

[0072] The specific method for step 5 is as follows:

[0073] Obtain the losses loss1 and loss2 based on the two tasks in step 4, and then define a learnable weight parameter α. The total loss function loss is composed of dynamic parameters and the two losses, which are used to initialize the autoencoder model update of node and action embedding.

[0074] In this embodiment, step 5 specifically includes the following steps:

[0075] (1) Define a learnable weight parameter α.

[0076] (2) Using the learnable weight parameter α and the two task loss functions loss1 and loss2 obtained, the total loss function loss is defined as follows:

[0077] loss=α*loss1+(1-α)*loss2loss=α*loss1+(1-α)*loss2

[0078] (3) Update the autoencoder model of nodes and edges using the total loss function loss.

[0079] Step 6: Obtain the autoencode model of node and action embedding obtained in Step 5. Based on the node autoencode model and action autoencode model, update the representation of nodes and edges in the power grid control knowledge graph to generate a more accurate power grid control cognition graph.

[0080] In this embodiment, in step 6, the power equipment nodes and fault handling actions in the power grid regulation historical dataset are re-encoded using the autoencoder model of nodes and edges obtained from step 5. Specifically, the method may be as follows:

[0081] (1) Obtain the power equipment nodes and fault handling actions from the historical data set of power grid regulation;

[0082] (2) Using nodes and actions as inputs to the node autoencoder model and the action autoencoder model respectively, we obtain the embedding representations of device nodes and fault handling actions;

[0083] Step 7: Based on the new power grid control knowledge graph obtained in Step 6, according to the representation of nodes and edges, perform association prediction on two non-adjacent nodes. Based on the prediction results with strong association, complete the edges of the corresponding nodes, and finally realize the completion of the power grid fault handling knowledge graph.

[0084] In step 7, based on the embedding representations of device nodes and fault handling actions obtained in step 6, and combined with the knowledge graph, the potential edges to be mined between nodes in the knowledge graph are completed. Specific methods can be divided into:

[0085] (1) Obtain the embedding of each node and all possible processing actions contained in that node;

[0086] (2) Based on the representation of a node and all the action spaces corresponding to that node, the target node that is highly matched in the cognitive knowledge graph is retrieved by calculating the sum of the node and the action. If it exists, edge completion is performed.

[0087] In the task of power grid fault control, most existing algorithms consider using model strategies to guide fault handling decisions, but rarely optimize these strategies using knowledge graphs. The innovation of this invention lies primarily in the completion of the power grid fault handling knowledge graph based on multi-task learning. On one hand, it utilizes Graph Embedding to calculate the loss between node pairs and action representations; on the other hand, it uses empirical meta-paths and expert meta-paths obtained from historical power grid control data to calculate adversarial losses. Then, a learnable parameter α is used to dynamically superimpose the two losses, jointly updating the autoencoder models of nodes and actions, thereby improving the accuracy of node and edge representations in the power grid control cognitive knowledge graph. Subsequently, based on known nodes and edges, it predicts whether there are edge connections between non-adjacent nodes, achieving the completion of the cognitive knowledge graph to better guide fault handling action recommendations and obtain more reasonable handling results.

[0088] The method in this invention is designed in two ways: firstly, it utilizes ReLU to convert the representation of edges between nodes into a weight module; secondly, it uses multi-task learning to update the parameters of the autoencoder model used for learning node and edge representations, thereby achieving a more accurate and robust representation of nodes and edges in the knowledge graph of power grid regulation. Specifically, the module for converting edge embeddings into edge weights uses the Euclidean distance between a node and its corresponding adjacent node as input to the ReLU function to convert the edge embeddings between adjacent nodes into weight values ​​that intuitively reflect the correlation between the two nodes. For the multi-task learning module for updating representation learning parameters, we use multi-task constraints—specifically, the simultaneous use of Graph Embedding and meta-path dual-task adversarial constraint functions—to update the parameters of the autoencoder network module for edges and nodes, thereby improving the accuracy and robustness of the autoencoder model.

[0089] Based on the above improvements, the proposed multi-task learning-based knowledge graph completion method for power grid fault handling is realized. This method can effectively improve the accuracy of node and edge representations in the power grid control cognitive graph, predict and complete the potential relationship edges between nodes, and the completed power grid control cognitive knowledge graph can better guide the current node to select a more reasonable fault handling action instruction set, improve the quality of fault handling actions, and ultimately make the decision-making effect more ideal.

