Two-stage fusion method and system for multi-source data

Abstract

The application discloses a two-stage fusion method and system for multi-source data, and the method comprises the following steps: acquiring multi-source heterogeneous data, wherein the multi-source heterogeneous data comprises communication network side data and power network side data; extracting initial standardized reconstruction features of the multi-source heterogeneous data; performing attribute level shallow fusion in the initial standardized reconstruction features based on the initial standardized reconstruction features, to obtain attribute level shallow features; performing deep feature aggregation based on the attribute level shallow features, to obtain a user aggregated feature relationship matrix; and constructing a fine-grained knowledge graph based on attribute features of a traditional power network and the user aggregated feature relationship matrix. The application proposes a two-stage data fusion method, which performs fine-grained fusion on features from the attribute level and the user level, respectively, to obtain more effective and accurate fusion information, which is provided to power distribution network dispatchers for auxiliary decision-making and maintenance record keeping.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network big data mining technology, specifically involving a two-stage fusion method for multi-source data. Background Technology

[0002] Power grid engineering projects are characterized by their large scale, long duration, and complex processes. This results in a vast amount of diverse and complex data throughout the entire process, from design and construction to acceptance. This data includes structured and semi-structured data (such as equipment inventories) and unstructured text data (such as design specifications). This massive amount of text data contains rich knowledge related to power grid topology, assets, and equipment, serving as a data source for departments such as operation and maintenance, dispatching, and finance within the power grid company. However, the text data generated in power grid infrastructure projects is primarily intended for reading by project personnel. Therefore, the unformatted text content is difficult for computers to process and understand, significantly increasing the difficulty and cost of extracting knowledge from this data. Therefore, applying knowledge graphs to the power industry can centrally process and analyze dispersed power data, thereby ensuring the universality and standardization of power grid data and guaranteeing the authenticity and consistency of data information in smart grid construction.

[0003] At present, the theory of power grid fault handling based on knowledge graphs is relatively mature and the application results are quite obvious. However, the following problems have been found in actual research: (1) The amount of text data such as dispatching procedures is relatively small, and traditional entity recognition models are difficult to meet the application requirements; (2) Power grid fault handling work has high requirements for response speed, so mastering real-time data can realize real-time auxiliary decision-making; (3) Power grid faults are complex, and the graph support provided for only a single type of traditional power grid data can no longer meet the needs of power grid dispatching and user experience requirements.

[0004] Meanwhile, with the increasing coverage of power grid users, the causes of power grid failures have become more complex. To ensure the safety of users' lives and property, there are increasing demands on power grid failure handling to achieve faster response times, shorter handling times, and higher accuracy. Efficient power grid failure handling requires dispatchers to rapidly and accurately grasp multi-source heterogeneous data, including power grid equipment topology, power grid failure handling plans, provincial and municipal dispatching procedures, and real-time equipment parameters. This poses a significant challenge to the current dispatching methods that rely primarily on manual experience analysis. Furthermore, text data in the power grid domain has its own domain characteristics, and some text data is relatively small, making it difficult for traditional entity recognition technologies to accurately identify entities in the power grid domain.

[0005] Therefore, in traditional power distribution networks, the working mode no longer fully meets the requirements for real-time, efficient power grid fault analysis and user experience due to problems such as high labor costs, limited diagnostic methods, and decentralized data management.

[0006] Among related technologies, Chinese invention patent document CN113947320A describes a power grid control method based on multimodal reinforcement learning, including the following steps: Step 1, inputting a power grid dataset; Step 2, constructing a knowledge graph containing the state of power grid equipment nodes and fault handling in the dataset; Step 3, obtaining a multimodal information fusion module that can embedding the current multimodal state of equipment nodes and fault handling actions; Step 4, selecting the knowledge graph constructed in Step 2, defining a multi-hop scoring function based on the current state to the target state to evaluate the correlation between two states; Step 5, constructing the meta-path of fault handling; Step 6, generating a reward function for reinforcement learning; Step 7, defining a Markov process for reinforcement learning to generate a policy-based reinforcement learning framework; Step 8, training to generate a power grid control strategy based on multimodal reinforcement learning.

[0007] This solution analyzes multimodal data of power grid equipment and studies the status of power grid equipment. It processes different modal information of a single device and also involves image information. It provides an encoder from the multimodal data of the device to extract the text information and image information features of the corresponding device, and then extracts the device status from the features. Since the information comes from the same device, they are all correlated and there is no need to perform data fusion.

[0008] Chinese invention patent document CN113254663A describes a knowledge graph joint representation learning method that integrates graph convolution and translation models, including the following steps: 1) Constructing the corresponding direct adjacency matrix and indirect adjacency matrix based on the knowledge graph; 2) Designing a graph convolutional network, including an input layer, two hidden layers and an output layer, optimizing the attention coefficients of adjacent nodes to the center node, and obtaining the vector representation of the nodes by learning the structural information of the direct and indirect neighboring nodes; 3) Using a translation model to learn the semantic information of relations, obtaining the vector representation of entities and relations; 4) Integrating the graph convolutional network and the translation model, and obtaining the final vector representation of the knowledge graph through continuous iterative learning.

[0009] The neural network used in this scheme is a graph convolutional network, whose main function is to fuse the features of neighboring nodes and perceive the features of nodes that are adjacent to each other. Essentially, it is a local neighbor perception at the node level. Its task mechanism is to know the connection matrix between the nodes, and then use a global attention mechanism and softmax to obtain more features from nodes that are closer to the node and fewer features from nodes that are farther away.

