Power user attribute recognition method and device based on space-time interaction fusion

Abstract

The application provides a power user attribute recognition method and device based on space-time interaction fusion. The method comprises: obtaining a power load sequence and a topology adjacency graph of a target user; wherein the power load sequence comprises power load of each sub-power consumption period of the target user within a preset time period, and the topology adjacency graph comprises at least one of an electrical connection relationship between a user node and a distribution transformer device, a spatial adjacency relationship between the user node and a neighbor node, and a semantic similarity relationship between the user node and the neighbor node; inputting the power load sequence into a time sequence feature extraction model to obtain a power load feature vector of the target user; inputting the topology adjacency graph into a relationship graph attention network to obtain a topology feature vector of the target user; and inputting the power load feature vector and the topology feature vector into a preset inference model to obtain attribute information of the target user. The technical scheme provided by the application can improve the accuracy of power user attribute recognition.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method and apparatus for identifying power user attributes based on spatiotemporal interactive fusion. Background Technology

[0002] With the continuous advancement of smart grid construction, the power system has accumulated massive amounts of user electricity consumption data. Accurately identifying the attributes of electricity users is a core foundation for distribution network operations such as load forecasting, demand response, line loss management, and fault repair. Existing methods for identifying electricity user attributes mostly focus on user electricity load time-series data. Through traditional machine learning or deep learning models, user time-series load data and basic attribute features are fused using vector concatenation and then input into a classification model for attribute prediction.

[0003] However, this simple vector concatenation method integrates multi-source features, ignoring the inherent electrical connections and spatial clustering effects among power users, resulting in low information utilization and incomplete feature representation, thereby reducing the accuracy of power user attribute identification. Summary of the Invention

[0004] The present invention provides a method, apparatus, electronic device, storage medium, and computer program product for identifying power user attributes based on spatiotemporal interactive fusion, which can improve the accuracy of power user attribute identification to a certain extent.

[0005] In a first aspect, the present invention provides a method for identifying power user attributes based on spatiotemporal interactive fusion, the method comprising:

[0006] Obtain the target user's electricity load sequence and topological adjacency graph; wherein, the electricity load sequence includes the target user's electricity load for each sub-electricity consumption period within a preset time period, and the topological adjacency graph includes at least one of the following: the electrical connection relationship between the user node and the distribution transformer equipment, the spatial adjacency relationship between the user node and its neighboring nodes, and the semantic similarity relationship between the user node and its neighboring nodes;

[0007] The electricity load sequence is input into the time-series feature extraction model to obtain the electricity load feature vector of the target user;

[0008] The topological adjacency graph is input into a relational graph attention network to obtain the topological feature vector of the target user;

[0009] The electrical load feature vector and the topology feature vector are input into a preset inference model to obtain the attribute information of the target user.

[0010] In one embodiment of the present invention, obtaining the electricity load sequence and topological adjacency graph of the target user includes:

[0011] Obtain the electricity load sequence of the target user, as well as the distribution transformers, points of interest, and user names within the target area where the target user is located;

[0012] Based on the electrical adjacency relationships of the distribution transformer equipment, an electrical topology diagram is constructed;

[0013] Based on the locational relationship between the target user and each of the points of interest, a geographic topology map is constructed;

[0014] A semantic relationship graph is constructed based on the semantic similarity between the name of the target user and the names of other users within the target area;

[0015] The electrical topology graph, the geographical topology graph, and the semantic relationship graph are merged into a topological adjacency graph.

[0016] In one embodiment of the present invention, constructing a geographic topology map based on the locational relationship between the target user and each of the points of interest includes:

[0017] Based on the positional relationship between the target user node and each point of interest, a Veno graph is constructed;

[0018] Each point of interest is traversed sequentially. If the target user node and the current point of interest have a common boundary in the Venn diagram, a spatial adjacency edge is established between the target user node and the current point of interest.

[0019] The weights of the spatial adjacent edges are determined based on the Euclidean distance between the target user node and the current point of interest to obtain a geographic topology map.

[0020] In one embodiment of the present invention, a semantic relationship graph is constructed based on the semantic similarity between the name of the target user and the names of other users within the target area, including:

[0021] Using the BERT model, target word vectors of the target user's name and neighbor word vectors of other users' names within the target region are extracted respectively;

[0022] For each neighbor word vector, calculate the cosine similarity between the target word vector and the neighbor word vectors;

[0023] When the cosine similarity is greater than or equal to the preset similarity, semantic similarity edges are constructed between the target user node and the neighboring nodes corresponding to the neighboring word vectors; wherein, multiple semantic similarity edges constitute a semantic relationship graph.

[0024] In one embodiment of the present invention, the electricity load sequence is input into a time-series feature extraction model to obtain the electricity load feature vector of the target user, including:

[0025] The electricity load sequence is input into a multi-branch time-series feature extraction model to obtain multiple branch feature vectors; wherein, different branches correspond to different expansion rates.

[0026] Multiple branch feature vectors are concatenated to obtain multi-scale feature vectors;

[0027] Based on the target time attention weight, attention is added to the multi-scale feature vector to obtain the target user's electricity load feature vector.

[0028] In one embodiment of the present invention, before the step of adding attention to the multi-scale feature vector based on the target time attention weight to obtain the target user's electricity load feature vector, the method further includes:

[0029] Based on the learnable weight matrix, the multi-scale feature vector is mapped to the attention latent space to obtain the attention intermediate feature vector;

[0030] The transpose of the attention scoring vector is multiplied by the attention intermediate feature vector to obtain the temporal attention weight vector;

[0031] Based on the time attention weight vector, a target time attention weight vector is determined; wherein the sum of each target time attention weight in the target time attention weight vector is 1.

[0032] In one embodiment of the present invention, the topological adjacency graph is input into a relational graph attention network to obtain the topological feature vector of the target user, including:

[0033] Based on each relationship type in the topological adjacency graph, linear transformations are performed on the current layer input features of the target user's target node and the current layer input features of all neighboring nodes connected to the target node through the relationship type to obtain the projection features of the target node and each neighboring node under the relationship type.

[0034] The projected features of the target node and the projected features of the corresponding neighbor node are concatenated into vectors. The learnable attention vector corresponding to the relation type is used to perform a linear mapping on the concatenated feature vector. Then, the LeakyReLU activation function is applied to obtain the attention score of the corresponding neighbor node relative to the target node under the relation.

[0035] The attention scores are normalized to obtain the normalized attention weights of each neighbor node under the relationship type.

[0036] Using the corresponding normalized attention weight as coefficient, the projected features of all neighbor nodes under the relation type are weighted and summed to obtain the single-relation neighbor aggregation features of the target node under the relation type.

[0037] The single-relation neighbor aggregation features corresponding to all relation types are summed to obtain the full-relation neighbor aggregation features; the self-connection projection matrix is ​​used to perform a linear transformation on the current layer input features of the target node to obtain the self-connection features; the full-relation neighbor aggregation features are added to the self-connection features to obtain the fused features.

[0038] The fused features are input into a nonlinear activation function for nonlinear transformation, and the updated node features of the target node in the current layer are output. At least two layers of single-layer graph feature aggregation are stacked, with the output of each layer serving as the input of the next layer. The node features output from the last layer are used as the topology feature vector of the target power user.

[0039] In one embodiment of the present invention, the electrical load feature vector and the topology feature vector are input into a preset inference model to obtain the attribute information of the target user, including:

[0040] The electrical load feature vector is input into the first linear transformation layer and linearly mapped to obtain the target time-series feature;

[0041] The topological feature vector is input into a second linear transformation layer for linear mapping to obtain the target topological feature; wherein the target temporal feature has the same dimension as the target topological feature;

[0042] The target temporal feature, the target topological feature, the element-wise interaction features of the target temporal feature and the target topological feature, and the absolute value features of the element-wise difference between the target temporal feature and the target topological feature are concatenated to obtain a fused feature vector;

[0043] The fused feature vector is input into a gated fully connected layer, and the gate weight coefficients are calculated using the Sigmoid activation function.

[0044] Using the gate weight coefficient as the weighting coefficient of the target temporal feature and the difference between 1 and the gate weight coefficient as the weighting coefficient of the target topological feature, the target temporal feature and the target topological feature are weighted and summed to obtain the spatiotemporal joint fusion feature of the target user;

[0045] The spatiotemporal joint fusion features are input into the classification fully connected layer to obtain the attribute information of the target user.

[0046] In one embodiment of the present invention, the method further includes:

[0047] The power load feature vectors of each power user are linearly mapped through the first linear transformation layer, and the topology feature vectors of each power user are linearly mapped through the second linear transformation layer. The mapped power load feature vectors and topology feature vectors are then normalized to obtain the normalized time series representation vectors and normalized topology representation vectors of each user.

[0048] Positive sample pairs are formed by the normalized temporal representation vector and the normalized topological representation vector of the same user, and negative sample pairs are formed by the normalized temporal representation vector of the current user and the normalized topological representation vectors of all other users. Based on cosine similarity and a preset temperature adjustment coefficient, the cross-view contrast loss is calculated using the information noise contrast estimation loss function. The temperature adjustment coefficient is used to adjust the granularity of the distinction between positive and negative samples.

[0049] Based on the degree of connection between user nodes and distribution transformers in the aforementioned topology adjacency graph, a topology consistency loss term is constructed;

[0050] The cross-view comparison loss and the topology consistency loss term are weighted and summed to obtain the joint total loss of the pre-trained optimization, where the weighting coefficients corresponding to the topology consistency loss term are the topology constraint weight coefficients.

