Power grid fault processing cognitive knowledge graph generation system, method and medium

Abstract

The present invention discloses a system, method and medium for generating a cognitive knowledge graph for power grid fault processing. The present invention adopts an attention mechanism combined with a collaborative factor decomposition machine to assign matching weights to different influencing factors, wherein two different forms of attention substructures are used to collaboratively model the influencing factor combination process, and then the combined features are combined with the knowledge graph generated by the power grid regulation history record to perform graph representation learning on the nodes and relationship edges in the knowledge graph, thereby optimizing the entity and relationship representation in the knowledge graph, and then generating a more accurate cognitive knowledge graph for power grid fault processing, especially for power grid regulation tasks affected by multiple factors, which can realize multi-level interaction of multiple influencing factors and then complete feature adaptive combination, so that the generated cognitive knowledge graph for power grid fault processing is more reasonable, accurate and robust, and at the same time provides a more favorable search environment for downstream decision-making tasks.

Description

Technical Field

[0001] The present invention relates to the technical field of cognitive knowledge graphs for guiding power grid regulation, and specifically to a system, method, and medium for generating a cognitive knowledge graph for power grid fault processing. Background Art

[0002] As the scale of power grids continues to expand and their operation modes become more flexible and diverse, grid regulation and control operations are becoming increasingly complex, and the workload of regulators is also increasing. This places higher demands on the automation and intelligence of regulation and control operations, and grid dispatch relies on the grid regulation knowledge graph for guidance. Existing grid regulation guidance models based on reinforcement learning combined with knowledge graphs can, to a certain extent, both recommend regulation actions based on the status of power equipment and provide a serialized representation of a series of operations to generate an interpretable instruction set. However, these are based on the assumption that the knowledge graph itself is relatively accurate. However, the status of device nodes in the knowledge graph is more susceptible to factors such as device temperature, device sound, and device current. The combination of these influencing factors will affect the representation of the power equipment status. Therefore, studying the combination of factors that affect the status of device nodes is a very necessary task for generating a reasonable cognitive knowledge graph for power grid fault handling. Summary of the Invention

[0003] The purpose of the present invention is to provide a system, method and medium for generating a cognitive knowledge graph for power grid fault processing. The present invention adopts an attention mechanism combined with a collaborative factor decomposition machine to assign matching weights to different influencing factors, wherein two different forms of attention substructures are used to collaboratively model the influencing factor combination process, and then the combined features are combined with the knowledge graph generated by the power grid regulation history record to perform graph representation learning on the nodes and relationship edges in the knowledge graph, thereby optimizing the entity and relationship representation in the knowledge graph, and then generating a more accurate power grid fault processing cognitive knowledge graph, especially for power grid regulation tasks affected by multiple factors, which can realize multi-level interaction of multiple influencing factors and then complete feature adaptive combination, so that the generated power grid fault processing cognitive knowledge graph is more reasonable, accurate and robust, while providing a more favorable search environment for downstream decision-making tasks.

[0004] To achieve this goal, the present invention designs a cognitive knowledge graph generation system for power grid fault processing, which is characterized by comprising a knowledge graph construction module, a triple construction module, an embedding (feature embedding, a tensor for storing information) acquisition module, an embedding triple acquisition module, a parameter update module, and a cognitive knowledge graph construction module.

[0005] The knowledge graph construction module uses prior knowledge of fault handling and combines the entity status of power grid equipment and corresponding fault handling actions in the historical data set of power grid control to construct a knowledge graph;

[0006] The triple construction module uses the knowledge graph and the relationship between the state transitions of each device entity in the power grid control historical data set to extract the device nodes and fault handling actions in the power grid control historical data set in triple format. According to the factor information affecting the state of the power grid device node, each device node is assigned associated multi-factor information to obtain the device node and fault handling action triple;

[0007] The Embedding Acquisition Module uses the device node and fault handling action triples to build a conditional variational autoencoder model. This model is used to obtain the relationship between the multi-factor Embedding information associated with the current device node state and the fault handling action.

[0008] The Embedding triplet acquisition module uses the relationship between the multi-factor Embedding information associated with the current device node status and the fault handling action Embedding to build an attention mechanism combined with a collaborative factor decomposition machine model. The Embedding set of the associated factor group is input into the attention mechanism combined with the collaborative factor decomposition machine model to obtain the feature expression of the device node associated with the associated factor, and obtain the node and fault handling action Embedding triplet;

[0009] The parameter update module calculates the loss of the node and fault handling action Embedding triples, and uses the calculation results to update the attention mechanism combined with the collaborative factorization machine model;

[0010] The cognitive knowledge graph construction module uses the embeddings of associated factor groups and the relationship embeddings of fault handling actions, and uses the updated attention mechanism combined with the collaborative factor decomposition machine model to adaptively fuse the multiple factors that constitute the device node status to generate a cognitive knowledge graph for power grid fault handling.