[0010] Chinese invention patent document CN113704500A discloses a knowledge graph community partitioning method based on graph neural networks, including the following steps: 1) constructing a knowledge graph and its adjacency matrix; 2) using graph neural networks to learn the representation of nodes in the knowledge graph to obtain the vector representation of the nodes; 3) calculating the similarity between nodes based on cosine similarity; 4) constructing an undirected weighted node relationship network based on the adjacency relationship between nodes of the same type and the node similarity; 5) setting modularity to evaluate the cohesion within the community; 6) partitioning the target nodes in the knowledge graph into communities using a community partitioning algorithm.

[0011] The knowledge graph constructed by this scheme only establishes relationships between nodes based on their similarity, representing coarse-grained associations between user nodes, and does not involve attribute relationships between nodes. Summary of the Invention

[0012] The technical problem to be solved by this invention is how to more effectively and accurately integrate multi-source data to provide auxiliary decision-making and maintenance records for power distribution network dispatchers.

[0013] The present invention solves the above-mentioned technical problems through the following technical means:

[0014] This invention proposes a two-stage fusion method for multi-source data, the method comprising the following steps:

[0015] Acquire multi-source heterogeneous data, including communication network side data and power grid side data;

[0016] Extract the initial standardized reconstruction features of the multi-source heterogeneous data;

[0017] Based on the initial standardized reconstruction features, attribute-level shallow fusion is performed within the initial standardized reconstruction features to obtain attribute-level shallow features.

[0018] Based on the aforementioned attribute-level shallow features, deep feature aggregation is performed to obtain the user aggregated feature relationship matrix;

[0019] A fine-grained knowledge graph is constructed based on the attribute characteristics of the traditional power grid and the relationship matrix of the user aggregation characteristics.

[0020] To fully utilize multi-source heterogeneous information, this invention proposes a two-stage data fusion method. This method performs fine-grained fusion of features at both the shallow attribute level and the deep user level. Furthermore, by leveraging the relationships between data attributes, a fine-grained knowledge graph of the user network can be constructed without additional computational load. This provides a reliable foundation for further analysis of user power aggregation characteristics. It effectively integrates auxiliary data sources from any other system on top of traditional distribution network data, resulting in more efficient and accurate fused information, which is then provided to distribution network dispatchers for decision-making and maintenance record-keeping.

[0021] Furthermore, the communication network-side data includes heterogeneous data generated by user-side network equipment, and the power grid-side data includes heterogeneous data generated by the PMS system and heterogeneous data generated by the electricity consumption information collection system.

[0022] Furthermore, after acquiring multi-source heterogeneous data, which includes communication network-side data and power grid-side data, the method further includes:

[0023] Text data preprocessing is performed on the multi-source heterogeneous data, and the multi-source heterogeneous data is structured type classification and data filtering are performed to obtain preprocessed multi-source heterogeneous data.

[0024] Accordingly, the extraction of the initial standardized reconstruction features of the multi-source heterogeneous data specifically includes:

[0025] Extract the initial standardized reconstruction features of the preprocessed multi-source heterogeneous data.

[0026] Furthermore, the extraction of the initial standardized reconstruction features of the multi-source heterogeneous data includes:

[0027] Data with a confidence level higher than a set threshold in the multi-source heterogeneous data are used as filter data, and the filter data is stored in the knowledge base in a networked manner using a graph database.

[0028] Using an NLP feature model, features are extracted from the filtered data corresponding to the communication network side data and the filtered data corresponding to the power grid side data, respectively, to obtain the corresponding first attribute feature set and second attribute feature set.

[0029] The first attribute feature set and the second attribute feature set are concatenated at the attribute level to obtain the initial standardized reconstructed features.

[0030] Further, the step of performing attribute-level shallow fusion within the initial standardized reconstructed features to obtain attribute-level shallow features includes:

[0031] A multilayer perceptron is used to perform attribute-level shallow fusion within the initial standardized reconstructed features to obtain attribute-level shallow features. The multilayer perceptron includes an input layer, a hidden layer, and an output layer. The hidden layer includes h hidden units, wherein:

[0032] The output of the hidden layer is:

[0033] H = XW h +b h

[0034] The output of the output layer is:

[0035] O = HW o +b o

[0036] In the formula: the weights and biases of the hidden layer are respectively represented by W. h ∈R n×h ,b h ∈R 1×h The output dimension of the hidden layer is q; the weights and bias parameters of the output layer are W and W, respectively. o ∈R n×q ,b o ∈R 1×q .

[0037] Furthermore, each of the hidden units is followed by a non-linear activation function, and the output of the activation function is:

[0038] H=σ(XW h +b h )

[0039] In the formula: σ() represents the activation function.

[0040] Furthermore, the step of performing deep feature aggregation based on the attribute-level shallow features to obtain the user aggregated feature relationship matrix includes:

[0041] Graph attention deep encoding is performed on the attribute-level shallow features to construct an image. As graphical data, each user feature vector in the attribute-level shallow features serves as a node in the image, and for each node v i For all Add a from v j Pointing to v i The edge e ji , For node v i The K nearest neighbors;

[0042] Based on the graph data, features from neighboring nodes are aggregated using a graph attention layer, and information is exchanged between the nodes to obtain the user aggregated feature relationship matrix. The computational operations within the graph attention layer are as follows:

[0043]

[0044] In the formula: Aggregate() and Update() are respectively the features through W agg and W update These two matrices are multiplied to perform feature aggregation and feature update operations, W agg and W update Here, represents the learnable weights of a set of aggregation operations and the learnable weights of a set of update operations, respectively.