[0051] Using the joint total loss as the optimization objective, the network parameters of the temporal feature extraction model, the relational graph attention network, and the first and second linear transformation layers are updated synchronously through the backpropagation algorithm. The training is iterated until the loss converges, thus completing the pre-training of the spatiotemporal feature encoder.

[0052] In one embodiment of the present invention, the method further includes:

[0053] When a fault is detected in the power equipment node of the target user, the power equipment node of the target user is used as the starting point for retrieval. A topological traversal search is performed based on the electrical connection edge between the user node and the distribution transformer equipment to filter out all power user nodes that have a direct or indirect electrical connection relationship with the power equipment node, thus forming a set of users affected by the power outage.

[0054] Based on the spatial coverage area corresponding to the set of users affected by the power outage, an extended search is performed along the spatial adjacency edges between the power user nodes and their neighboring nodes to filter out sensitive interest nodes within the affected spatial range.

[0055] By integrating the user attribute information of the set of users affected by the power outage with the location attribute information of the sensitive interest nodes, a comprehensive power outage impact analysis report is generated, including statistics on affected users and a list of sensitive locations.

[0056] Secondly, the present invention provides a power user attribute identification device based on spatiotemporal interactive fusion, the device comprising:

[0057] The data acquisition module is used to acquire the electricity load sequence and topology adjacency graph of the target user; wherein, the electricity load sequence includes the electricity load of the target user in each sub-electricity consumption period within a preset time period, and the topology adjacency graph includes at least one of the following: the electrical connection relationship between the user node and the distribution transformer equipment, the spatial adjacency relationship between the user node and the neighboring node, and the semantic similarity relationship between the user node and the neighboring node;

[0058] The time-series feature extraction module is used to input the electricity load sequence into the time-series feature extraction model to obtain the electricity load feature vector of the target user;

[0059] The topology feature extraction module is used to input the topology adjacency graph into the relation graph attention network to obtain the topology feature vector of the target user;

[0060] The attribute recognition module is used to input the power load feature vector and the topology feature vector into a preset inference model to obtain the attribute information of the target user.

[0061] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the power user attribute identification method based on spatiotemporal interactive fusion as described above.

[0062] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the power user attribute identification method based on spatiotemporal interactive fusion as described above.

[0063] Fifthly, the present invention provides a computer program product, including a computer program, which, when executed by a processor, causes the computer to perform the power user attribute identification method based on spatiotemporal interactive fusion as described above.

[0064] This invention provides a method for identifying electricity user attributes based on spatiotemporal interactive fusion. By collaboratively fusing temporal feature data of electricity load sequences and diverse heterogeneous topological feature data, it integrates a complete feature chain from user's own electricity consumption behavior to heterogeneous graph multi-relationship neighborhood information. The temporal feature extraction model provides multi-scale electricity consumption pattern representation, while the relational graph attention network provides topological structure representation under multiple relationships. By incorporating both into a pre-defined inference model, the attribute information of the target user is obtained. Even in complex scenarios such as mixed commercial and residential areas, it can output relatively accurate user attribute identification results, providing reliable technical support for improving the intelligent operation and maintenance level of the power grid. Attached Figure Description

[0065] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0066] Figure 1 This is a flowchart illustrating a method for identifying power user attributes based on spatiotemporal interactive fusion, as provided in an embodiment of the present invention.

[0067] Figure 2 This is an architecture diagram of a power user attribute identification system based on spatiotemporal interaction fusion, provided as an embodiment of the present invention.

[0068] Figure 3 This is a schematic diagram of the structure of a power user attribute identification device based on spatiotemporal interactive fusion provided in an embodiment of the present invention.

[0069] Figure 4 This is a schematic diagram of an electronic device provided according to an embodiment of the present invention. Detailed Implementation

[0070] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0071] Currently, power user attribute identification mainly employs methods based on deep learning and traditional GIS-assisted analysis. Specifically, these include:

[0072] Vector concatenation fusion method: The user's time-series load data and basic attribute features are fused by vector concatenation and then input into the classification model for attribute prediction.

[0073] Traditional supervised learning models rely on a large amount of labeled data for training and use models such as support vector machines, decision trees, and convolutional neural networks for user classification.

[0074] GIS Fixed Radius Buffer Analysis: Based on the coordinates of users and geographic points of interest, fixed radius buffer analysis is used to achieve spatial correlation and assist in user attribute judgment.

[0075] However, existing methods typically fuse multi-source features using simple vector concatenation, neglecting the inherent electrical connections and spatial clustering effects among power users, resulting in low information utilization and incomplete feature representation. Furthermore, special user labels in the power system rely on manual annotation, leading to small sample sizes and historical labeling errors. Traditional supervised learning models are prone to overfitting under small sample sizes or noisy labels, exhibiting poor generalization ability. Additionally, traditional GIS analysis relies solely on geometric distance for spatial association, failing to model the nonlinear mapping between spatial semantics and electricity consumption behavior in complex geographical environments, thus affecting identification accuracy. Finally, existing deep learning models often lack physical constraints, making it difficult to correlate outputs with actual power grid operation patterns, hindering understanding and decision support for business personnel.

[0076] Based on this, the present invention provides a power user attribute identification method based on spatiotemporal interactive fusion. By synergistically fusing temporal feature data of electricity load sequences and multi-dimensional heterogeneous topological feature data, it integrates a complete feature chain from the user's own electricity consumption behavior to heterogeneous graph multi-relationship neighborhood information. The temporal feature extraction model provides multi-scale electricity consumption pattern representation, and the relational graph attention network provides topological structure representation under multiple relationships. By incorporating both into a preset inference model, the attribute information of the target user is obtained. Even in complex scenarios such as mixed commercial and residential areas, it can output relatively accurate user attribute identification results, providing reliable technical support for improving the intelligent operation and maintenance level of the power grid.

[0077] Please see Figure 1 and Figure 2 The embodiments of this invention provide a method for identifying power user attributes based on spatiotemporal interactive fusion. This method can be used to automatically identify user attribute categories in a power system.

[0078] Specifically, the power user attribute identification method based on spatiotemporal interaction fusion can be applied to power Internet of Things (IoT) sensing and analysis systems, such as power marketing system user profile construction, distribution network refined operation and maintenance, electricity consumption behavior analysis, and power outage impact assessment. The embodiments in this application are not specifically limited herein. The method may include the following steps.

[0079] Step S110: Obtain the power load sequence and topology adjacency graph of the target user; wherein, the power load sequence includes the power load of the target user in each sub-power consumption period within a preset time period, and the topology adjacency graph includes at least one of the following: the electrical connection relationship between the user node and the distribution transformer equipment, the spatial adjacency relationship between the user node and the neighboring node, and the semantic similarity relationship between the user node and the neighboring node.

[0080] In this embodiment, the electricity load sequence is time-series data used to characterize the electricity consumption behavior patterns of the target user over time. Specifically, the electricity load sequence of the target user can be represented as follows: ,in This represents the power load value at the t-th sampling time, where T is the number of sub-power consumption periods within the preset time period.

[0081] Electricity load sequences can be obtained from electricity marketing systems or electricity information collection systems. The sampling interval can be 15 minutes, 30 minutes, or 1 hour, and the corresponding preset time period can be several days, weeks, or months. The electricity load sequence records users' electricity intensity, electricity fluctuation, electricity periodicity, and other behavioral characteristics, providing a data foundation for subsequent time series feature extraction.

[0082] In this embodiment, the topological adjacency graph is a heterogeneous graph structure used to simultaneously characterize the multi-dimensional relationships between target users and multiple types of entities. The topological adjacency graph contains at least three types of connections: electrical connections, which describe the physical connection between user nodes and distribution transformers based on the power grid topology, reflecting the user's electrical affiliation in the graph; spatial adjacency, which describes the proximity between user nodes and neighboring nodes such as geographic points of interest, reflecting the user's geographic spatial clustering characteristics in the graph; and semantic similarity, which describes the semantic association between user nodes and other user nodes based on name text, reflecting the user's potential business connections in the graph.

[0083] Topological adjacency graphs extend traditional single-relationship graphs into multi-relationship heterogeneous networks, enabling user attribute identification to utilize complementary information from three dimensions: electrical, spatial, and semantic, providing structured input for subsequent topological feature aggregation.

[0084] Step S120: Input the electricity load sequence into the time series feature extraction model to obtain the electricity load feature vector of the target user.

[0085] In this embodiment, the temporal feature extraction model is a computational model used to automatically extract multi-scale temporal features from electricity load sequences. Specifically, the temporal feature extraction model adopts a structure based on a multi-scale attention temporal convolutional network, capturing electricity consumption patterns at different time spans in the electricity load sequence through multiple parallel convolutional branches. Compared to traditional recurrent neural networks, the temporal feature extraction model can significantly improve the processing efficiency of long sequence data by utilizing the parallel computation characteristics of convolution. (Electricity load feature vector) It is a high-dimensional feature representation of the electricity consumption behavior pattern of target users, which can simultaneously characterize short-term load changes, intraday electricity consumption cycle changes, and long-term electricity consumption trends across days, providing a time-series dimension for subsequent user attribute identification.

[0086] Step S130: Input the topological adjacency graph into the relational graph attention network to obtain the topological feature vector of the target user.

[0087] In this embodiment, the relational graph attention network is a graph neural network model used for adaptive aggregation of neighbor features for different types of connections in heterogeneous graphs. Unlike traditional graph attention networks that use a uniform attention mechanism for all edges, the relational graph attention network can learn independent linear transformation matrices and attention weights for electrical connections, spatial adjacencies, and semantic similarities in the topological adjacency graph, thereby automatically distinguishing the degree of differentiated influence of different types of neighbor nodes on the target user attributes.