[0011] The present invention first processes the device nodes and the corresponding actions in the fault handling action space in the power grid control historical data set, uses the prior knowledge of fault handling to construct a knowledge graph based on the processed power grid control historical data set, and then collects factor information affecting the state of the power grid device nodes, such as current, sound decibels, device temperature, etc., to construct a conditional variational autoencoder model framework and a multi-factor adaptive combination model framework based on an attention mechanism combined with a collaborative factor decomposition machine. The conditional variational autoencoder model is used to embed the power grid fault handling action and information such as device current, device sound decibels, and device temperature. The state of the power grid device node is adaptively represented by the multi-influencing factor embedding through the attention mechanism combined with the collaborative factor decomposition machine model, and the fault handling action is represented by the action autoencoder network embedding. After all nodes and action instances in the knowledge graph are initialized and represented, they are used as input for downstream tasks, wherein the downstream task is a graph embedding representation training model based on power grid device nodes and fault handling actions, and the model parameters based on the attention mechanism combined with the collaborative factor decomposition machine are updated through the computational loss of the graph embedding representation model. Based on the obtained model, the fault handling actions in the power grid control knowledge graph are more accurately represented, and the multi-factor features that constitute the device node status are more reasonably combined, ultimately generating a more accurate and robust power grid fault handling cognitive knowledge graph, thereby better guiding the handling of power grid device node faults.

[0012] Beneficial effects of the present invention:

[0013] 1. This invention uses a conditional variational autoencoder to separately encode the multiple factors that constitute the state of a device node, making fuller use of environmental information and enriching the feature representation of power grid device nodes. The conditional variational autoencoder uses factor data and corresponding label information as input to generate a latent feature space. A decoder and labels are then used to generate the output. The output and input are then combined to perform loss calculations to update model parameters, ultimately resulting in a conditional encoder. Downstream tasks can then sample corresponding features from the latent feature space based on the labels.

[0014] 2. The present invention uses the attention mechanism combined with the collaborative factor decomposition machine to adaptively fuse the multiple factors that constitute the device node state to obtain important information in the combined features, that is, using two different forms of attention substructures to collaboratively model feature combinations, which includes the low-order interaction layer of the attention sub-network and the feature extraction layer of the empowerment module, enriching the device node information and improving the effectiveness of feature acquisition, which is conducive to the more reasonable generation of cognitive knowledge graphs in the later stage.

[0015] 3. The present invention proposes a general framework for completing the knowledge graph for power grid fault processing based on multi-task learning. This framework first uses multi-tasks to effectively update the parameters of the representation learning model, then uses the representation learning model to re-encode historical data in power grid regulation, and uses the relationship between node embedding and edge embedding to predict and add some potential edges in the knowledge graph, thereby achieving the completion of the knowledge graph for power grid fault processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a structural schematic diagram of the present invention;

[0017] Figure 2 Schematic diagram of a factor encoding module based on conditional variational autoencoding of the present invention;

[0018] Figure 3 Schematic diagram of a multi-factor adaptive fusion module based on an attention mechanism combined with a collaborative factorization machine according to the present invention;

[0019] Figure 4 Schematic diagram of the overall framework for generating cognitive knowledge graphs for power grid fault processing based on multi-factor adaptive fusion combined with graph embedding.

[0020] Figure 2 The factor encoding module based on conditional variational autoencoder uses a variational autoencoder network to obtain the embedding of the power grid equipment status under the constraints of current factor, sound factor, and temperature factor.

[0021] Figure 2 In the figure, X represents the input grid equipment information, Y represents the current, temperature, and sound information, Encoder represents the encoder, and Q represents the grid equipment information combined with the current, temperature, or sound information.

[0022] Figure 2 In the equation, μ represents the mean, KL represents the KL divergence of the distribution, ε(Y,X) represents the variance, and Sampleεfrom N(0,I) represents that the distribution of ε is sampled from the standard normal distribution. The above process is reparameterization. The reparameterized result z and Y features are connected as the input information of Decoder(Q) to obtain f(z,Y). f(z,Y) represents the output information of Decoder. Finally, the input power grid equipment information X is squared with f(z,Y), that is, ||Xf(z,Y)|| 2 , and then update the variational autoencoder model parameters.

[0023] Figure 3In this paper, a multi-factor adaptive fusion module based on the attention mechanism combined with the collaborative factor decomposition machine is used to adaptively fuse the device embeddings generated under the conditional variational autoencoder network based on three grid state influencing factors (current, sound, and temperature).

[0024] Figure 3 In the figure, Input layer represents the input layer, and the input data is the grid equipment information and the corresponding grid state influencing factors. Encoder (Q) represents the encoding layer, and the obtained data is the device embedding generated under the conditional variational autoencoder network based on the three grid state influencing factors.

[0025] Figure 3 In the figure, the Weighting Module is a weighting module used to assign importance to features, and the Weighting Data is the data processed by the Weighting Module. The Attention Layer is used to perform attention processing on the encoding obtained by the Encoder (Q). The Low-level Feature Interaction Layer is used to interact with various features obtained from the lower layers. The Feature Combination Module (the rightmost layer of the third layer) is used to combine features of the Weighting Data. The Network Layer module is used to perform high-level feature interaction on the features formed after feature combination. The features are then fused with the features of the Input Layer and Low-level Feature Interaction Layer modules through the Feature with Multi-factor module to obtain the state Embedding of the power grid equipment.