[0045] Furthermore, the feature aggregation operation is used to compute the representation of the current node by aggregating the features of neighboring nodes:

[0046]

[0047] In the formula: It is x i The set of adjacent nodes;

[0048]

[0049] h(·)=x′ i =x″ i W update

[0050] In the formula: GAT represents a graph attention network;

[0051] The feature update operation is used to update h heads with different weights, each head being a pair of aggregated features x″. i After splitting, we get:

[0052]

[0053] In the formula: head 1 head 2 head h Each head represents a different head; These represent the weights corresponding to each head.

[0054] Furthermore, the construction of a fine-grained knowledge graph based on the attribute features of the traditional power grid and the user aggregation feature relationship matrix includes:

[0055] Based on the attribute characteristics of the traditional power grid, a knowledge graph of the traditional power grid is constructed;

[0056] Based on the user aggregation feature relationship matrix and the traditional power grid knowledge graph, a relational interconnection is constructed from the graph attributes of the traditional power grid to the user aggregation feature relationship matrix to obtain the fine-grained knowledge graph.

[0057] Furthermore, this invention also proposes a two-stage fusion system for multi-source data, the system comprising:

[0058] The data acquisition module is used to acquire multi-source heterogeneous data, including communication network side data and power grid side data;

[0059] The feature extraction module is used to extract the initial standardized reconstruction features of the multi-source heterogeneous data;

[0060] The shallow fusion module is used to perform attribute-level shallow fusion within the initial standardized reconstruction features based on the initial standardized reconstruction features, to obtain attribute-level shallow features.

[0061] The deep fusion module is used to perform deep feature aggregation based on the attribute-level shallow features to obtain the user aggregated feature relationship matrix;

[0062] The graph construction module is used to construct a fine-grained knowledge graph based on the attribute characteristics of the traditional power grid and the user aggregation feature relationship matrix.

[0063] The advantages of this invention are:

[0064] (1) In order to make full use of multi-source heterogeneous information, this invention proposes a two-stage data fusion method, which performs fine-grained fusion of features at the shallow attribute level and the deep user level respectively. Furthermore, through the correlation between data attributes, a fine-grained knowledge graph of the user network can be constructed without additional computation. This provides a reliable foundation for further analysis of user power aggregation features. It can effectively integrate auxiliary data sources from any other system on the basis of traditional distribution network data to obtain more effective and accurate fusion information, which can be provided to distribution network dispatchers for auxiliary decision-making and maintenance record keeping.

[0065] (2) Since text data in the power grid field has its own domain characteristics and some text data is relatively small, traditional entity recognition technology is difficult to accurately identify entities in the power grid field. This invention introduces a data source, namely the real-time user-side equipment data network of the operator, to provide simple and effective real-time information, realize the intelligent construction technology of power grid fault handling knowledge graph, provide dispatchers with real-time fault handling auxiliary knowledge graph, and provide support for dispatching and maintenance work.

[0066] (3) The present invention uses a multi-layer perceptual neural network to perform feature fusion within a node, and globally perceives all attribute dimensions within the node, enabling the perception of more fine-grained features, and realizing the use of neural networks to perceive the feature hierarchy of fine-grained attributes within a node.

[0067] (4) Under the condition of unknown relationships between nodes, this invention utilizes graph attention mechanism to integrate node features while paying more attention to adaptive relationship construction.

[0068] (5) The construction of the knowledge graph in this invention involves the relationship between the attributes of each node. After learning the distribution between each user node through graph attention, the distribution between the nodes is used, and then combined with the inherent relationship between the attributes of the nodes, the various attributes in the traditional power grid equipment are further associated to form a fine-grained knowledge graph, rather than being limited to the coarse-grained relationship between user nodes.

[0069] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating a two-stage fusion method for multi-source data proposed in an embodiment of the present invention.

[0071] Figure 2 This is a schematic diagram of the two-stage data fusion principle in one embodiment of the present invention;

[0072] Figure 3 This is a schematic diagram of the shallow fusion principle of a multilayer perceptual neural network in one embodiment of the present invention;

[0073] Figure 4 This is a diagram of the user-level interaction architecture for attention encoding in one embodiment of the present invention;

[0074] Figure 5 This is an example of a multi-source heterogeneous data knowledge graph combining user-side network equipment and traditional power grid systems in one embodiment of the present invention;

[0075] Figure 6 This is a schematic diagram of the structure of a two-stage fusion system for multi-source data proposed in an embodiment of the present invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] like Figures 1 to 2 As shown, the first embodiment of the present invention proposes a two-stage fusion method for multi-source data, the method comprising the following steps:

[0078] S10. Acquire multi-source heterogeneous data, wherein the multi-source heterogeneous data includes communication network side data and power grid side data;

[0079] It should be noted that multi-source heterogeneous data refers to the real-time collection of relevant raw data from various power grid devices, such as electricity consumption, daily limit voltage, power outage signals, etc. The raw data is generated after these data are collected and fed back.

[0080] S20. Extract the initial standardized reconstruction features of the multi-source heterogeneous data;

[0081] S30. Based on the initial standardized reconstruction features, perform attribute-level shallow fusion within the initial standardized reconstruction features to obtain attribute-level shallow features.

[0082] S40. Based on the attribute-level shallow features, perform deep feature aggregation to obtain the user aggregated feature relationship matrix;

[0083] S50. Based on the attribute characteristics of the traditional power grid and the user aggregation feature relationship matrix, construct a fine-grained knowledge graph.