[0088] The topological feature vector output by the relational graph attention network is generated through single-layer or multi-layer neighbor feature aggregation and node feature updating. It integrates the multi-relationship neighborhood information of target users in heterogeneous graphs, and can characterize the load association characteristics reflected by electrical connections, the geographical clustering characteristics reflected by spatial adjacency, and the business association characteristics reflected by semantic similarity, providing a topological dimension for subsequent user attribute identification.

[0089] Step S140: Input the electrical load feature vector and the topology feature vector into a preset inference model to obtain the attribute information of the target user.

[0090] In this embodiment, the preset inference model is an inference module used to adaptively fuse temporal features and topological features and output attribute classification results. The electricity load feature vector carries information about the target user's own electricity consumption behavior pattern, while the topological feature vector carries information about the target user's multi-dimensional neighborhood relationships in the heterogeneous graph. The two are complementary in the attribute identification task.

[0091] The pre-defined inference model nonlinearly fuses feature vectors from two dimensions, comprehensively utilizing electricity consumption behavior patterns and spatial topological relationships to output attribute labels for target users. Attribute information can be user categories of interest to the electricity business, such as residential, general industrial and commercial users, shopping mall users, hospital users, or high-loss users.

[0092] This approach addresses the shortcomings of existing technologies in electricity user attribute identification, such as shallow multi-source data fusion, lack of spatiotemporal interaction modeling, and insufficient spatial semantic mining. In practical applications, attribute identification does not rely solely on morphological analysis of electricity load curves or single geospatial analysis. Instead, it integrates a complete feature chain, from user's own electricity consumption behavior to heterogeneous graph multi-relationship neighborhood information, through the synergistic fusion of electricity load sequence features and topological features of the topological adjacency graph. The temporal feature extraction model provides multi-scale electricity consumption pattern representation, the relational graph attention network provides topological structure representation under multiple relationships, and the pre-defined inference model adaptively balances the contribution of the two types of features to the final attribute determination.

[0093] By incorporating both into a unified reasoning framework, even in complex scenarios such as mixed-use commercial and residential areas, relatively accurate user attribute identification results can be output, providing reliable technical support for improving the level of intelligent operation and maintenance of the power grid.

[0094] In some embodiments, obtaining the target user's electricity load sequence and topological adjacency graph in step S110 may include the following steps.

[0095] Step S111: Obtain the electricity load sequence of the target user, as well as the distribution transformers, points of interest, and user names within the target area where the target user is located.

[0096] Step S112: Construct an electrical topology diagram based on the electrical adjacency relationships of the distribution transformer equipment.

[0097] Step S113: Construct a geographic topology map based on the locational relationship between the target user and each of the points of interest.

[0098] Step S114: Construct a semantic relationship graph based on the semantic similarity between the name of the target user and the names of other users in the target area.

[0099] Step S115: Merge the electrical topology graph, the geographical topology graph, and the semantic relationship graph into a topological adjacency graph.

[0100] In this embodiment, the target area is the geographical area and power grid management unit where the target user is located, specifically it can be a transformer substation, feeder power supply area, or administrative region. In step S111, the system obtains the basic elements required to construct the heterogeneous map from multi-source data. The electricity load sequence is obtained from the power marketing system or electricity information collection system, reflecting the historical electricity consumption behavior of the target user. The distribution transformer equipment refers to the distribution transformers and their associated lines and transformer substation information that are electrically connected to the user within the target area, which comes from the distribution network GIS system and is used to construct the electrical connection relationships.

[0101] Points of interest (POIs) are locations within a target area that possess specific geographic semantic attributes, such as schools, hospitals, shopping malls, industrial parks, or government agencies. They are derived from external GIS or POI databases and are used to construct spatial proximity relationships. Usernames are derived from user profiles in the marketing system and mainly include text fields such as company names, names of government agencies and public institutions, shop names, school / hospital names, or park / unit names. They are used to construct semantic similarity relationships.

[0102] In this embodiment, the electrical topology diagram is a graph structure used to represent the physical connection relationships between users and distribution transformers based on the power grid topology. Specifically, the electrical topology diagram uses user nodes and distribution transformer nodes as vertices, and establishes electrical connection edges based on the power grid hierarchy of "user-distribution transformer-line". If a target user is connected to a distribution transformer through a transformer substation, then there is an electrical connection edge between the user node and the distribution transformer node. The electrical connection edge reflects the user's electrical affiliation in the graph, enabling users under the same distribution transformer or the same line to be interconnected through electrical paths.

[0103] In this embodiment, the geographic topology graph is a graph structure used to represent the spatial proximity relationship between user nodes and geographic points of interest. Specifically, the geographic topology graph uses user nodes and point of interest nodes as vertices, and establishes spatial adjacency edges based on their spatial positions. The construction method is not simply to use a fixed-radius buffer, but rather to spatially partition the coordinates of users and points of interest using computational geometry methods, treating nodes with common boundaries as spatial neighbors, thereby truly reflecting geographic proximity.

[0104] In this embodiment, the semantic relationship graph is a graph structure used to represent the semantic associations between user nodes based on their name texts. Specifically, the semantic relationship graph uses each user node as a vertex. By extracting the semantic vectors of the user name texts and calculating the similarity between the vectors, a semantic similarity edge is established between two user nodes when the similarity is higher than a preset threshold. The semantic similarity edge is used to connect users who may be in the same group, chain institutions, or have similar business attributes, thereby introducing business semantic dimension association information into the heterogeneous graph.

[0105] In this embodiment, the topological adjacency graph includes three edge types: electrical connection edges, spatial adjacency edges, and semantically similar edges, as well as three node types: user nodes, distribution transformer nodes, and points of interest nodes. By merging these three types of graphs, the topological adjacency graph can uniformly carry complementary information from three dimensions: electrical topology, geospatial data, and business semantics, providing structured input for the adaptive neighbor aggregation of the subsequent relational graph attention network.

[0106] Specifically, Represents an electrical topology view; Represents a GIS spatial adjacency view; These represent semantically similar views. These three views can be generated separately during the construction phase, but will ultimately be merged into a single heterogeneous graph. .

[0107] In this system, user nodes are the core shared nodes. The power network is responsible for describing the physical connections of users within the distribution network; the GIS network is responsible for describing the spatial relationships between users and their surrounding POIs and geographic neighbors; and the semantic network is responsible for describing the semantic similarity relationships between user names, industries, and organizational attributes. These three elements are not judged separately, but rather participate in RGAT aggregation together in the same heterogeneous graph using different edge types.

[0108] This approach addresses the issues of shallow multi-source data fusion and lack of spatiotemporal interaction modeling in existing technologies. In practical applications, user attribute identification does not solely rely on electricity load curve analysis or single geospatial analysis, but rather requires the comprehensive utilization of information from three dimensions: electrical affiliation, spatial clustering, and semantic association.

[0109] The electrical topology graph provides the electrical affiliation of users within the graph, the geographic topology graph provides the spatial clustering of users and points of interest, and the semantic relationship graph provides the business semantic associations between users. By merging these three into a unified heterogeneous graph, the relationship graph attention network can learn aggregation weights for different types of edges, adaptively capturing the differentiated impact of different neighbors on user attributes, thereby significantly improving the comprehensiveness and accuracy of user attribute identification.

[0110] In some embodiments, in step S113, constructing a geographic topology map based on the locational relationship between the target user and each of the points of interest may include the following steps.

[0111] Step S1131: Construct a Veno map based on the positional relationship between the target user node and each point of interest.

[0112] Step S1132: Iterate through each point of interest in sequence. If the target user node and the current point of interest have a common boundary in the Venn diagram, establish a spatial adjacency edge between the target user node and the current point of interest.

[0113] Step S1133: Determine the weight of the spatial adjacent edge based on the Euclidean distance between the target user node and the current point of interest to obtain a geographic topology map.

[0114] In this embodiment, the Venograph is a computational geometry used for adaptively partitioning planar space. Specifically, the Venograph is constructed based on the geographic coordinates of the target user node and various interest point nodes. For a given set of points, the Venograph divides the planar space into several polygonal units, each unit corresponding to a node. The distance from any point within a unit to that node is less than its distance to any other node. By constructing the Venograph, spatial adjacency relationships are no longer determined by a fixed-radius buffer, but rather adaptively determined by the geometric boundaries of the spheres of influence of each node.

[0115] In this embodiment, the common boundary is a line segment shared between two Vino units. If two nodes have a common boundary in the Vino graph, it indicates that their spheres of influence are directly adjacent in space and have a true geographical proximity relationship. In this case, a spatial adjacency edge is established between the target user node and the current point of interest node.

[0116] Compared to the fixed-radius buffer method used in traditional GIS analysis, the boundary construction method based on the public boundary of the Veno map can automatically adjust the spatial influence range according to the differences in urban and rural population density and building layout: in densely populated urban areas, the nodes are densely distributed, and the area of ​​the Veno unit naturally shrinks to avoid introducing irrelevant noise; in sparsely populated rural areas, the nodes are sparsely distributed, and the area of ​​the Veno unit naturally expands to avoid missing effective neighbors.

[0117] For each spatial adjacency edge established in step S1132, the Euclidean distance between the target user node and the corresponding point of interest node is calculated, and the reciprocal of the distance is used as the initial weight of the spatial adjacency edge. This weight is used to characterize the degree of spatial proximity between two nodes: the closer the distance, the greater the weight, indicating a stronger mutual influence between the two in geographic space. By assigning a weight based on the reciprocal of the distance to each spatial adjacency edge, the geographic topology map not only records whether spatial neighbor relationships exist between nodes, but also quantitatively characterizes the degree of closeness of spatial neighbor relationships.