[0026] Figure 4 In, e i Represents a node in the knowledge graph, i.e., a power grid equipment node; r i "[Function]" represents an edge in a knowledge graph, i.e., a fault handling action. "[TransR]" is a neural network that can be used for graph embedding and learning the embedding of nodes and edges in a knowledge graph. "[Loss]" represents the loss function used for backpropagation to update neural network parameters. "[Feature Fusion with Multi-factor Module]" represents a multi-factor adaptive fusion module based on the attention mechanism combined with a collaborative factorization machine. DETAILED DESCRIPTION

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0028] like Figures 1 to 4The power grid fault processing cognitive knowledge graph generation system is characterized by comprising a data acquisition module, a knowledge graph construction module, a triple construction module, an embedding acquisition module, an embedding triple acquisition module, a parameter update module, a cognitive knowledge graph construction module, and a decision module;

[0029] The data acquisition module is used to obtain historical data sets of power grid regulation and collect information on factors that affect the status of power grid equipment nodes. The historical data sets of power grid regulation include information on equipment nodes in the power grid, a set of fault handling actions, and information on factors that affect the status of power grid equipment nodes, including current, sound decibels, and equipment temperature.

[0030] The knowledge graph construction module uses prior knowledge of fault handling and combines the physical status of power grid equipment and corresponding fault handling actions in the historical data of power grid regulation to construct a knowledge graph. Prior knowledge of fault handling is a processing guide provided by historical processing experience and knowledge bases such as books.

[0031] The triple construction module uses the knowledge graph and the relationship between the state transitions of each device entity in the power grid control historical data set to extract the device nodes and fault handling actions in the power grid control historical data set in triple format (the triple format includes a first device node, a fault handling action, and a second device node obtained after the fault handling action is performed on the first device node). Based on the factor information that affects the state of the power grid device node, each device node is assigned associated multi-factor information to obtain a triple of the device node and the fault handling action;

[0032] The Embedding Acquisition Module uses the device node and fault handling action triples to build a conditional variational autoencoder model. This model is used to obtain the relationship between the multi-factor Embedding information associated with the current device node state and the fault handling action.

[0033] The Embedding triplet acquisition module uses the relationship between the multi-factor Embedding information associated with the current device node state and the fault handling action Embedding to construct an attention mechanism combined with a collaborative factorization machine model. The multi-factor adaptive fusion model framework based on the attention mechanism combined with the collaborative factorization machine is used to adaptively fuse the device Embeddings generated by the conditional variational autoencoder network based on three grid state influencing factors (current, sound, and temperature). The Embedding set of the associated factor group is input into the attention mechanism combined with the collaborative factorization machine model to obtain the feature expression of the device node associated with the associated factor, and obtain the node and fault handling action Embedding triplet;

[0034] The triplet uses a multi-factor adaptive fusion model framework based on an attention mechanism combined with a collaborative factorization machine to obtain the state embedding of power grid device nodes. Fault handling actions are also initialized through a neural network to obtain a triplet of embeddings. This triplet includes the starting state of the power grid device, the required handling action for that state, and the post-handling state of the power grid device. This triplet is equivalent to ("the starting state of the power grid device under the influence of the three factors (current, sound, and temperature)", the required handling action for that state, and "the state of the power grid device under the influence of the three factors (current, sound, and temperature)").

[0035] The parameter update module uses Graph Embedding to calculate the loss of the node and fault handling action Embedding triples. This loss is recorded as loss, and the calculation result is used to update the attention mechanism combined with the collaborative factorization machine model.

[0036] The cognitive knowledge graph construction module uses the embeddings of associated factor groups and the relationship embeddings of fault handling actions, and uses the updated attention mechanism combined with the collaborative factor decomposition machine model to adaptively fuse the multiple factors that constitute the device node state to obtain the state expression of the device node, generating a more accurate expression of the power grid fault handling cognitive knowledge graph.

[0037] The decision module is used to obtain the most reasonable decision path based on the representation of nodes and edges and using the source nodes and target nodes in the power grid fault processing cognitive knowledge graph.

[0038] In the above technical solution, the device node and fault handling action triplet is a triplet of the first device node characteristic factor, the fault handling action, and the second device node characteristic factor, which can be fused by the three parts (current factor, sound factor, temperature factor). The (current factor, sound factor, temperature factor) of the previous node is the fusion of the (current factor, sound factor, temperature factor) possessed by the previous node, and the (current factor, sound factor, temperature factor) of the latter node is the fusion of the (current factor, sound factor, temperature factor) of the latter node. The state representation of the power grid equipment using multimodal features can more accurately express the state of the power grid equipment. The representation angle of the power grid equipment under a single mode is too single. The features under multimodality represent the state of the power grid equipment in multiple dimensions, wherein the multi-dimensional representation can achieve the complementary and enhanced effect of the state features, thereby enriching the features of the power grid equipment state and providing more accurate and effective judgment and guidance for downstream tasks.

[0039] In the above technical solution, the Embedding acquisition module uses a conditional variational autoencoder model to obtain the relationship between the multi-factor Embedding information associated with the current device node state and the fault handling action Embedding. The specific method is:

[0040] The multi-factor information and corresponding labels associated with the node device status in the triplet are input into the conditional variational autoencoder model for feature sampling, and the embeddings corresponding to the multiple factors are generated respectively. The fault handling action is initialized using a neural network to generate a relational embedding, and finally the multi-factor embedding information associated with the current device node status and the relational embedding of the fault handling action are obtained.