[0084] To fully utilize multi-source heterogeneous information, this invention proposes a two-stage data fusion method. This method performs fine-grained fusion of features at both the shallow attribute level and the deep user level, improving the robustness of user network features and achieving an effective and analyzable feature set. Furthermore, through the correlation between data attributes, a fine-grained knowledge graph of the user network can be constructed without additional computational load. This provides a reliable foundation for further analysis of user power aggregation features. It effectively integrates auxiliary data sources from any other system on top of traditional distribution network data, obtaining more effective and accurate fused information, which is then provided to distribution network dispatchers for auxiliary decision-making and maintenance record keeping.

[0085] In one embodiment, the communication network-side data includes heterogeneous data generated by user-side network equipment, and the power grid-side data includes heterogeneous data generated by the PMS system and heterogeneous data generated by the electricity consumption information collection system.

[0086] It should be noted that the traditional operating mode in power distribution networks, due to high labor costs, limited diagnostic methods, and decentralized data management, no longer fully meets the requirements for real-time, efficient power grid fault analysis and user experience. This embodiment adds an auxiliary data source—the operator's real-time user-monitored equipment data network—to provide simple and effective real-time information. Real-time network data stream feedback, combined with traditional power data research, enhances the accuracy and timeliness of traditional power grid system judgments, providing dispatchers with effective fault handling reference data.

[0087] It should be noted that the embodiments of the present invention are aimed at the text information of all devices in the distribution network, involving multiple power grid devices rather than a single power grid device. To solve the problem of fusion of these heterogeneous data from different sources and to simultaneously construct the relationship between these heterogeneous data, and in order to solve the real-time and accuracy of feedback sensing in the power grid, data from user-side network devices is introduced, which further increases the difficulty of data fusion analysis.

[0088] In one embodiment, after step S10: acquiring multi-source heterogeneous data, wherein the multi-source heterogeneous data includes communication network-side data and power grid-side data, the method further includes the following steps:

[0089] Text data preprocessing is performed on the multi-source heterogeneous data, and the multi-source heterogeneous data is structured type classification and data filtering are performed to obtain preprocessed multi-source heterogeneous data.

[0090] Accordingly, step S20: extracting the initial standardized reconstruction features of the multi-source heterogeneous data specifically includes:

[0091] Extract the initial standardized reconstruction features of the preprocessed multi-source heterogeneous data.

[0092] This embodiment performs text data preprocessing on data from two different data sources: the communication network and the power grid. The approach can be referenced from general text processing models, which involves structuring and classifying the initial data source text and filtering the data.

[0093] In one embodiment, step S20: extracting the initial standardized reconstruction features of the multi-source heterogeneous data includes the following steps:

[0094] S21. Data with a confidence level higher than a set threshold in the multi-source heterogeneous data are used as filtering data, and the filtering data are stored in the knowledge base in a networked mode using a graph database.

[0095] It should be noted that due to the different text data structures originating from user-side network equipment and the power grid system, the quality standards for data management vary across systems. The challenge in achieving a model that represents and stores user information from distribution network operation scenarios lies in how to encode large amounts of text and other data in a unified format. Traditional distribution network knowledge graphs, due to the domain-specific nature of their data, rely on a network composed of overhead lines, cables, towers, distribution transformers, disconnect switches, reactive power compensators, and some ancillary facilities, which plays a role in distributing electrical energy within the power grid. Throughout the entire construction process of a distribution network, related business activities—from planning, procurement, construction, operation, dispatching, maintenance, decommissioning, to scrapping—are managed by different professional departments. The corresponding information is managed and maintained in different information systems. The departments involved mainly include planning and construction, materials, production, dispatching, and technology. The main information includes basic equipment and facility information, financial asset information, operational data collection, maintenance and repair information, dispatching and operation information, and outage and scrapping information.

[0096] The data structure of user-side network equipment is quite different from that of the power grid system. It involves information such as network throughput and data flow, and therefore lacks standards and specifications for information technology construction and a unified and standardized basis.

[0097] Therefore, this embodiment uses a top-down knowledge graph construction technique to complete the construction of the ontology model and data layer. Information with high confidence is selected from the text and other data collected from the distribution network and user-side network equipment as filtering data and added to the knowledge base to describe the concepts of distribution network attributes and their interrelationships, as shown in Table 1. Then, the filtering data is stored in a networked mode using a graph database.

[0098] Table 1. Attributes of Traditional Power Systems

[0099]

[0100] S22. Using an NLP feature model, feature extraction is performed on the filtered data corresponding to the communication network side data and the filtered data corresponding to the power grid side data, respectively, to obtain the corresponding first attribute feature set and second attribute feature set.

[0101] It should be noted that this embodiment utilizes NLP-related knowledge to extract corresponding attribute features from different text data from user-side network equipment and power grid system using NLP feature models, and extracts the initial standardized features of users; thereby obtaining the user-side network equipment data attribute feature set, i.e., the first attribute feature set S(u), and the power grid data attribute feature set, i.e., the second attribute feature set S(e).

[0102] S23. Perform a feature concatenation at the attribute level between the first attribute feature set and the second attribute feature set to obtain the initial standardized reconstructed features.

[0103] It should be noted that in this embodiment, the first attribute feature set and the second attribute feature set are concatenated at the attribute level to obtain the initial standardized reconstruction feature x for the heterogeneous data of power grid data and user-side network equipment:

[0104] x = S (u) +S (e) .