[0118] The geographic topology map constructed through the above steps enables dynamic and adaptive division of spatial influence ranges. This approach can address the problems of insufficient spatial semantic mining, omission of neighbors in sparse areas, or introduction of noise in dense areas in traditional GIS fixed-radius buffer analysis.

[0119] In practical applications, the construction of spatial adjacency edges is no longer based on a one-size-fits-all distance threshold, but rather on an adaptive partitioning of the Venn diagram. This allows the geographic topology map to truly reflect the geographic interaction logic between users and points of interest, providing a more reasonable and accurate graph structure foundation for the subsequent learning of the influence weights of spatial neighbors on user attributes by the relational graph attention network.

[0120] In some embodiments, in step S114, constructing a semantic relationship graph based on the semantic similarity between the name of the target user and the names of other users in the target area may include the following steps.

[0121] Step S1141: Using the BERT model, extract the target word vector of the target user's name and the neighbor word vectors of the names of other users in the target region.

[0122] Step S1142: For each neighbor word vector, calculate the cosine similarity between the target word vector and the neighbor word vectors.

[0123] Step S1143: When the cosine similarity is greater than or equal to the preset similarity, construct semantic similarity edges between the target user node and the neighboring nodes corresponding to the neighboring word vectors; wherein, multiple semantic similarity edges constitute a semantic relationship graph.

[0124] In this embodiment, the BERT model is a pre-trained language model used to convert natural language text into fixed-dimensional semantic vectors. Specifically, through pre-training on a large-scale corpus, the BERT model can capture deep semantic information of text, making semantically similar texts closer together in the vector space. Usernames include text fields such as company names, government agency names, shop names, school / hospital names, or park / unit names, sourced from user profiles in the marketing system. Before being input into the BERT model, the system first performs text cleaning and standardization on the usernames, removing irrelevant characters and format differences. Then, the processed name text is input into the BERT model to extract the corresponding name semantic vectors. The target user's name, after being extracted by the BERT model, yields the target word vector. The names of other users within the target area are extracted by the BERT model to obtain their respective neighbor word vectors. .

[0125] In this embodiment, word vectors are high-dimensional feature representations of user name text in semantic space. Cosine similarity is a metric used to measure the directional consistency between two word vectors in semantic space, with a value ranging from [-1, 1]. The closer the value is to 1, the more consistent the directions of the two vectors, and the more semantically similar the corresponding user names. Specifically, the target word vector... Word vectors of neighbors The formula for calculating cosine similarity is: By calculating cosine similarity, it is possible to quantitatively assess the degree of association between a target user and other users at the semantic level of their names.

[0126] In this embodiment, the preset similarity is a threshold used to determine whether two user names have a significant semantic relationship. Specifically, when the cosine similarity between the target word vector and a neighboring word vector is greater than or equal to the preset similarity, it indicates that the two user names are highly similar semantically and may belong to the same group, chain organization, or have similar business attributes.

[0127] At this point, a semantically similar edge is established between the target user node and the neighbor node corresponding to the neighbor's word vector. This semantically similar edge connects potential semantically related users in the heterogeneous graph, and cosine similarity can be used as the initial weight of the edge. When multiple neighbor nodes within the target region are semantically highly similar to the target user's name, the target user node will have multiple semantically similar edges, which together constitute the semantic relationship graph.

[0128] The semantic relationship graph constructed through the above steps enables the heterogeneous graph to capture potential business relationships between users based on name semantics. This approach can solve the problem of traditional power user attribute identification relying solely on electrical and geographical information while neglecting user business semantic relationships.

[0129] In practical applications, multiple enterprise users under the same group or multiple store users under the same chain brand, although they may be independent in electrical topology and relatively dispersed in spatial distribution, share a high degree of semantic consistency in their names. By introducing semantically similar edges, the relational graph attention network can include the information of these semantically related users in the neighbor aggregation range during the subsequent topological feature aggregation process, thereby helping to identify user groups with similar business attributes and improving the comprehensiveness and accuracy of user attribute identification.

[0130] In some embodiments, step S120, inputting the electricity load sequence into a time-series feature extraction model to obtain the electricity load feature vector of the target user, may include the following steps.

[0131] Step S121: Input the electricity load sequence into the multi-branch time series feature extraction model to obtain multiple branch feature vectors; wherein, different branches correspond to different expansion rates.

[0132] Step S122: Concatenate multiple branch feature vectors to obtain multi-scale feature vectors.

[0133] Step S123: Based on the target time attention weight, add attention to the multi-scale feature vector to obtain the target user's electricity load feature vector.

[0134] In this embodiment, the multi-branch temporal feature extraction model is a computational structure used to extract electricity load sequence features in parallel from different time scales. Specifically, the temporal feature extraction model uses a temporal convolutional network as the basic unit, sets up multiple parallel convolutional branches, and each branch is configured with a different dilation rate. The dilation rate is a parameter used to control the receptive field span of the convolutional kernel; the larger the dilation rate, the longer the time span that the convolutional layer can cover in one operation.

[0135] For the electricity load sequence of target user u The system simultaneously inputs this data to each branch. For example, three parallel branches can be set up with expansion rates of d=1, d=2, and d=4, respectively. The branch with d=1 is used to capture short-term load mutation characteristics, such as instantaneous power jumps within a few sampling points; the branch with d=2 is used to capture intraday periodic variation characteristics, such as regular fluctuations in electricity consumption at different times of the day; and the branch with d=4 is used to capture long-term trend characteristics across daily or weekly scales, such as electricity consumption pattern cycles over multiple consecutive days. Each branch independently outputs a branch feature vector, representing the load behavior pattern at different time scales.

[0136] Through a multi-branch parallel structure, the temporal feature extraction model can analyze the same load curve from multiple time observation scales simultaneously, avoiding the limitation that a single convolution scale can only capture a certain time pattern.

[0137] In this embodiment, the multi-scale feature vector is a joint representation of load features across multiple time scales. Specifically, the branch feature vectors output from all branches in step S121 are concatenated along the feature dimension to form a multi-scale feature vector F_i. The concatenation operation preserves all scale information extracted from each branch, enabling the multi-scale feature vector to simultaneously encompass electricity consumption patterns at three levels: short-term fluctuations, medium-term cycles, and long-term trends.

[0138] In this embodiment, the target temporal attention weight is a weighted coefficient vector used to dynamically distinguish the contribution of different time steps to attribute identification. Specifically, the goal of the temporal attention mechanism is to automatically assign importance weights to each time step in the multi-scale feature vector.

[0139] For the t-th time step, the system maps the multi-scale features of that time step to the attention latent space based on the learnable weight matrix, then converts them into scalar scores via the attention scoring vector, and finally obtains the target temporal attention weights for that time step through normalization. ,satisfy The weights are ≥0 and the sum of the weights for all time steps is 1. A higher weight indicates that the model considers the load characteristics at that time step more important for determining user attributes. The system weights the target time attention weights for each time step with the corresponding multi-scale features to obtain the target user's electricity load feature vector. .

[0140] Through the above steps, the electricity load feature vector is no longer a simple encoding of the original load value, but a deep feature representation that integrates short-term, medium-term, and long-term multi-scale features and automatically filters key time step information through a time attention mechanism. This approach can solve the problem in existing technologies where it is difficult to process long sequences in parallel using a single time scale or recurrent neural networks.

[0141] In practical applications, electricity users with different attributes exhibit significant differences in their electricity consumption patterns: residential users show a bimodal load distribution (morning and evening peaks), shopping mall users' load is highly correlated with their business hours, and industrial users' load may have continuous cyclical characteristics. By using a multi-scale attention-based temporal convolutional network, the temporal feature extraction model can simultaneously capture these differentiated features across multiple time spans and automatically focus on the time segments with the highest discriminative power for attribute identification using an attention mechanism, thereby improving the representational ability of temporal features and the accuracy of subsequent attribute classification.

[0142] In some embodiments, prior to step S123, attention is added to the multi-scale feature vector based on the target time attention weight to obtain the target user's electricity load feature vector, and the following steps may also be included.

[0143] Step S12301: Based on the learnable weight matrix, map the multi-scale feature vector to the attention latent space to obtain the attention intermediate feature vector.

[0144] Step S12302: Multiply the transpose of the attention scoring vector with the attention intermediate feature vector to obtain the time attention weight vector.

[0145] Step S12303: Based on the time attention weight vector, determine the target time attention weight vector; wherein the sum of each target time attention weight in the target time attention weight vector is 1.

[0146] In this embodiment, the learnable weight matrix is ​​a linear transformation parameter matrix used to map multi-scale feature vectors to the attention latent space. Specifically, for the multi-scale feature vector of target user i at time step t... Learnable weight matrix Map it to attention intermediate feature vector ,in, Here, is the bias term, and tanh is the nonlinear activation function. The attention latent space is an intermediate representation space obtained after nonlinear transformation, in which the relative importance of features at different time steps is more easily perceived by the model automatically.

[0147] In this embodiment, the attention scoring vector is a learnable parameter vector used to compress the intermediate attention feature vector into a single importance score. Specifically, the attention scoring vector... Pay attention to the intermediate feature vector Perform a linear mapping to obtain the unnormalized importance score for that time step. This score is a scalar value; a larger value indicates that the model considers the load characteristics at time step t to be more critical for user attribute identification. The above operation is performed sequentially for all T time steps to obtain the time attention weight vector. , ,…, ].

[0148] In this embodiment, the target temporal attention weight vector is a normalized attention weight distribution. Specifically, a Softmax normalization operation is performed on each element of the temporal attention weight vector to obtain the target temporal attention weight vector. , ,…, ],in, After normalization, the attention weights for each objective time satisfy the following conditions: ≥0 and This allows attention weights to be interpreted as a probabilistic allocation of the importance of each time step by the model.