[0041] In the above technical solution, the specific method for the knowledge graph construction module to construct the knowledge graph by using prior knowledge of fault handling combined with the entity status of power grid equipment and corresponding fault handling actions in the power grid control historical data set is as follows:

[0042] Obtaining a fault handling record of each grid device node in its initial state, the fault handling record including the grid device entity state and the corresponding fault handling action;

[0043] Each power grid equipment entity state in the fault handling record is regarded as an entity node in the knowledge graph, and the fault handling action taken for each power grid equipment entity state is regarded as the association edge between entity nodes;

[0044] The grid equipment node states in the entire grid control historical data set are associated according to the edges corresponding to the fault handling actions. At the same time, the corresponding fault handling actions in the historical records are corrected based on the prior knowledge of fault handling, and finally a knowledge graph containing the grid equipment node states and fault handling actions in the data set is formed.

[0045] Using entities in the dataset as nodes and action items as edges, the edges connect corresponding nodes according to actual circumstances, forming a basic knowledge graph. The various connections in the graph are then corrected using prior knowledge to obtain a corrected knowledge graph. This generates a knowledge graph generated from historical records, corrected by prior knowledge, which serves as the foundation for subsequent embedding of entities (grid equipment node status) and fault handling actions in the knowledge graph.

[0046] In the above technical solution, the specific method for the triplet construction module to obtain the device node and fault handling action triplet is:

[0047] Create a data structure of node information, which includes the device temperature, device current and device sound information associated with the node;

[0048] Based on the original device node and fault handling action triples in the power grid control historical data set, information reconstruction is performed to obtain a set of device node and fault handling action triples containing three correlation factors in each node, namely ((current factor, sound factor, temperature factor), action, (current factor, sound factor, temperature factor)).

[0049] In the above technical solution, the specific method for the Embedding acquisition module to obtain the relationship Embedding of the multi-factor Embedding information and the fault handling action associated with the current device node status is:

[0050] Based on the state of the power grid device node (the node state is formed by the fusion of (current factor, sound factor, temperature factor)), the entity class corresponding to each power grid device node state is defined (referring to the type of fault of the power grid device node. Different faults have different names, such as different types of short circuit faults, different types of open circuit faults, etc.); at the same time, the dimension size of each node state is embed_size, that is, the feature embedding dimension (the dimension of the node state is determined by the amount of feature information to be stored, and the specific value is defined through experimental trials and experience). The dimension of the impact factor associated with each device node is defined as embed_size (each device node is formed by the fusion of the triple (current factor, sound factor, temperature factor) features. The size is embed_size).

[0051] Construct a conditional variational autoencoder model framework. Use the correlation factor information and label information corresponding to the device node as the input of the conditional variational autoencoder to generate a latent feature space. The latent feature space and the normal distribution are constrained by KL divergence. Then, use the decoder and the corresponding label information to obtain the output. The output and input are combined to calculate the loss and update the conditional variational autoencoder model parameters, ultimately obtaining a conditional variational autoencoder.

[0052] The association factor of the device node and its corresponding label information are input into the conditional variational autoencoder. The feature sampling of the association factor is performed in the latent feature space through the label information to achieve embedding of the association factor of the device node.

[0053] Define the node self-encoding model and initialize it using the node self-encoding model according to the device node information in the power grid dataset. The dimension of the initialization vector is embed_size.

[0054] Use a neural network to initialize the embedding of the fault handling action, and set the dimension of the initialization vector of the fault handling action to 1*embed_size, which is one times the feature embedding dimension;

[0055] Finally, the relationship between the multi-factor embedding information associated with the current device node status and the fault handling action is obtained, that is, the device node is embedded and the fault handling action is embedded.

[0056] The conditional variational autoencoder (CVAE) model framework is used to encode device nodes. In this paper, CVAE maps the encodings of device nodes under three correlation factors (current factor, sound factor, and temperature factor) into the same encoding space. This facilitates the subsequent fusion of the resulting mappings, ultimately resulting in an encoding for the node with all three correlation factors (current factor, sound factor, and temperature factor). The CVAE model framework outputs the node encoding, or embedding.

[0057] In the correlation factor information corresponding to the device node, (current factor, sound factor, temperature factor), action, (current factor, sound factor, temperature factor) "can be equivalent to (device node A, action, device node B), so device node A is composed of the three correlation factors (current factor, sound factor, temperature factor);

[0058] The label information is three related factors as labels, namely "current factor" label, "sound factor" label, and "temperature factor" label.

[0059] In the above technical solution, the specific method for the Embedding triplet acquisition module to acquire the node and fault handling action Embedding triplet is: obtaining the embedded device node and fault handling action;

[0060] A multi-factor adaptive fusion model framework based on the attention mechanism combined with a collaborative factorization machine is constructed. This model framework uses two different forms of attention substructures to collaboratively model feature combinations. It consists of four parts: a data encoding layer, a low-order interaction layer of the attention subnetwork, a feature extraction layer of the empowerment module, and an output layer.