[0105] In one embodiment, step S30: based on the initial normalized reconstructed features, performing attribute-level shallow fusion within the initial normalized reconstructed features to obtain attribute-level shallow features, includes the following steps:

[0106] A multilayer perceptron is used to perform attribute-level shallow fusion within the initial standardized reconstructed features to obtain attribute-level shallow features. The multilayer perceptron includes an input layer, a hidden layer, and an output layer. The hidden layer includes h hidden units, wherein:

[0107] The output of the hidden layer is:

[0108] H = XW h +b h

[0109] The output of the output layer is:

[0110] O = HW o +b o

[0111] In the formula: the weights and biases of the hidden layer are respectively represented by W. h ∈R n×h ,b h ∈R 1×h The output dimension of the hidden layer is q; the weights and bias parameters of the output layer are W, respectively. o ∈R n×q ,b o ∈R 1×q Initial standardized reconstructed feature set Where D is the feature dimension and N is its batch size.

[0112] It should be noted that, after obtaining the initial standardized reconstructed features x of heterogeneous data from power grid data and user-side network equipment, this embodiment further improves the robustness of heterogeneous data features through data fusion. In the first stage, a multilayer perceptron neural network is used to perform interactive processing within each user feature at the attribute level, such as... Figure 4 As shown.

[0113] Furthermore, to leverage the potential of the multi-layered architecture, a non-linear activation function σ is applied to each hidden unit after the affine transformation to enhance the model's expressive power. The output after the activation function is:

[0114] H=σ(XW h +b h )

[0115] In the formula: σ() represents the activation function.

[0116] Specifically, suppose the output of the l-th layer (l = 1…L) before passing through the activation function is z. l The activated form is represented as a. l σ represents the activation function. The output of the previous layer after activation becomes the input of this layer, and the output of this layer before activation is:

[0117] z l =W l a l-1 +b l

[0118] a l =σ(z) l )

[0119] In the formula: W l b l This represents the weights and biases of the l-th hidden layer; a l-1 This represents the representation of the (l-1)th hidden layer after the activation function.

[0120] It should be noted that this embodiment utilizes a fully connected perceptron network to perform feature fusion within a node, enabling global perception of all attribute dimensions within the node and more fine-grained features. Unlike the perception of local adjacent nodes at the node level, this embodiment uses a neural network to perform fine-grained perception of attribute-level features within the node.

[0121] In one embodiment, step S40: performing deep feature aggregation based on the attribute-level shallow features to obtain the user aggregated feature relationship matrix specifically includes the following steps:

[0122] S41. Perform graph attention deep encoding on the attribute-level shallow features to construct an image. As graphical data, each user feature vector in the attribute-level shallow features serves as a node in the image, and for each node v i For all Add a from v j Pointing to v i The edge e ji , For node vi The K nearest neighbors;

[0123] It should be noted that after performing attribute-level shallow fusion of the initial standardized reconstructed features x of the power grid data and heterogeneous data of user-side network equipment through neural networks, attribute-level shallow features O of user-level power grid data and heterogeneous data of user-side network equipment are obtained; now, graph attention deep encoding is performed on attribute-level shallow features O, and deep feature aggregation is performed at the user level.

[0124] like Figure 4 As shown, for a set of initial standardized reconstruction features O = {x1, ..., x}, n}, each user feature vector Where D is the feature dimension, and i = 1, 2, ..., N. These features can be viewed as a set of unordered nodes, represented as... For each node v i Determine node v i K nearest neighbors For all Add a from v j Pointing to v i The edge e ji We construct a graph G = (V, E), where E represents all edges.

[0125] S42. Based on the graph data, features from neighboring nodes are aggregated using a graph attention layer, and information is exchanged between the nodes to obtain the user aggregated feature relationship matrix, wherein the computational operations within the graph attention layer are as follows:

[0126]

[0127] In the formula: Aggregate() and Update() are respectively the features through W agg and W update These two matrices are multiplied to perform feature aggregation and feature update operations, W agg and W update Here, represents the learnable weights of a set of aggregation operations and the learnable weights of a set of update operations, respectively.

[0128] Furthermore, the feature aggregation operation is used to compute the representation of the current node by aggregating the features of neighboring nodes:

[0129]

[0130] In the formula: It is x i The set of adjacent nodes;

[0131]

[0132] h(·)=x′ i =x″ i W update

[0133] In the formula: GAT represents a graph attention network, designed to achieve its simplicity and efficiency;

[0134] The feature update operation is used to update h heads with different weights, each head being a pair of aggregated features x″. i After splitting, we get:

[0135]

[0136] In the formula: head 1 head 2 head h Each head represents a different head; These represent the weights corresponding to each head.

[0137] It should be noted that this embodiment constructs a knowledge graph under the condition of unknown relationships between nodes. Therefore, the fusion of node features and relationship weight matrix is ​​constructed through adaptive learning and is a learnable weight.

[0138] It should be noted that, firstly, the aggregated feature x″ i The algorithm splits the data into h heads: head 1, head 2, ..., head h, and then updates each head with different weights. All heads can be updated in parallel and concatenated to form the final value. This multi-head update operation allows the model to update information in multiple representation subspaces, which is beneficial for feature diversity. Simultaneously, by utilizing the similarity matrix in the graph attention encoding layer, a graph relationship aggregation network between electricity users can be obtained naturally and without increasing computational cost.

[0139] In the deep feature fusion stage, this embodiment uses graph attention mechanism encoding, which can not only better obtain user-level relevance, but also utilize its distribution matrix and the inherent correlation between feature attributes to construct a user network knowledge graph that combines user-side network equipment and power grid multi-source data without increasing the amount of computation.