[0149] Through the above steps, the system completes the entire calculation process for time attention weights. The generation process of the target time attention weights is specifically defined as three sub-steps: "mapping-scoring-normalization". In practical applications, not all time steps in the target user's electricity load sequence contribute equally to attribute identification. For example, the load characteristics of shopping mall users during weekday and holiday business hours are significantly different from other times, and abnormal electricity consumption by high-loss users may only be concentrated in specific time periods.

[0150] Through the aforementioned attention weighting mechanism, the model can automatically identify and enhance the contribution of key time steps while suppressing interference from irrelevant or redundant time steps, thus improving the final generated electricity load feature vector. It is more discriminative and robust.

[0151] In some embodiments, step S130, inputting the topological adjacency graph into a relational graph attention network to obtain the topological feature vector of the target user, may include the following steps.

[0152] Step S131: Based on each relationship type in the topological adjacency graph, perform linear transformation on the current layer input features of the target user's target node and the current layer input features of all neighboring nodes connected to the target node through the relationship type to obtain the projection features of the target node and each neighboring node under the relationship type.

[0153] Step S132: Concatenate the projection features of the target node with the projection features of the corresponding neighbor node, perform linear mapping on the concatenated feature vector using the learnable attention vector corresponding to the relation type, and then process it with the LeakyReLU activation function to obtain the attention score of the corresponding neighbor node relative to the target node under the relation.

[0154] Step S133: Normalize the attention score to obtain the normalized attention weight of each neighbor node under the relationship type.

[0155] Step S134: Using the corresponding normalized attention weight as coefficient, perform a weighted summation of the projected features of all neighbor nodes under the relation type to obtain the single-relationship neighbor aggregation features of the target node under the relation type.

[0156] Step S135: Summate the single-relation neighbor aggregation features corresponding to all relation types to obtain the full-relation neighbor aggregation features; use the self-connection projection matrix to perform a linear transformation on the current layer input features of the target node to obtain the self-connection features; add the full-relation neighbor aggregation features and the self-connection features to obtain the fused features.

[0157] Step S136: Input the fused features into a nonlinear activation function for nonlinear transformation, and output the updated node features of the target node in the current layer; stack at least two layers of single-layer graph feature aggregation, with the output of each layer serving as the input of the next layer, and use the node features output by the last layer as the topology feature vector of the target power user.

[0158] In this embodiment, the relationship type is a classification identifier for edge types in the topological adjacency graph, specifically including three categories: electrical connection relationships, spatial adjacency relationships, and semantic similarity relationships. For each relationship type, the relationship graph attention network is configured with independent learnable parameters to achieve differentiated processing of neighbor features under different relationships.

[0159] In this embodiment, the projected features are linear transformations of the node input features under a specific relation type. Specifically, the relation graph attention network sets a corresponding relation transformation matrix for each relation type r. ,in The input node feature dimension, Let r be the spatial dimension of the topological embedding. For the target node u and its neighboring nodes v connected through relation r, the current layer input features are respectively... and The projected features are obtained after linear transformation by the relation transformation matrix. and .in, The initial characteristics of user nodes may include user profile characteristics, contract capacity, industry category, geographic coordinate encoding, or load statistics characteristics.

[0160] Electrical connectivity, spatial adjacency, and semantic similarity each have independent relationship transformation matrices, enabling the model to distinguish the meaning of neighbor node features under different relationships. For example, user neighbors on the same line are more likely to reflect electrical load associations, while spatially proximate points of interest are more likely to reflect commercial attributes or sensitive user attributes.

[0161] In this embodiment, the attention score is an unnormalized metric used to quantify the importance of neighboring nodes to the target node under a specific relationship type. Specifically, for relationship type r, the projected features of the target node are... Projection features of neighboring nodes Perform vector concatenation to obtain the concatenated vector [ || ].

[0162] Subsequently, the learnable attention vector corresponding to this relation type is used. A linear mapping is performed on the concatenated vector, followed by non-linear processing using the LeakyReLU activation function, to obtain the attention score of the neighbor node v relative to the target node u under this relationship. The LeakyReLU activation function outputs a small gradient instead of zero when the input is negative, which helps alleviate the vanishing gradient problem and makes the attention mechanism training more stable. Attention score. The larger the value, the more valuable the neighbor node v's attribute identification of the target node u is under this relationship type.

[0163] In this embodiment, the normalized attention weight is the probabilistic distribution of attention scores across all neighboring nodes under the same relation type. Specifically, for relation type r, the attention scores of all neighboring nodes obtained in step S132 are subjected to Softmax normalization to obtain the normalized attention weight of each neighboring node. ,in Let represent the set of all neighboring nodes connected to the target node u through relation type r. After normalization, the attention weights of all neighboring nodes under the same relation type satisfy . ≥0 and This allows attention weights to have an interpretable probabilistic meaning.

[0164] In this embodiment, the single-relationship neighbor aggregation feature is a weighted fusion representation of the target node's information across all neighbor nodes under a single relationship type. Specifically, it uses normalized attention weights. The coefficients are the projected features of all neighboring nodes under relation type r. By performing a weighted summation, we obtain the single-relation neighbor aggregation feature of the target node under this relation type. This aggregation feature integrates the contributions of all neighboring nodes to the target node under this relationship type, and the attention weight automatically highlights important neighbors and suppresses secondary neighbors.

[0165] In this embodiment, the full-relation neighbor aggregation feature is a comprehensive representation of the target node's neighbor information across all relation types. Specifically, the full-relation neighbor aggregation feature is obtained by summing the single-relation neighbor aggregation features corresponding to all relation types. Simultaneously, a self-connected projection matrix is ​​employed. Input features of the current layer of the target node Perform a linear transformation to obtain the self-connected features. Self-connect features are used to preserve the target node's own information, preventing the original features of the target node from being excessively diluted or lost during neighbor aggregation. The full-relationship neighbor aggregation features are added to the self-connect features to obtain the fused features.

[0166] In this embodiment, the fused features are input into a nonlinear activation function. Perform a nonlinear transformation and output the updated node features of the target node in the current layer. Nonlinear activation functions introduce nonlinear expressive power into the model, enabling the learning of more complex feature transformations after multiple layers are stacked.

[0167] Furthermore, at least two layers of the aforementioned single-layer graph feature aggregation are performed in the stack. Single-layer graph feature aggregation can only aggregate information of directly adjacent first-order neighbors. However, through multi-layer stacking, the output of each layer serves as the input of the next layer, allowing the target node to progressively aggregate information of second-order, third-order, and even higher-order neighbors, thereby capturing multi-hop relationships in the heterogeneous graph. For example, the first layer aggregates information of other users under the same transformer, while the second layer can further aggregate information of users under adjacent transformers, thus identifying electrical association patterns over a wider range. The node features output from the last layer are used as the topological feature vector of the target power user. .

[0168] Through the steps described above, the relational graph attention network completes the entire process of adaptively aggregating neighbor features for different types of connections on heterogeneous graphs. This approach addresses the problem in existing graph neural networks that use a uniform attention mechanism for all edges and cannot distinguish the differentiated impacts of different types of neighbors.

[0169] In practical applications, when determining the attribute of "shopping mall users," spatial neighbors, i.e., the interest points of densely distributed surrounding shops, should be given higher weight; while when determining "high-loss users," electrical neighbors, i.e., users with load-related connections on the same line, should be given higher weight. By setting independent relationship transformation matrices and attention parameters for different relationship types, the relational graph attention network can adaptively learn the differentiated influence weights of different types of neighbors on user attributes, thereby outputting more discriminative and business-interpretable topological feature vectors.

[0170] In some embodiments, step S140 may include the following steps: inputting the electrical load feature vector and the topology feature vector into a preset inference model to obtain the attribute information of the target user.

[0171] Step S141: Input the electrical load feature vector into the first linear transformation layer and perform linear mapping to obtain the target time-series feature.

[0172] Step S142: Input the topological feature vector into the second linear transformation layer for linear mapping to obtain the target topological feature; wherein, the target temporal feature has the same dimension as the target topological feature.

[0173] Step S143: Concatenate the target temporal feature, the target topological feature, the element-wise interaction features of the target temporal feature and the target topological feature, and the absolute value features of the element-wise difference between the target temporal feature and the target topological feature to obtain a fused feature vector.

[0174] Step S144: Input the fused feature vector into the gated fully connected layer, and calculate the gate weight coefficients using the Sigmoid activation function.

[0175] Step S145: Using the gating weight coefficient as the weighting coefficient of the target temporal feature and the difference between 1 and the gating weight coefficient as the weighting coefficient of the target topological feature, perform a weighted summation on the target temporal feature and the target topological feature to obtain the spatiotemporal joint fusion feature of the target user.

[0176] Step S146: Input the spatiotemporal joint fusion features into the classification fully connected layer to obtain the attribute information of the target user.

[0177] In this embodiment, the first linear transformation layer and the second linear transformation layer are linear projection layers used to map feature vectors of different dimensions to a unified dimensional space. Specifically, after pre-training, the encoder parameters of the temporal feature extraction model and the relational graph attention network are fixed. The first linear transformation layer uses the electrical load feature vector... Using the input as input, the target temporal features are obtained through linear mapping. The second linear transformation layer uses topological feature vectors. Using the input as input, the target topological features are obtained through linear mapping. ,in, , The weight matrix is ​​a learnable matrix. , This is a bias term.