[0061] The three embeddings obtained by conditional variational autoencoding under the constraints of current factor, sound factor, and temperature factor are input into the multi-factor adaptive fusion model framework based on the attention mechanism combined with the collaborative factorization machine. The multi-factor features are fused and the feature dimension obtained by fusion is called embed_size, which is used as the state embedding of the relevant device node.

[0062] Based on the acquired device node embeddings and fault action embeddings, representation learning and updating are performed on the nodes and fault handling actions in the historical records of power grid regulation and control. Ultimately, a triplet of node and fault handling action embeddings is obtained. The learning and updating process involves assigning the newly acquired embeddings to the nodes, replacing the previous embeddings with the newly acquired embeddings. This design applies the accurate results obtained after training to the grid node feature embeddings and fault handling action feature embeddings, thereby updating the embeddings.

[0063] A multi-factor adaptive fusion model framework based on an attention mechanism combined with a collaborative factorization machine is used to fuse the embeddings generated for a node under the constraints of three corresponding correlation factors to obtain the node's final embedding. The input information of this multi-factor adaptive fusion model framework based on an attention mechanism combined with a collaborative factorization machine is the node's embedding under the current factor constraint, the node's embedding under the sound factor constraint, and the node's embedding under the temperature factor constraint. The output information is the final node embedding obtained by fusing the embeddings generated under the three correlation factor constraints. This solution obtains multi-dimensional feature embeddings for power grid equipment nodes, enriches the power grid equipment status characteristics, and generates a more precise cognitive knowledge graph.

[0064] The multi-factor adaptive fusion model framework based on the attention mechanism combined with the collaborative factor decomposition machine performs attention mechanism interaction and multi-scale feature fusion on the power equipment state feature embedding obtained based on the "current factor, sound factor, and temperature factor", effectively complementing and fusing multi-dimensional features.

[0065] In the above technical solution, the parameter update module calculates the loss of the node and fault handling action embedding triples, uses the calculation results to initialize the relational embedding of the fault handling action, and updates the attention mechanism combined with the collaborative factorization machine model;

[0066] The specific method for initializing the relationship embedding of fault handling actions and updating the attention mechanism combined with the collaborative factorization machine model is as follows:

[0067] The loss of the node and fault handling action Embedding triple is calculated using Graph Embedding. This loss is recorded as loss and is used to initialize the Embedding for the fault action and update the parameters of the collaborative factorization machine model based on the attention mechanism.

[0068] First, we define a Graph Embedding model based on TransR and use it to calculate the representation loss of triples in the knowledge graph.

[0069] Input the embedded node and fault handling action triples into the TransR-based GraphEmbedding model. The output is the sum of the node embedding and edge embedding calculated by TransR, and the distance loss between them and the embeddings of adjacent nodes, denoted as loss. For example, for (node ​​A's embedding, edge embedding, node B's embedding), the loss function is calculated as (node ​​A's embedding + edge embedding + distance loss between node B's embedding).

[0070] The loss function is used to update the parameters of the fault action initialization embedding and the attention-based collaborative factorization machine model. This method is a native feature of the deep learning framework. The backpropagation mechanism is used to update the parameters of the neural network to obtain a trained model.

[0071] The specific method for the cognitive knowledge graph construction module to generate the power grid fault processing cognitive knowledge graph is:

[0072] Obtain the embedding and attention mechanism combined with the collaborative factorization machine model of the fault handling action. This model is a neural network after the parameter update. When the parameter update is completed, the model training is completed. The attention mechanism combined with the collaborative factorization machine model is a model obtained after the training is completed. This model has the feature fusion function required by the present invention;

[0073] The device node association factors are sampled from the latent feature space generated by the conditional variational autoencoder model encoder through the corresponding labels to generate an embedding representation of the association factors.

[0074] The three embeddings of the device node obtained through conditional variational autoencoding under the constraints of current factor, sound factor, and temperature factor are input into the attention mechanism combined with the collaborative factorization machine model. The embedding representation of the device node is obtained through the adaptive fusion of the correlation factors.

[0075] By combining embeddings for fault handling actions and device nodes, a cognitive knowledge graph for power grid fault handling is constructed. This generated cognitive knowledge graph can guide power grid regulation. For example, when encountering relevant issues, the cognitive knowledge graph can be used to recommend regulatory decisions.

[0076] This method first constructs a knowledge graph for power grid fault handling and then embeds the entity nodes (power grid equipment status) and edges (fault handling actions) within the knowledge graph. This knowledge graph is then given a knowledge representation, where both edges and nodes have practical meaning, thereby forming a cognitive knowledge graph for power grid fault handling. This cognitive knowledge graph can then be used to generate power grid equipment fault handling sequences, whether through manual analysis, simple graph distance calculation, or reinforcement learning.

[0077] In the above technical solution, the decision module is used to obtain the most reasonable decision path based on the representation of nodes and edges, and using the source nodes (current problems and status) and target nodes (target status to be achieved) in the power grid fault processing cognitive knowledge graph. The specific method is as follows:

[0078] Obtain the device node status and target status that require adjustment (obtain the grid problem faced by the current employee (assuming the "grid administrator"). This problem inevitably points to a current grid state that requires adjustment. If adjustment is not required, the administrator will not propose it. If it is proposed, adjustment is required).