[0090] The working principle of this invention is:

[0091] This invention first processes the equipment nodes and corresponding fault handling actions in the historical power grid control dataset. It then constructs a knowledge graph based on the processed historical power grid control dataset using prior knowledge of fault handling, and defines expert meta-paths based on expert definitions. Next, it uses an autoencoder method to embedding the power grid equipment node states and fault handling actions, where the power grid equipment node states are represented by a node autoencoder network embedding, and the fault handling actions are represented by an action autoencoder network embedding. After initializing the representations of all nodes and action instances in the knowledge graph, these are used as inputs for two downstream tasks. The first downstream task is to learn how to represent these nodes and actions using a graph embedding learning network. The first task updates the data by constructing a loss function `loss1` based on nodes and action edges. The second downstream task, based on the initialized node and action representations, transforms the action representation into transition probabilities (weights) between two adjacent nodes. A larger weight between two nodes indicates that the current node is more likely to execute the action and reach that node. Then, a cognitive graph of historical power grid control data is constructed based on the previous node representations and weights. Subsequently, the Dijkstra algorithm is used to calculate the shortest instance path from a node in the cognitive graph to the target node. Then, the shortest entity path is abstracted based on the obtained shortest instance path dataset. This set of shortest entity paths is then used for adversarial learning against expert meta-paths defined by experts, thus constructing the loss function `loss2` for task 2. Finally, an adaptive dynamic weight is defined for the two task loss functions based on the importance of tasks 1 and 2 to the upstream task. The parameters of the node autoencoder network and action autoencoder network are updated based on the two loss constraints. Based on the representations of equipment nodes and fault handling actions in the obtained historical power grid control dataset, the potential relationships between entities in the knowledge graph are predicted and edge completion is performed, thereby better guiding the handling of power grid equipment node faults.

[0092] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A power grid fault handling knowledge graph completion method based on multi-task learning, characterized by: The method comprises the following steps: Step 1, input power grid regulation historical data set; Step 2, construct a knowledge graph for power grid fault handling and define expert meta-paths for power grid regulation; Step 3, select the device node state and the triple of fault handling in step 2 to initialize the Embedding of the device node and the handling action, and finally obtain the initialized Embedding of the current device node state and fault handling action and the auto-encoding model for initializing the node and action Embedding; Step 4, input the triple in step 3 which has been initialized with Embedding as the input of two downstream tasks, and obtain the calculation loss functions of the two downstream tasks; Step 5, based on the calculation loss functions of the two downstream tasks obtained in step 4, define a learnable weight parameter α to dynamically superimpose the calculation loss functions of the two downstream tasks, and then update the auto-encoding model of the node and edge; Step 6, obtain the auto-encoding model of the node and action Embedding in step 5 to generate a more accurate power grid regulation cognitive graph; Step 7, based on the new power grid regulation cognitive graph obtained in step 6, finally realize the completion of the power grid fault handling knowledge graph; The specific method of step 3 is: The knowledge graph constructed in step 2 and the relationship between the state transitions of each device entity in the power grid data set are used to extract the format of the device node and fault handling action in the power grid data set in step 1 based on triple, wherein the node device state is input into the node auto-encoding network for Embedding initialization, and the fault handling action is input into the relationship auto-encoding network for Embedding initialization, and finally the initialized Embedding of the current device node state and fault handling action and the auto-encoding model for initializing the node and action Embedding are obtained; The specific method of step 4 is: Select the triple in step 3 which has been initialized with Embedding; then define two downstream tasks, wherein task 1 is to calculate the loss using Graph Embedding for the triple, and the loss is recorded as loss1, and task 2 is to convert the edge representation of the cognitive knowledge graph constructed based on the power grid regulation historical data set into a weight representing the correlation degree between adjacent nodes, and then use the Dijkstra method based on the cognitive knowledge graph with correlation weight to obtain a set of all shortest instance paths in the power grid regulation historical data set, then abstract the instances into entities and further abstract all sets of shortest instance paths into experience shortest meta-paths based on entities, and finally calculate the adversarial loss between the set of experience shortest meta-paths obtained from the power grid regulation historical data set and the expert meta-paths defined in step 2, and the loss is recorded as loss2.

2. The power grid fault handling knowledge graph completion method based on multi-task learning according to claim 1, characterized in that: The power grid regulation historical data set in step 1 includes device node information in the power grid and a set of fault handling actions.

3. The method of claim 1, wherein the method is based on multi-task learning. The specific method of step 2 is: The prior knowledge of fault handling is combined with the state of power grid equipment entities and the corresponding execution actions in the power grid data set to construct a corresponding knowledge graph structure, and an expert meta-path for power grid regulation is defined by experts related to power grid regulation.

4. The method of claim 1, wherein the method is based on multi-task learning. The specific method of step 5 is: The two task-based losses loss1 and loss2 in step 4 are obtained, and then a learnable weight parameter a is defined, and the total loss function loss is composed of the dynamic parameter and the two losses, which is used to initialize the self-encoding model update of the node and action Embedding.

5. The power grid fault handling knowledge graph completion method based on multi-task learning according to claim 1, characterized in that: The specific method of step 6 is: The self-encoding model of the node and action Embedding obtained in step 5 is obtained, and the representation of the nodes and edges in the power grid regulation knowledge graph is updated based on the node self-encoding model and the action self-encoding model, respectively, to generate a more accurate power grid regulation cognitive graph.

6. The method of claim 1, wherein the method is based on multi-task learning. The specific method of step 7 is: Based on the new power grid regulation cognitive graph obtained in step 6, the correlation prediction of two non-adjacent nodes is performed according to the representation of the nodes and edges, and the edge completion of the corresponding nodes is performed based on the prediction results with strong correlation, finally realizing the completion of the power grid fault handling knowledge graph.

Images