[0178] After linear transformation, the target temporal features and target topological features are mapped to the same dimensional space, which gives the two feature vectors the basis for fusion and comparison in the same semantic space.

[0179] In this embodiment, the element-wise interaction feature is a cross-representation used to capture the cooperative relationship between the target temporal feature and the target topological feature. Specifically, the element-wise interaction feature is obtained by element-wise multiplication of the target temporal feature and the target topological feature, i.e. Element-wise multiplication can capture the consistency or co-occurrence patterns of two features along corresponding dimensions. The absolute value of the element-wise difference is a comparative representation used to measure the degree of difference between target temporal features and target topological features, i.e. The difference feature can reflect inconsistencies or abnormal signals between user electricity consumption behavior patterns and topological characteristics.

[0180] The four types of features are concatenated to obtain a fused feature vector. , where || denotes the vector concatenation operation. The fused feature vector integrates the independent information of temporal features, the independent information of topological features, the interaction information between the two, and the difference information between the two, providing rich discriminative basis for subsequent gating fusion.

[0181] In this embodiment, the gated fully connected layer is a computational layer used to adaptively learn the fusion ratio of temporal and topological features from the fused feature vector. The gate weight coefficients are adaptive coefficients used to control the respective contribution ratios of temporal and topological features in the final fusion.

[0182] Specifically, feature vector fusion The gating weights are calculated by inputting the gated fully connected layer and then using the Sigmoid activation function. ,in For learnable weight matrix, For bias terms, This is the Sigmoid activation function. The Sigmoid activation function constrains the values ​​of the gating weights to the interval (0,1). Gating weights The closer a value is to 1, the greater the contribution of temporal features to the identification of current user attributes, according to the model; the closer a value is to 0, the greater the contribution of topological features, according to the model.

[0183] In this embodiment, the spatiotemporal joint fusion feature is a comprehensive representation of the target user's temporal features and topological features after adaptive weighted fusion through a gating mechanism. Specifically, the system uses gating weight coefficients... The weighting coefficients of the target time series features, with 1- The weighted coefficients of the target topological features are used to sum the target temporal features and the target topological features in a weighted manner, thus obtaining the spatiotemporal joint fusion features. .

[0184] This gating weighting mechanism ensures that the fusion ratio is not a fixed value set manually, but rather adaptively determined by the model based on the specific characteristics of the input users. For residential users with strong regularity in electricity consumption patterns and relatively simple topological relationships, the model can automatically increase the weight of time-series features; for shopping mall users with significant spatial clustering effects and greater influence from the surrounding commercial environment, the model can automatically increase the weight of topological features.

[0185] In this embodiment, the classification fully connected layer is the output layer used to map the spatiotemporal joint fusion features to the user attribute category space. Specifically, it maps the spatiotemporal joint fusion features... The input is a fully connected layer for classification, which is then normalized using Softmax to obtain the probability distribution of the target user belonging to each attribute tag. The category with the highest probability is selected as the attribute information of the target user. The attribute information can be user categories that the power business is concerned with, such as residential users, general industrial and commercial users, shopping mall users, hospital users, or high-loss users.

[0186] Through the above steps, the pre-defined inference model completes the end-to-end inference process from temporal and topological dual-stream features to user attribute labels. This approach addresses the problem in existing technologies where simple splicing and fusion methods cannot adaptively balance the contributions of multi-source features.

[0187] In practical applications, target temporal features carry information about the user's own electricity consumption behavior patterns, while target topological features carry information about the user's multi-dimensional neighborhood relationships in heterogeneous graphs. The two are complementary in attribute identification tasks. Through multi-dimensional feature concatenation and gating fusion mechanisms, the pre-defined inference model can comprehensively utilize electricity consumption behavior patterns and spatial topological associations, and adaptively learn the optimal temporal-topological fusion ratio for different types of users. This allows it to output high-precision user attribute identification results even in complex scenarios such as mixed-use commercial and residential areas.

[0188] In some embodiments, prior to step S140, the method may further include the following steps.

[0189] Step S210: Linearly map the power load feature vectors of each power user through the first linear transformation layer, and linearly map the topology feature vectors of each power user through the second linear transformation layer; and normalize the mapped power load feature vectors and topology feature vectors respectively to obtain the normalized time series representation vectors and normalized topology representation vectors of each user.

[0190] Step S220: A positive sample pair is formed by using the normalized temporal representation vector and the normalized topological representation vector of the same user; a negative sample pair is formed by using the normalized temporal representation vector of the current user and the normalized topological representation vectors of all other users; based on cosine similarity and a preset temperature adjustment coefficient, the cross-view contrast loss is calculated using the information noise contrast estimation loss function; the temperature adjustment coefficient is used to adjust the granularity of the distinction between positive and negative samples.

[0191] Step S230: Based on the degree of connection between user nodes and distribution transformers in the topology adjacency graph, construct a topology consistency loss term.

[0192] Step S240: The cross-view comparison loss and the topology consistency loss term are weighted and summed to obtain the joint total loss of the pre-trained optimization, wherein the weighting coefficient corresponding to the topology consistency loss term is the topology constraint weight coefficient.

[0193] Step S250: Using the joint total loss as the optimization objective, the network parameters of the temporal feature extraction model, the relational graph attention network, and the first and second linear transformation layers are updated synchronously through the backpropagation algorithm. The training is iterated until the loss converges, thus completing the pre-training of the spatiotemporal feature encoder.

[0194] In this embodiment, the above steps constitute a graph contrast learning pre-training stage, which is performed before the task adaptive inference stage, and is used to learn robust and physically interpretable user feature representations without the need for user attribute labels.

[0195] In this embodiment, the normalized time-series representation vector and the normalized topology representation vector are standardized feature representations used to map the electricity load feature vector and the topology feature vector to a unified, comparable latent space. Specifically, the first linear transformation layer and the second linear transformation layer are the linear mapping layers in steps S141 and S142. For user u, its electricity load feature vector... After the first linear transformation layer After mapping, the projected temporal features are obtained. Its topological feature vector After the second linear transformation layer The projected topological features are obtained after mapping. .

[0196] Subsequently, the projected feature vectors were subjected to L2 normalization to obtain normalized time series representation vectors. and normalized topological representation vector Normalization unifies the feature vectors to a unit length, making subsequent cosine similarity calculations equivalent to vector dot products, thus facilitating the measurement of semantic distance between features on a uniform scale.

[0197] In this embodiment, the cross-view contrast loss is an optimization objective function used to narrow the distance between two view features of the same user and widen the distance between features of different users through self-supervised learning. Its core idea is that the electricity consumption behavior pattern and topological features of the same user are descriptions of the same entity from different perspectives, and their feature representations should be similar in the latent space; while the feature representations of different users should be far apart.

[0198] Specifically, a positive sample pair consists of the normalized temporal representation vector and the normalized topological representation vector of the same user. For a batch of training samples containing B users, the positive sample pairs are {( ),( ),…,( Negative sample pairs are represented by the normalized temporal representation vector of the current user u. Normalized topological representation vectors of all other users v in the batch (v≠u) constitutes the training sample set for contrastive learning. Positive sample pairs and negative sample pairs together constitute the training sample set for contrastive learning.

[0199] In this embodiment, cosine similarity is a metric used to measure the directional consistency of two normalized feature vectors in the latent space. Specifically, the cosine similarity between the normalized temporal representation vector of user u and the normalized topological representation vector of user v can be expressed as: Since the feature vectors have been normalized, the similarity value ranges from [-1, 1]. The closer the value is to 1, the more consistent the directions of the two feature vectors are.

[0200] In this embodiment, the temperature adjustment coefficient τ is a hyperparameter used to control the granularity of distinguishing between positive and negative samples. Specifically, the temperature adjustment coefficient τ is introduced into the Softmax normalization operation. When the value of τ is small, the similarity difference is amplified, the model more strictly distinguishes between positive and negative samples, and pays more attention to the difficult-to-distinguish negative samples; when the value of τ is large, the similarity distribution is smoother, the training process is more stable, but the distinguishing strength between positive and negative samples will be weakened.

[0201] Based on the above definition, the cross-view contrast loss is calculated using the information-noise contrast estimation loss function, specifically expressed as follows: The numerator is the similarity index of the current user's positive sample pairs, and the denominator is the sum of the similarity indices between the current user and all users in the batch, including the current user as a positive sample and other users as negative samples. Minimizing this loss function is equivalent to maximizing the mutual information between the two views of the same user, enabling the model to learn the intrinsic relationship between user electricity consumption behavior and topology under unlabeled conditions.

[0202] In this embodiment, the topology consistency loss term is a penalty term used to incorporate the connection patterns of the physical power grid as a constraint into model training. In a real power grid, users with close electrical connections, such as those connected to the same transformer or junction box, typically exhibit natural similarities in their electricity consumption attributes and load characteristics due to sharing the same power supply environment. If the feature representations learned by the model violate this physical law—that is, despite close electrical connections, there are significant differences in the representations in the feature space—it indicates that the model may be overfitting to noisy data or failing to capture effective physical priors.

[0203] Specifically, the formula for calculating the topology consistency loss term is as follows: .in, Let be the set of electrically connected edges in the topological adjacency graph. and These are the predicted probability distributions or feature representations of user nodes u and v, respectively. This loss term penalizes node pairs that are electrically connected but have large feature differences, forcing the model output to conform to the feature representations of the power grid's physical connections.

[0204] In this embodiment, the joint total loss is a weighted combination of the cross-view comparison loss and the topology consistency loss. Specifically, the formula for calculating the joint total loss is: .in, These are the topology constraint weight coefficients, used to control the contribution of topology consistency constraints to joint optimization. The topology constraint weight coefficients can be set by dividing a validation set on a small number of labeled samples, searching the candidate value set {0.01, 0.05, 0.1, 0.5, 1.0}, and selecting the value that best achieves the classification effect on the validation set as the final weight coefficient.