[0079] In the cognitive graph, fault handling actions are recommended for the device node information and target status; the power grid fault handling cognitive knowledge graph is equivalent to a dictionary or tool library for fault handling. Whether using manual or simple distance calculation of the graph or reinforcement learning, power grid equipment fault handling sequences can be generated on this cognitive knowledge graph to complete or guide the automation of power grid fault handling.

[0080] The most reasonable decision path is obtained based on the recommended fault handling actions. Specifically, each node and edge in the cognitive knowledge graph has a precise embedding representation. From the problem node state to the target state, the path with minimized loss can be obtained through traversal calculation or the optimal path can be obtained through reinforcement learning traversal. Through comparison, the most reasonable decision path is obtained.

[0081] The cognitive knowledge graph is connected, enabling path generation. Each node and edge in the cognitive knowledge graph has feature embeddings, enabling direct distance calculation. For connected nodes in the graph, there is at least one path between the current node and the target node. Therefore, multiple paths can be generated for comparison, allowing the most appropriate fault resolution path to be selected.

[0082] A method for generating a cognitive knowledge graph for power grid fault processing includes the following steps:

[0083] Step 1: The knowledge graph construction module uses prior knowledge of fault handling and combines the entity status of power grid equipment and corresponding fault handling actions in the power grid control historical data set to construct a knowledge graph;

[0084] Step 2: The triple construction module uses the knowledge graph and the relationship between the state transitions of each device entity in the power grid control historical dataset to extract the device nodes and fault handling actions in the power grid control historical dataset in triple format. Based on the factor information affecting the state of the power grid device node, each device node is assigned associated multi-factor information to obtain the device node and fault handling action triple;

[0085] Step 3: The embedding acquisition module uses the device node and fault handling action triples to build a conditional variational autoencoder model. The conditional variational autoencoder model is used to obtain the relationship embeddings between the multi-factor embedding information associated with the current device node state and the fault handling action.

[0086] Step 4: The Embedding triplet acquisition module uses the multi-factor Embedding information associated with the current device node state and the relationship Embedding of the fault handling action to build an attention mechanism combined with a collaborative factorization machine model. The Embedding set of the associated factor group is input into the attention mechanism combined with the collaborative factorization machine model to obtain the feature expression of the device node associated with the associated factor, and obtain the node and fault handling action Embedding triplet;

[0087] Step 5: The parameter update module calculates the loss for the node and fault handling action Embedding triples, and uses the calculation results to update the attention mechanism combined with the collaborative factorization machine model;

[0088] Step 6: The cognitive knowledge graph construction module uses the embeddings of the associated factor groups and the relationship embeddings of the fault handling actions, and uses the updated attention mechanism combined with the collaborative factor decomposition machine model to adaptively fuse the multiple factors that constitute the device node status to generate a cognitive knowledge graph for power grid fault handling.

[0089] A computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the above-described method. In the task of power grid fault control, most existing algorithms consider using model strategies to guide fault handling decisions. Even when innovating in the knowledge graph direction, they often ignore the impact of multiple factors associated with the state of power grid device nodes on the feature expression of the device nodes. The present invention primarily generates a cognitive knowledge graph for power grid fault handling based on an attention mechanism combined with a collaborative factorization machine. First, a conditional variational autoencoder model is used to encode the multiple factors that constitute the state of the device node, fully utilizing the factor information associated with the power grid device node to enrich the feature expression of the power grid device node. Then, an attention mechanism combined with a collaborative factorization machine's multi-factor adaptive fusion module is used to screen important information from the multi-factor combination features and adaptively fuse the multiple factors that constitute the device node state to obtain the feature expression of the device node state. An autoencoder is used to encode the fault handling action. Finally, the encoded information of nodes and edges in the power grid control historical dataset is embedded in the knowledge graph. The loss is calculated and the model parameters based on the attention mechanism combined with the collaborative factorization machine are updated in reverse. Finally, a multi-factor adaptive fusion model is generated, which can effectively fuse the multiple factors of the related power grid equipment node status and generate more information-rich equipment node information, thereby improving the accuracy of node representation in the power grid fault processing cognitive knowledge graph. The power grid fault processing cognitive knowledge graph constructed based on this can better guide the selection of fault handling actions and obtain more reasonable fault handling effects.

[0090] On the one hand, the present invention utilizes a conditional variational autoencoder to encode the multiple factors constituting the state of a device node separately. On the other hand, a multi-factor adaptive fusion module that utilizes an attention mechanism combined with a collaborative factor decomposition machine is utilized to complete the conversion of the multi-factor encoding into the device node representation. The multi-factor encoding module based on the conditional variational autoencoder uses the factor data information and the corresponding label information as the input of the conditional variational autoencoder to generate a latent feature space, and then uses a decoder and a label to obtain the output. The output and the input are subjected to loss calculation to update the model parameters, and finally a conditional encoder is obtained. The target data can be sampled from the latent feature space according to the label corresponding to the data. For the multi-factor adaptive fusion module that utilizes the attention mechanism combined with the collaborative factor decomposition machine, two different forms of attention substructures are used to collaboratively model feature combinations, which include a low-order interaction layer of the attention subnetwork and a feature extraction layer of the weighting module, so as to enrich the information expression of the device node and improve the effectiveness of feature acquisition.