[0140] It should also be noted that the data to be fused in this embodiment all come from different data sources. Therefore, in order to explore the direct correlation between these data, the problem of fusion between heterogeneous data must first be solved in order to further obtain correlation information to construct a knowledge graph. At the same time, data from user-side network devices is introduced, which increases the difficulty of heterogeneous data analysis. This embodiment proposes a two-stage data fusion strategy to achieve deep fusion of heterogeneous data from different sources, and further provides a foundation for knowledge graph construction.

[0141] The two-stage data fusion strategy involves first preprocessing multi-source heterogeneous data by attribute to extract primary attribute features, obtaining initial standard features for the multi-source data. Then, at the attribute level, a deep learning multilayer perceptron is used to perform attribute-level interactions within user features, learning the correlation between different data attributes. In the second stage, a graph attention mechanism is used to fuse the overall correlation information of user features at the user level after processing a large number of attribute-level features. Through the two-stage data feature fusion, a more refined interaction between multi-source heterogeneous data is achieved. At the same time, combined with the user aggregated feature relationship matrix encoded by the graph attention mechanism in the second stage, it can naturally and without increasing the computational load can be associated with traditional power system-related attributes, constructing a user-side network knowledge graph that combines user-side network equipment and power grid multi-source data.

[0142] In one embodiment, step S50: constructing a fine-grained knowledge graph based on the attribute features of the traditional power grid and the user aggregation feature relationship matrix, specifically includes the following steps:

[0143] S51. Based on the attribute characteristics of the traditional power grid, construct a knowledge graph of the traditional power grid;

[0144] S52. Based on the user aggregation feature relationship matrix and the traditional power grid knowledge graph, construct the relationship interconnection from the graph attributes of the traditional power grid to the user aggregation feature relationship matrix to obtain the fine-grained knowledge graph.

[0145] By constructing a knowledge graph, enabling knowledge graph queries to make associations like humans, it becomes crucial for enhancing model intelligence and autonomous fault diagnosis. This embodiment describes the process of constructing a distribution network information knowledge graph based on user-measured network equipment data as follows: Figure 5 As shown.

[0146] The main purpose of applying knowledge graphs is to assist in identifying existing problems or potential hazards in various equipment during the operation of a power distribution network, finding the causes of these hazards, and proposing corresponding measures and suggestions to improve the safety of power system operation. The construction of ontology rules is a crucial task in building a knowledge graph for power distribution network equipment. This construction includes steps such as defining concept categories and hierarchical structures, and defining concept attribute relationships. The ontology concept category division mainly involves classifying and defining equipment fault types. Based on their internal elements, they can be divided into the following categories: equipment, components, fault causes, and suggestions and measures. Defining concept attribute relationships allows for greater refinement of the ontology, resulting in a well-structured hierarchical classification system. Each fault category consists of equipment, components, fault causes, suggestions, and measures, all of which can be abstracted into entity and entity state descriptions. This leads to an accurately defined and comprehensive knowledge graph for power distribution network equipment.

[0147] This embodiment uses the user-side network equipment attributes provided by the operator, combined with the user aggregation feature relationship matrix obtained in the previous section, to obtain the user aggregation feature set U = {h1,…,h...}. n} and graph relation matrix User characteristics m represents the number of attributes a user contains, including attribute features from the traditional power grid and attribute features from user-side network devices.

[0148] Then, by leveraging the inherent relationships between knowledge graph attributes in existing traditional power grids and combining them with the user relevance matrix learned from graph attention, we can learn the implicit relationships between graph attributes of traditional power grids and other user attributes or network-side device attributes. This allows us to naturally and without increasing computational load obtain information about the relationships between data attributes, thereby achieving the interconnection of relationships between traditional power grids and user-side network-related device attributes.

[0149] like Figure 5 As shown, after learning the distribution of user nodes through graph attention, the distribution of relationships between nodes, combined with the inherent associations of the attributes between nodes, can be further associated with various attributes in traditional power grid equipment to form a fine-grained knowledge graph, rather than being limited to the coarse-grained associations between user nodes.

[0150] This embodiment combines the extracted entity information with relevant attributes of the traditional power system, thereby providing a reliable foundation for further analysis of user power aggregation characteristics, such as regional power outage sensing, regional power distribution equipment fault prediction, real-time disaster sensing, and auxiliary decision-making, enhancing the accuracy and timeliness of judgment in the traditional power grid system and providing dispatchers with effective fault handling reference data.

[0151] In addition, such as Figure 6 As shown in the second embodiment of the present invention, a two-stage fusion system for multi-source data is proposed, the system comprising:

[0152] Data acquisition module 10 is used to acquire multi-source heterogeneous data, including communication network side data and power grid side data;

[0153] Feature extraction module 20 is used to extract the initial standardized reconstruction features of the multi-source heterogeneous data;

[0154] The shallow fusion module 30 is used to perform attribute-level shallow fusion within the initial standardized reconstruction features based on the initial standardized reconstruction features to obtain attribute-level shallow features.

[0155] The deep fusion module 40 is used to perform deep feature aggregation based on the attribute-level shallow features to obtain a user aggregated feature relationship matrix.

[0156] The graph construction module 50 is used to construct a fine-grained knowledge graph based on the attribute characteristics of the traditional power grid and the user aggregation feature relationship matrix.