[0205] In this embodiment, the joint total loss L is used as the optimization objective. The parameters of the multi-scale temporal convolutional network of the temporal feature extraction model and the relational transformation matrix of the relational graph attention network are updated synchronously through the backpropagation algorithm. and attention parameters The parameters of the projection matrices of the first and second linear transformation layers are determined. Iterative training continues until the joint total loss converges, completing the pre-training of the spatiotemporal feature encoder. After pre-training, the encoder parameters are fixed, and only the mapping structures of the first and second linear transformation layers are retained for use in the subsequent adaptive inference stage.

[0206] Through the above steps, the graph contrastive learning pre-training module completes self-supervised joint training of the temporal feature encoder and the topological feature encoder without requiring user attribute labels. This approach can address the pain points of scarce labeled data and noise interference in power scenarios.

[0207] In practical applications, special user labels in power systems rely on manual annotation, resulting in small sample sizes and potential errors in historical annotations. Traditional supervised learning models are prone to overfitting under small sample sizes or noisy labels. By constructing a cross-view comparison task between temporal and topological views, the model can learn robust general node representations during the unsupervised pre-training stage.

[0208] Meanwhile, the topology consistency loss term utilizes the homogeneity of the physical power grid to constrain the output space of the neural network model. This not only serves as regularization to prevent overfitting but also makes the model's inference results physically interpretable. After pre-training, only a small number of labeled samples are needed to fine-tune the classifier, enabling high-precision user attribute identification even with low resource labels.

[0209] In some embodiments, after step S140, the identification of power user attributes based on spatiotemporal interaction fusion may further include the following steps.

[0210] Step S151: If a fault is detected in the power equipment node of the target user, a topological traversal search is performed based on the electrical connection edge between the user node and the distribution transformer equipment, using the power equipment node of the target user as the starting point. All power user nodes that have a direct or indirect electrical connection with the power equipment node are selected to form a set of users affected by the power outage.

[0211] Step S152: Based on the spatial coverage area corresponding to the set of users affected by the power outage, perform an extended search along the spatial adjacency edges between the power user nodes and their neighboring nodes to filter out sensitive interest nodes within the affected spatial range.

[0212] Step S153: Integrate the user attribute information of the set of users affected by the power outage with the location attribute information of the sensitive interest nodes to generate a comprehensive power outage impact analysis report including statistics of affected users and a list of sensitive locations.

[0213] In this embodiment, the above steps constitute the dynamic power outage assessment stage. This stage reuses the heterogeneous graph structure and user attribute identification results constructed in steps S110 to S140 to perform cascaded assessment of the power outage impact range when a power grid fault occurs.

[0214] In this embodiment, a power equipment node is a distribution transformer or line node that has an electrical connection with a user node in the topological adjacency graph. A fault refers to the operating state of the power equipment node where power supply is interrupted due to tripping, short circuit, or other abnormal reasons. When the system detects a fault in a power equipment node, it uses that node as the starting point and performs a topological traversal search along the electrical connection edges in the topological adjacency graph.

[0215] In this embodiment, the topology traversal search is a graph search algorithm that expands outwards layer by layer along the edges of the graph to visit neighboring nodes. It can be implemented using a breadth-first search or depth-first search strategy. Specifically, starting from the faulty equipment node, the algorithm sequentially visits the first-level power user nodes directly connected to it along the electrical connection edges. Then, it continues to visit upstream or downstream distribution transformer equipment nodes via the electrical connection edges of the user nodes, and so on, layer by layer, until all power user nodes that have a direct or indirect electrical connection relationship with the faulty equipment node have been traversed. These searched power user nodes together constitute the set of users affected by the power outage.

[0216] In this embodiment, the set of users affected by the power outage is the range of users affected by the fault determined at the electrical topology level. Compared with traditional methods that rely solely on spatial location to estimate the outage range, the topology traversal search based on the electrical connection edges of the heterogeneous graph can accurately track the propagation path of the fault state along the power grid topology, ensuring that no users connected downstream of the faulty line but spatially distant are missed, nor are users spatially adjacent but belonging to other independent lines misjudged.

[0217] In this embodiment, spatial expansion retrieval is a retrieval operation that further identifies sensitive locations within the spatial coverage area of ​​the set of users affected by the power outage. Sensitive interest nodes are POI nodes with specific social function attributes in the topological adjacency graph, such as hospitals, schools, government agencies, or large shopping malls.

[0218] Specifically, the system first determines the spatial range covered by the geographical coordinates of each power user node in the set of users affected by the power outage. Then, based on this spatial range, it performs an extended search along the spatial adjacency edges of each power user node in the topological adjacency graph to filter out sensitive interest nodes within the affected spatial range.

[0219] Spatial adjacency edges, based on the geographic topology map constructed in the preceding steps, record the spatial proximity relationships between user nodes and points of interest using the Vinograph common boundary and inverse distance weights. Spatial extended retrieval enables the system not only to identify users experiencing direct power outages but also to assess the secondary impacts of power outage events on surrounding sensitive locations.

[0220] In this embodiment, the comprehensive power outage impact analysis report is a structured output document that integrates power outage user attribute statistics and a list of sensitive locations. Specifically, it integrates the attribute information of each user in the set of users affected by the power outage, such as category tags like residents, general industrial and commercial, hospitals, and schools, as well as the location attribute information of sensitive interest nodes, such as location name, location type, and address, to generate the comprehensive power outage impact analysis report. This report includes at least the statistics of affected users, such as the number of affected households categorized by user attributes, and a list of sensitive locations, such as a detailed list of sensitive locations such as hospitals and schools within the power outage area.

[0221] Through the above steps, the system simultaneously achieves accurate user attribute identification and dynamic collaborative assessment of the impact of power outages within a heterogeneous graph inference framework. This approach addresses the insufficient business collaboration between power outage assessment and user attribute identification in existing technologies.

[0222] In practical applications, traditional power outage impact analysis often relies solely on simple spatial overlay analysis based on GIS spatial buffers. This fails to distinguish the essential differences between electrical connection relationships and spatial proximity relationships, and also makes it difficult to associate user attribute tags for refined impact assessment.

[0223] This invention unifies user attribute identification and power outage assessment within the same heterogeneous graph reasoning framework. It achieves a dual-dimensional cascaded search of electrical and spatial dimensions by performing fault propagation traversal based on electrical connection edges and evaluating secondary impacts based on spatial adjacency edges. The final comprehensive power outage impact analysis report not only includes statistics on the number and types of users affected by the outage but also marks information on sensitive locations within the outage area. This provides more comprehensive and accurate business support for prioritizing repair resource scheduling, making power supply decisions for critical users, and emergency response.

[0224] Please see Figure 3 One embodiment of the present invention can provide a power user attribute identification device based on spatiotemporal interactive fusion. The power user attribute identification device based on spatiotemporal interactive fusion may include: a data acquisition module, a time-series feature extraction module, a topology feature extraction module, and an attribute identification module.

[0225] The data acquisition module is used to acquire the electricity load sequence and topology adjacency graph of the target user; wherein, the electricity load sequence includes the electricity load of the target user in each sub-electricity consumption period within a preset time period, and the topology adjacency graph includes at least one of the following: the electrical connection relationship between the user node and the distribution transformer equipment, the spatial adjacency relationship between the user node and the neighboring node, and the semantic similarity relationship between the user node and the neighboring node.

[0226] The time-series feature extraction module is used to input the electricity load sequence into the time-series feature extraction model to obtain the electricity load feature vector of the target user.

[0227] The topology feature extraction module is used to input the topology adjacency graph into the relation graph attention network to obtain the topology feature vector of the target user.

[0228] The attribute recognition module is used to input the power load feature vector and the topology feature vector into a preset inference model to obtain the attribute information of the target user.

[0229] The specific functions and effects of the power user attribute identification device based on spatiotemporal interactive fusion can be explained by referring to other embodiments in this specification, and will not be repeated here. Each module in the power user attribute identification device based on spatiotemporal interactive fusion can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0230] Please see Figure 4 One embodiment of the present invention can provide an electronic device, the electronic device comprising:

[0231] A memory, and one or more processors communicatively connected to the memory;

[0232] The memory stores instructions that can be executed by the one or more processors, which, when executed by the one or more processors, enable the one or more processors to implement the power user attribute identification method based on spatiotemporal interactive fusion as described in any of the above embodiments.

[0233] One embodiment of the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the power user attribute identification method based on spatiotemporal interactive fusion described in any of the above embodiments.

[0234] The embodiments of this specification also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the power user attribute identification method based on spatiotemporal interactive fusion described in any of the above embodiments.

[0235] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments described herein, and are not intended to limit the scope of the invention.

[0236] It is understood that in the various embodiments described in this specification, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments described in this specification.

[0237] It is understood that the various implementation methods described in this specification can be implemented individually or in combination, and the implementation methods in this specification are not limited in this respect.

[0238] Unless otherwise stated, all technical and scientific terms used in the embodiments of this specification have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this specification. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0239] It is understood that the processor in this invention can be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method implementation can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this specification. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this specification can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0240] It is understood that the memory in this invention can be volatile memory or non-volatile memory, or may include both. Specifically, the non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM). It should be noted that the memory in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0241] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in the embodiments of the present invention are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0242] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this specification.

[0243] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the aforementioned method implementations, and will not be repeated here.