[0091] The present invention fully considers the multi-factor information that constitutes the node status of power grid equipment, and can effectively improve the richness and accuracy of node expression in the power grid fault processing cognitive knowledge graph. The generated more reasonable power grid fault processing cognitive knowledge graph can better guide the current node to select a more reasonable fault processing action instruction set, improve the quality of fault handling actions, and ultimately make the decision-making effect more ideal.

[0092] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A cognitive knowledge graph generation system for power grid fault processing, characterized by: It includes a knowledge graph construction module, a triple construction module, an embedding acquisition module, an embedding triple acquisition module, a parameter update module and a cognitive knowledge graph construction module; The knowledge graph construction module uses prior knowledge of fault handling and combines the entity status of power grid equipment and corresponding fault handling actions in the historical data set of power grid control to construct a knowledge graph; The triple construction module uses the knowledge graph and the relationship between the state transitions of each device entity in the power grid control historical data set to extract the device nodes and fault handling actions in the power grid control historical data set in triple format. According to the factor information affecting the state of the power grid device node, each device node is assigned associated multi-factor information to obtain the device node and fault handling action triple; The Embedding Acquisition Module uses the device node and fault handling action triples to build a conditional variational autoencoder model. This model is used to obtain the relationship between the multi-factor Embedding information associated with the current device node state and the fault handling action. The Embedding triplet acquisition module uses the relationship between the multi-factor Embedding information associated with the current device node status and the fault handling action Embedding to build an attention mechanism combined with a collaborative factor decomposition machine model. The Embedding set of the associated factor group is input into the attention mechanism combined with the collaborative factor decomposition machine model to obtain the feature expression of the device node associated with the associated factor, and obtain the node and fault handling action Embedding triplet; The parameter update module calculates the loss of the node and fault handling action Embedding triples, and uses the calculation results to update the attention mechanism combined with the collaborative factorization machine model; The cognitive knowledge graph construction module uses the embeddings of associated factor groups and the relationship embeddings of fault handling actions, and uses the updated attention mechanism combined with the collaborative factor decomposition machine model to adaptively fuse the multiple factors that constitute the device node status to generate a cognitive knowledge graph for power grid fault handling.

2. The power grid fault processing cognitive knowledge graph generation system according to claim 1 is characterized by: It also includes a data acquisition module, which is used to obtain historical data sets of power grid regulation and control, and collect information on factors that affect the status of power grid equipment nodes; It also includes a decision module, which is used to obtain the most reasonable decision path by using the source nodes and target nodes in the power grid fault processing cognitive knowledge graph.

3. The power grid fault processing cognitive knowledge graph generation system according to claim 1, characterized in that: The Embedding acquisition module uses a conditional variational autoencoder model to obtain the relationship between the multi-factor Embedding information associated with the current device node state and the fault handling action. The specific method is: The multi-factor information and corresponding labels associated with the node device status in the triplet are input into the conditional variational autoencoder model for feature sampling, and the embeddings corresponding to the multiple factors are generated respectively. The fault handling action is initialized using a neural network to generate a relational embedding, and finally the multi-factor embedding information associated with the current device node status and the relational embedding of the fault handling action are obtained.

4. The power grid fault processing cognitive knowledge graph generation system according to claim 1, characterized in that: The specific method for constructing the knowledge graph by using the prior knowledge of fault handling combined with the entity status of the power grid equipment and the corresponding fault handling actions in the power grid control historical data set is as follows: Obtaining a fault handling record of each grid device node in its initial state, the fault handling record including the grid device entity state and the corresponding fault handling action; Each power grid equipment entity state in the fault handling record is regarded as an entity node in the knowledge graph, and the fault handling action taken for each power grid equipment entity state is regarded as the association edge between entity nodes; The grid equipment node states in the entire grid control historical data set are associated according to the edges corresponding to the fault handling actions. At the same time, the corresponding fault handling actions in the historical records are corrected based on the prior knowledge of fault handling, and finally a knowledge graph containing the grid equipment node states and fault handling actions in the data set is formed.

5. The power grid fault processing cognitive knowledge graph generation system according to claim 1, characterized in that: The specific method for the triplet construction module to obtain the device node and fault handling action triplet is: Create a data structure of node information, which includes the device temperature, device current and device sound information associated with the node; Based on the original device node and fault handling action triples in the power grid control historical data set, information reconstruction is performed to obtain a set of device node and fault handling action triples containing three correlation factors in each node.

6. The power grid fault processing cognitive knowledge graph generation system according to claim 1, characterized in that: The specific method for the embedding acquisition module to obtain the multi-factor embedding information associated with the current device node status and the relationship embedding of the fault handling action is: According to the state of the power grid device node, define the entity class corresponding to each power grid device node state; at the same time, the dimension size of each node state is defined as embed_size, and the dimension of the impact factor associated with each device node is defined as embed_size; Construct a conditional variational autoencoder model framework. Use the correlation factor information and label information corresponding to the device node as the input of the conditional variational autoencoder to generate a latent feature space. The latent feature space and the normal distribution are constrained by KL divergence. Then, use the decoder and the corresponding label information to obtain the output. The output and input are combined to calculate the loss and update the conditional variational autoencoder model parameters, ultimately obtaining a conditional variational autoencoder. The association factor of the device node and its corresponding label information are input into the conditional variational autoencoder. The feature sampling of the association factor is performed in the latent feature space through the label information to achieve embedding of the association factor of the device node. Define the node self-encoding model and initialize it using the node self-encoding model according to the device node information in the power grid dataset. The dimension of the initialization vector is embed_size. Use neural networks to initialize embedding for fault handling actions; Finally, the relationship embedding between the multi-factor embedding information associated with the current device node status and the fault handling action is obtained.