[0157] To fully utilize multi-source heterogeneous information, this invention proposes a two-stage data fusion method. It performs fine-grained fusion of features at both the shallow attribute level and the deep user level. Furthermore, through the correlation between data attributes, a fine-grained knowledge graph of the user network can be constructed without additional computational load. This provides a reliable foundation for further analysis of user power aggregation characteristics. It effectively integrates auxiliary data sources from any other system on top of traditional distribution network data, resulting in more efficient and accurate fused information, which is then provided to distribution network dispatchers for auxiliary decision-making and maintenance record-keeping.

[0158] In one embodiment, the communication network-side data includes heterogeneous data generated by user-side network equipment, and the power grid-side data includes heterogeneous data generated by the PMS system and heterogeneous data generated by the electricity consumption information collection system.

[0159] This embodiment provides simple and effective real-time information by adding the operator's real-time user measurement equipment data network as an auxiliary data source. It uses real-time network data stream feedback, combined with traditional power data research, to enhance the accuracy and timeliness of judgment in traditional power grid systems, and provides dispatchers with effective fault handling reference data.

[0160] In one embodiment, the system further includes:

[0161] The text preprocessing module is used to perform text data preprocessing on the multi-source heterogeneous data, perform structured type classification and data filtering on the multi-source heterogeneous data, and obtain preprocessed multi-source heterogeneous data.

[0162] Accordingly, the feature extraction module 20 is specifically used to: extract the initial standardized reconstruction features of the preprocessed multi-source heterogeneous data.

[0163] In one embodiment, the feature extraction module 20 includes:

[0164] The filtering unit is used to select data with a confidence level higher than a set threshold from the multi-source heterogeneous data as filtering data, and to store the filtering data in a knowledge base in a networked manner using a graph database.

[0165] The feature extraction unit is used to extract features from the filtered data corresponding to the communication network side data and the filtered data corresponding to the power grid side data using an NLP feature model, respectively, to obtain the corresponding first attribute feature set and second attribute feature set.

[0166] The feature concatenation unit is used to concatenate the first attribute feature set and the second attribute feature set at the attribute level to obtain the initial standardized reconstructed features.

[0167] In one embodiment, the shallow fusion module 30 is specifically used for:

[0168] A multilayer perceptron is used to perform attribute-level shallow fusion within the initial standardized reconstructed features to obtain attribute-level shallow features. The multilayer perceptron includes an input layer, a hidden layer, and an output layer. The hidden layer includes h hidden units, wherein:

[0169] The output of the hidden layer is:

[0170] H = XW h +b h

[0171] The output of the output layer is:

[0172] O = HW o +b o

[0173] In the formula: the weights and biases of the hidden layer are respectively represented by W. h ∈R n×h ,b h ∈R 1×h The output dimension of the hidden layer is q; the weights and bias parameters of the output layer are W, respectively. o ∈R n×q ,b o ∈R 1×q Initial standardized reconstructed feature set Where D is the feature dimension and N is its batch size.

[0174] Furthermore, to leverage the potential of the multi-layered architecture, a non-linear activation function σ is applied to each hidden unit after the affine transformation to enhance the model's expressive power. The output after the activation function is:

[0175] H=σ(XW h +b h )

[0176] In the formula: σ() represents the activation function.

[0177] In one embodiment, the deep fusion module 40 includes:

[0178] Graph construction units are used to perform graph attention deep encoding on the attribute-level shallow features to construct an image. As graphical data, each user feature vector in the attribute-level shallow features serves as a node in the image, and for each node v i For all Add a from v j Pointing to v i The edge e ji , For node v i The K nearest neighbors;

[0179] An aggregation unit is used to aggregate features from neighboring nodes based on the graph data using a graph attention layer, and exchange information between the nodes to obtain the user aggregated feature relationship matrix. The computational operations within the graph attention layer are as follows:

[0180]

[0181] In the formula: Aggregate() and Update() are respectively the features through W agg and W update These two matrices are multiplied to perform feature aggregation and feature update operations, W agg and W update Here, represents the learnable weights of a set of aggregation operations and the learnable weights of a set of update operations, respectively.

[0182] Furthermore, the feature aggregation operation is used to compute the representation of the current node by aggregating the features of neighboring nodes:

[0183]

[0184] In the formula: It is x i The set of adjacent nodes;

[0185]

[0186] h(·)=x′ i =x″ i W update

[0187] In the formula: GAT represents a graph attention network, designed to achieve its simplicity and efficiency;

[0188] The feature update operation is used to update h heads with different weights, each head being a pair of aggregated features x″. i After splitting, we get:

[0189]

[0190] In the formula: head 1 head2 head h Each head represents a different head; These represent the weights corresponding to each head.

[0191] The multi-head update operation in this embodiment allows the model to update information in multiple representation subspaces, which is beneficial for feature diversity. Simultaneously, by utilizing the similarity matrix in the graph attention encoding layer, a graph relationship aggregation network among electricity users can be obtained naturally and without increasing computational cost.

[0192] In the deep feature fusion stage, this embodiment uses graph attention mechanism encoding, which can not only better obtain user-level relevance, but also utilize its distribution matrix and the inherent correlation between feature attributes to construct a user network knowledge graph that combines user-side network equipment and power grid multi-source data without increasing the amount of computation.

[0193] In one embodiment, the map construction module 50 includes:

[0194] The first knowledge graph construction unit is used to construct a traditional power grid knowledge graph based on the attribute characteristics of the traditional power grid.

[0195] The fine-grained knowledge graph construction unit is used to construct the relational interconnection from the graph attributes of the traditional power grid to the user aggregated feature relation matrix based on the user aggregated feature relation matrix and the traditional power grid knowledge graph, thereby obtaining the fine-grained knowledge graph.