[0244] In the several embodiments provided in this specification, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0245] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0246] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0247] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of this specification, in essence, or the parts that contribute to the prior art, or parts of the technical solutions, can be embodied in the form of software products. These computer software products are stored in a storage medium and include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this specification. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0248] The above description is merely a specific embodiment of this specification, but the scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this specification should be included within the scope of protection of this specification. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for identifying power user attributes based on spatiotemporal interactive fusion, characterized in that, The method includes: Obtain the target user's electricity load sequence and topological adjacency graph; wherein, the electricity load sequence includes the target user's electricity load for each sub-electricity consumption period within a preset time period, and the topological adjacency graph includes at least one of the following: the electrical connection relationship between the user node and the distribution transformer equipment, the spatial adjacency relationship between the user node and its neighboring nodes, and the semantic similarity relationship between the user node and its neighboring nodes; The electricity load sequence is input into the time-series feature extraction model to obtain the electricity load feature vector of the target user; The topological adjacency graph is input into a relational graph attention network to obtain the topological feature vector of the target user; The electrical load feature vector and the topology feature vector are input into a preset inference model to obtain the attribute information of the target user.

2. The method according to claim 1, characterized in that, Obtain the target user's electricity load sequence and topological adjacency graph, including: Obtain the electricity load sequence of the target user, as well as the distribution transformers, points of interest, and user names within the target area where the target user is located; Based on the electrical adjacency relationships of the distribution transformer equipment, an electrical topology diagram is constructed; Based on the locational relationship between the target user and each of the points of interest, a geographic topology map is constructed; A semantic relationship graph is constructed based on the semantic similarity between the name of the target user and the names of other users within the target area; The electrical topology graph, the geographical topology graph, and the semantic relationship graph are merged into a topological adjacency graph.

3. The method according to claim 2, characterized in that, Based on the locational relationships between the target user and each of the points of interest, a geographic topology map is constructed, including: Based on the positional relationship between the target user node and each point of interest, a Veno graph is constructed; Each point of interest is traversed sequentially. If the target user node and the current point of interest have a common boundary in the Venn diagram, a spatial adjacency edge is established between the target user node and the current point of interest. The weights of the spatial adjacent edges are determined based on the Euclidean distance between the target user node and the current point of interest to obtain a geographic topology map.

4. The method according to claim 2, characterized in that, Based on the semantic similarity between the target user's name and the names of other users within the target region, a semantic relationship graph is constructed, including: Using the BERT model, target word vectors of the target user's name and neighbor word vectors of other users' names within the target region are extracted respectively; For each neighbor word vector, calculate the cosine similarity between the target word vector and the neighbor word vectors; When the cosine similarity is greater than or equal to the preset similarity, semantic similarity edges are constructed between the target user node and the neighboring nodes corresponding to the neighboring word vectors; wherein, multiple semantic similarity edges constitute a semantic relationship graph.

5. The method according to claim 1, characterized in that, The electricity load sequence is input into a time-series feature extraction model to obtain the electricity load feature vector of the target user, including: The electricity load sequence is input into a multi-branch time-series feature extraction model to obtain multiple branch feature vectors; wherein, different branches correspond to different expansion rates. Multiple branch feature vectors are concatenated to obtain multi-scale feature vectors; Based on the target time attention weight, attention is added to the multi-scale feature vector to obtain the target user's electricity load feature vector.

6. The method according to claim 5, characterized in that, Before the step of adding attention to the multi-scale feature vector based on the target time attention weight to obtain the target user's electricity load feature vector, the method further includes: Based on the learnable weight matrix, the multi-scale feature vector is mapped to the attention latent space to obtain the attention intermediate feature vector; The transpose of the attention scoring vector is multiplied by the attention intermediate feature vector to obtain the temporal attention weight vector; Based on the time attention weight vector, a target time attention weight vector is determined; wherein the sum of each target time attention weight in the target time attention weight vector is 1.

7. The method according to claim 1, characterized in that, The topological adjacency graph is input into a relational graph attention network to obtain the topological feature vector of the target user, including: Based on each relationship type in the topological adjacency graph, linear transformations are performed on the current layer input features of the target user's target node and the current layer input features of all neighboring nodes connected to the target node through the relationship type to obtain the projection features of the target node and each neighboring node under the relationship type. The projected features of the target node and the projected features of the corresponding neighbor node are concatenated into vectors. The learnable attention vector corresponding to the relation type is used to perform a linear mapping on the concatenated feature vector. Then, the LeakyReLU activation function is applied to obtain the attention score of the corresponding neighbor node relative to the target node under the relation. The attention scores are normalized to obtain the normalized attention weights of each neighbor node under the relationship type. Using the corresponding normalized attention weight as coefficient, the projected features of all neighbor nodes under the relation type are weighted and summed to obtain the single-relation neighbor aggregation features of the target node under the relation type. The single-relation neighbor aggregation features corresponding to all relation types are summed to obtain the full-relation neighbor aggregation features; the self-connection projection matrix is ​​used to perform a linear transformation on the current layer input features of the target node to obtain the self-connection features; the full-relation neighbor aggregation features are added to the self-connection features to obtain the fused features. The fused features are input into a nonlinear activation function for nonlinear transformation, and the updated node features of the target node in the current layer are output. At least two layers of single-layer graph feature aggregation are stacked, with the output of each layer serving as the input of the next layer. The node features output from the last layer are used as the topology feature vector of the target power user.

8. The method according to claim 1, characterized in that, The electrical load feature vector and the topology feature vector are input into a preset inference model to obtain the attribute information of the target user, including: The electrical load feature vector is input into the first linear transformation layer and linearly mapped to obtain the target time-series feature; The topological feature vector is input into a second linear transformation layer for linear mapping to obtain the target topological feature; wherein the target temporal feature has the same dimension as the target topological feature; The target temporal feature, the target topological feature, the element-wise interaction features of the target temporal feature and the target topological feature, and the absolute value features of the element-wise difference between the target temporal feature and the target topological feature are concatenated to obtain a fused feature vector; The fused feature vector is input into a gated fully connected layer, and the gate weight coefficients are calculated using the Sigmoid activation function. Using the gate weight coefficient as the weighting coefficient of the target temporal feature and the difference between 1 and the gate weight coefficient as the weighting coefficient of the target topological feature, the target temporal feature and the target topological feature are weighted and summed to obtain the spatiotemporal joint fusion feature of the target user; The spatiotemporal joint fusion features are input into the classification fully connected layer to obtain the attribute information of the target user.

9. The method according to claim 8, characterized in that, The method further includes: The power load feature vectors of each power user are linearly mapped through the first linear transformation layer, and the topology feature vectors of each power user are linearly mapped through the second linear transformation layer. The mapped power load feature vectors and topology feature vectors are then normalized to obtain the normalized time series representation vectors and normalized topology representation vectors of each user. Positive sample pairs are formed by the normalized temporal representation vector and the normalized topological representation vector of the same user, and negative sample pairs are formed by the normalized temporal representation vector of the current user and the normalized topological representation vectors of all other users. Based on cosine similarity and a preset temperature adjustment coefficient, the cross-view contrast loss is calculated using the information noise contrast estimation loss function. The temperature adjustment coefficient is used to adjust the granularity of the distinction between positive and negative samples. Based on the degree of connection between user nodes and distribution transformers in the aforementioned topology adjacency graph, a topology consistency loss term is constructed; The cross-view comparison loss and the topology consistency loss term are weighted and summed to obtain the joint total loss of the pre-trained optimization, where the weighting coefficients corresponding to the topology consistency loss term are the topology constraint weight coefficients. Using the joint total loss as the optimization objective, the network parameters of the temporal feature extraction model, the relational graph attention network, and the first and second linear transformation layers are updated synchronously through the backpropagation algorithm. The training is iterated until the loss converges, thus completing the pre-training of the spatiotemporal feature encoder.

10. The method according to claim 1, characterized in that, The method further includes: When a fault is detected in the power equipment node of the target user, the power equipment node of the target user is used as the starting point for retrieval. A topological traversal search is performed based on the electrical connection edge between the user node and the distribution transformer equipment to filter out all power user nodes that have a direct or indirect electrical connection relationship with the power equipment node, thus forming a set of users affected by the power outage. Based on the spatial coverage area corresponding to the set of users affected by the power outage, an extended search is performed along the spatial adjacency edges between the power user nodes and their neighboring nodes to filter out sensitive interest nodes within the affected spatial range. By integrating the user attribute information of the set of users affected by the power outage with the location attribute information of the sensitive interest nodes, a comprehensive power outage impact analysis report is generated, including statistics on affected users and a list of sensitive locations.

11. A power user attribute identification device based on spatiotemporal interactive fusion, characterized in that, The device includes: The data acquisition module is used to acquire the electricity load sequence and topology adjacency graph of the target user; wherein, the electricity load sequence includes the electricity load of the target user in each sub-electricity consumption period within a preset time period, and the topology adjacency graph includes at least one of the following: the electrical connection relationship between the user node and the distribution transformer equipment, the spatial adjacency relationship between the user node and the neighboring node, and the semantic similarity relationship between the user node and the neighboring node; The time-series feature extraction module is used to input the electricity load sequence into the time-series feature extraction model to obtain the electricity load feature vector of the target user; The topology feature extraction module is used to input the topology adjacency graph into the relation graph attention network to obtain the topology feature vector of the target user; The attribute recognition module is used to input the power load feature vector and the topology feature vector into a preset inference model to obtain the attribute information of the target user.

12. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the power user attribute identification method based on spatiotemporal interactive fusion as described in any one of claims 1 to 10.

13. A computer storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the power user attribute identification method based on spatiotemporal interactive fusion as described in any one of claims 1 to 10.

Images