7. The power grid fault processing cognitive knowledge graph generation system according to claim 1, characterized in that: The specific method for the Embedding triplet acquisition module to acquire the node and fault handling action Embedding triplet is: obtaining the embedded device node and fault handling action; A multi-factor adaptive fusion model framework based on the attention mechanism combined with a collaborative factorization machine is constructed. This model framework uses two different forms of attention substructures to collaboratively model feature combinations. It consists of four parts: a data encoding layer, a low-order interaction layer of the attention subnetwork, a feature extraction layer of the empowerment module, and an output layer. The three embeddings obtained by conditional variational autoencoding under the constraints of current factor, sound factor, and temperature factor are input into the multi-factor adaptive fusion model framework based on the attention mechanism combined with the collaborative factorization machine. The multi-factor features are fused and the feature dimension obtained by fusion is called embed_size, which is used as the state embedding of the relevant device node. Based on the obtained device node embedding and fault action embedding, the nodes and fault handling actions in the power grid control history are represented and learned and updated, and finally the embedding triples of nodes and fault handling actions are obtained.

8. The power grid fault processing cognitive knowledge graph generation system according to claim 1, characterized in that: The parameter update module calculates the loss of the node and fault handling action embedding triples, uses the calculation results to initialize the relational embedding of the fault handling action, and updates the attention mechanism combined with the collaborative factorization machine model; The specific method for initializing the relationship embedding of fault handling actions and updating the attention mechanism combined with the collaborative factorization machine model is as follows: The loss of the node and fault handling action Embedding triple is calculated using Graph Embedding. This loss is recorded as loss and is used to initialize the Embedding for the fault action and update the parameters of the collaborative factorization machine model based on the attention mechanism. First, we define a Graph Embedding model based on TransR and use it to calculate the representation loss of triples in the knowledge graph. Input the embedded node and fault handling action triples into the TransR-based GraphEmbedding model. The output is the TransR-based calculation of the node and edge embedding and the distance loss between the node and the adjacent node, which is recorded as loss. Use the loss function to update the parameters of the fault action initialization embedding and the attention mechanism combined with the collaborative factorization machine model; The specific method for the cognitive knowledge graph construction module to generate the power grid fault processing cognitive knowledge graph is: Obtaining the embedding and attention mechanism of fault handling actions combined with the collaborative factorization machine model; The device node association factors are sampled from the latent feature space generated by the conditional variational autoencoder model encoder through the corresponding labels to generate an embedding representation of the association factors. The three embeddings of the device node obtained through conditional variational autoencoding under the constraints of current factor, sound factor, and temperature factor are input into the attention mechanism combined with the collaborative factorization machine model. The embedding representation of the device node is obtained through the adaptive fusion of the correlation factors. Combine the embeddings of fault handling actions and device nodes to build a cognitive knowledge graph for power grid fault handling.

9. A method for generating a cognitive knowledge graph for power grid fault processing, characterized in that: It includes the following steps: Step 1: The knowledge graph construction module uses prior knowledge of fault handling and combines the entity status of power grid equipment and corresponding fault handling actions in the power grid control historical data set to construct a knowledge graph; Step 2: The triple construction module uses the knowledge graph and the relationship between the state transitions of each device entity in the power grid control historical dataset to extract the device nodes and fault handling actions in the power grid control historical dataset in triple format. Based on the factor information affecting the state of the power grid device node, each device node is assigned associated multi-factor information to obtain the device node and fault handling action triple; Step 3: The embedding acquisition module uses the device node and fault handling action triples to build a conditional variational autoencoder model. The conditional variational autoencoder model is used to obtain the relationship embeddings between the multi-factor embedding information associated with the current device node state and the fault handling action. Step 4: The Embedding triplet acquisition module uses the multi-factor Embedding information associated with the current device node state and the relationship Embedding of the fault handling action to build an attention mechanism combined with a collaborative factorization machine model. The Embedding set of the associated factor group is input into the attention mechanism combined with the collaborative factorization machine model to obtain the feature expression of the device node associated with the associated factor, and obtain the node and fault handling action Embedding triplet; Step 5: The parameter update module calculates the loss for the node and fault handling action Embedding triples, and uses the calculation results to update the attention mechanism combined with the collaborative factorization machine model; Step 6: The cognitive knowledge graph construction module uses the embeddings of the associated factor groups and the relationship embeddings of the fault handling actions, and uses the updated attention mechanism combined with the collaborative factor decomposition machine model to adaptively fuse the multiple factors that constitute the device node status to generate a cognitive knowledge graph for power grid fault handling.

10. A computer-readable storage medium storing a computer program, wherein: When the computer program is executed by a processor, the steps of the method according to claim 9 are implemented.

Images