[0196] This embodiment leverages the inherent relationships between knowledge graph attributes in existing traditional power grids. By combining these relationships with the user relevance matrix learned from graph attention, it can learn implicit relationships between graph attributes of traditional power grids and other user attributes or network-side device attributes. This allows for the natural acquisition of information about relationships between data attributes without increasing computational load, thereby achieving the interconnection of relationships between traditional power grids and user-side network-related device attributes.

[0197] It should be noted that other embodiments or implementation methods of the two-stage fusion system for multi-source data described in this invention can refer to the above-described method embodiments, and will not be repeated here.

[0198] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0199] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0200] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0201] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0202] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A two-stage fusion method for multi-source data, characterized in that, The method includes the following steps: Acquire multi-source heterogeneous data, including communication network side data and power grid side data; Extracting initial standardized reconstruction features from the multi-source heterogeneous data includes: using data with a confidence level higher than a set threshold as filter data; storing the filter data in a knowledge base in a networked manner using a graph database; using an NLP feature model to extract features from the filter data corresponding to the communication network side data and the filter data corresponding to the power grid side data, respectively, to obtain corresponding first attribute feature sets and second attribute feature sets; and concatenating the first attribute feature sets and the second attribute feature sets at the attribute level to obtain the initial standardized reconstruction features. Based on the initial standardized reconstructed features, attribute-level shallow fusion is performed within the initial standardized reconstructed features to obtain attribute-level shallow features. This includes using a multilayer perceptron neural network to perform attribute-level shallow fusion within the initial standardized reconstructed features to obtain attribute-level shallow features. The multilayer perceptron neural network includes an input layer, a hidden layer, and an output layer. The hidden layer includes h hidden units, wherein: The output of the hidden layer is: The output of the output layer is: In the formula: the weights and biases of the hidden layer are respectively expressed as: The output dimension of the hidden layer is q; the weights and bias parameters of the output layer are respectively... Initial standardized reconstructed feature set , where D is the feature dimension and N is its batch size; Based on the attribute-level shallow features, deep feature aggregation is performed to obtain a user aggregated feature relationship matrix, including graph attention deep encoding of the attribute-level shallow features to construct an image. As graphical data, each user feature vector in the attribute-level shallow features serves as a node in the image. For each node... For all Add a from point to edge , For nodes The K nearest neighbors; Based on the graph data, features from neighboring nodes are aggregated using a graph attention layer, and information is exchanged between the nodes to obtain the user aggregated feature relationship matrix. The computational operations within the graph attention layer are as follows: : In the formula: and The features are respectively passed through and and Multiplying these two matrices performs feature aggregation and feature update operations. and The two sets of learning weights are a set of aggregation operations and a set of learning weights for update operations, respectively. A fine-grained knowledge graph is constructed based on the attribute characteristics of the traditional power grid and the relationship matrix of the user aggregation characteristics.

2. The two-stage fusion method for multi-source data as described in claim 1, characterized in that, The communication network side data includes heterogeneous data generated by user-side network equipment, and the power grid side data includes heterogeneous data generated by the PMS system and heterogeneous data generated by the electricity consumption information collection system.

3. The two-stage fusion method for multi-source data as described in claim 1, characterized in that, After acquiring multi-source heterogeneous data, including communication network-side data and power grid-side data, the method further includes: Text data preprocessing is performed on the multi-source heterogeneous data, and the multi-source heterogeneous data is structured type classification and data filtering are performed to obtain preprocessed multi-source heterogeneous data. The extraction of the initial standardized reconstruction features of the multi-source heterogeneous data specifically includes: Extract the initial standardized reconstruction features of the preprocessed multi-source heterogeneous data.

4. The two-stage fusion method for multi-source data as described in claim 1, characterized in that, Each of the hidden units is followed by a non-linear activation function, and the output of the activation function is: In the formula: This represents the activation function.

5. The two-stage fusion method for multi-source data as described in claim 1, characterized in that, The feature aggregation operation is used to calculate the representation of the current node by aggregating the features of neighboring nodes: In the formula: yes The set of adjacent nodes; In the formula: GAT represents a graph attention network; The feature update operation is used to update h heads with different weights, where each head is a set of aggregated features. After splitting, we get: In the formula: , , Each head represents a different head; , , These represent the weights corresponding to each head.

6. The two-stage fusion method for multi-source data as described in claim 1, characterized in that, The fine-grained knowledge graph is constructed based on the attribute features of the traditional power grid and the user aggregation feature relationship matrix, including: Based on the attribute characteristics of the traditional power grid, a knowledge graph of the traditional power grid is constructed; Based on the user aggregation feature relationship matrix and the traditional power grid knowledge graph, a relational interconnection is constructed from the graph attributes of the traditional power grid to the user aggregation feature relationship matrix to obtain the fine-grained knowledge graph.

7. A two-stage fusion system for multi-source data, characterized in that, The system for performing the method as described in any one of claims 1-6 includes: The data acquisition module is used to acquire multi-source heterogeneous data, including communication network side data and power grid side data; The feature extraction module is used to extract the initial standardized reconstruction features of the multi-source heterogeneous data; The shallow fusion module is used to perform attribute-level shallow fusion within the initial standardized reconstruction features based on the initial standardized reconstruction features, to obtain attribute-level shallow features. The deep fusion module is used to perform deep feature aggregation based on the attribute-level shallow features to obtain the user aggregated feature relationship matrix; The graph construction module is used to construct a fine-grained knowledge graph based on the attribute characteristics of the traditional power grid and the user aggregation feature relationship matrix.

Images