Community crowd flow prediction method driven by community geographic knowledge graph

Abstract

The invention belongs to the technical field of crowd flow prediction, and particularly discloses a community crowd flow prediction method driven by a community geographic knowledge graph, and the method comprises the following steps: obtaining historical crowd flow data of a target community, and constructing entrance and exit crowd flow tensors according to the historical crowd flow data; obtaining multi-source geographic information data of a target community, constructing a community geographic knowledge graph according to the multi-source geographic information data, and learning low-dimensional vector representation corresponding to each entity in the community geographic knowledge graph by adopting a knowledge graph embedding model to obtain an entity embedded vector; extracting spatial features and time features of community crowd flow from the entrance and exit crowd flow tensors, and generating spatial-temporal feature embedding; and by using an attention mechanism, carrying out adaptive weighted fusion on the spatio-temporal feature embedding and entity embedding vectors, and outputting a community crowd flow prediction result through a gating mechanism and linear transformation on a fusion result. The method can effectively improve the precision of crowd flow prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of crowd flow prediction, and more particularly relates to a community crowd flow prediction method driven by a community geographic knowledge graph. BACKGROUND

[0002] Crowd flow is an important geographic spatial data representing the dynamic changes of regional population, and is widely used in many fields such as resident travel behavior analysis, public resource scheduling, urban planning and management, etc. In the community governance scenario, accurately predicting the trend of crowd flow change has important practical value for improving community management efficiency, optimizing public facility layout, and promoting community business activities, etc.

[0003] Current mainstream crowd flow prediction methods are mostly based on traffic network or deep learning driven spatio-temporal modeling techniques, and usually construct models at the city scale, which perform well in large-scale city crowd flow prediction. However, the community spatial structure is complex and the functional areas are highly heterogeneous, and existing models often have difficulty in effectively handling the heterogeneous correlation between micro-spatial units, and also fail to fully capture the complex interaction between crowd activities and community environmental factors (such as building layout, land use function, etc.), resulting in limited prediction performance in the community range.

[0004] In addition, traditional methods mostly rely only on historical crowd flow data and spatial adjacency relations, and lack of structured integration and deep semantic modeling of multi-source geographic semantic information such as regional boundaries, building attributes, functional zoning, etc., which limits the spatial expression ability and prediction interpretability of the model.

[0005] Therefore, how to improve the accuracy of crowd flow prediction is a problem that needs to be solved at present. SUMMARY

[0006] In view of the defects of the prior art, the purpose of the present application is to provide a community crowd flow prediction method driven by a community geographic knowledge graph, which can effectively improve the accuracy of crowd flow prediction.

[0007] To achieve the above purpose, in a first aspect, the present application provides a community crowd flow prediction method driven by a community geographic knowledge graph, comprising the following steps: S10, acquiring and according to the historical crowd flow data of the target community, dividing the community into S10, acquiring and according to the historical crowd flow data of the target community, dividing the community into S20, acquire and construct a community geographic knowledge graph containing community area entities, POI entities, building entities and the relationships therebetween according to multi-source geographic information data of the target community, including community area boundaries, POI information and building information, and learn a low-dimensional vector representation corresponding to each entity in the community geographic knowledge graph using a knowledge graph embedding model to obtain entity embedding vectors; S30, extract spatial and temporal features of community population flow from the entrance and exit population flow tensors to generate spatiotemporal feature embeddings; S40, use an attention mechanism to adaptively weight and fuse the spatiotemporal feature embeddings and the entity embedding vectors, and output community population flow prediction results through a gating mechanism and linear transformation.

[0008] The community geographic knowledge graph driven community population flow prediction method provided in the application has the beneficial effects that: by fusing community geographic information and population flow data, a community geographic knowledge graph is constructed to realize structured representation of the environment; meanwhile, spatiotemporal features are extracted from historical flow data, and an attention mechanism is used to effectively fuse geographic semantic features and spatiotemporal dynamic features. This method helps to alleviate the problems of insufficient depiction of micro community environment heterogeneity and insufficient utilization of semantic information in existing population flow prediction models, thereby improving prediction accuracy and generalization performance.

[0009] As a further optimization, in step S10, the step of regularly dividing the community is specifically: Based on the latitude and longitude information of the community geographic boundary, the target community is regularly divided according to a preset grid size, and the entire community is divided into community sub-areas; For each community sub-area, the personnel entry and exit flow and in a given time window T are counted to obtain the population entry and exit flow values of each sub-area at different time periods T, and a multi-dimensional tensor is constructed based on the time dimension and the spatial area to represent the changes of the entry and exit population flow of the community in the spatiotemporal dimension.

[0010] As a further optimization, in step S20, the step of constructing the community geographic knowledge graph and extracting the entity embedding vectors is specifically: The community geographic knowledge graph is constructed, including defining entity types and relationship types, the defined entity types in the knowledge graph including community areas, POIs, buildings and POI category entities, and the relationship types including the containing relationship between areas and POIs, the subordinate relationship between POIs and buildings, the spatial adjacency relationship between areas and buildings, and the attribution relationship between POIs and categories; The constructed community geographic knowledge graph is trained by using a knowledge graph embedding model, and embedding vectors corresponding to each entity in the graph are generated by using a graph embedding method.

[0011] As a further preferred, in step S30, the step of extracting the spatial and temporal features of the community population flow is as follows: The spatial features of the community population flow data are extracted by using a relational graph convolution network, which aggregates the information of the entity neighbors in the graph based on different types of relationships to obtain and fuse the spatial correlation features of the region; The temporal features of the community population flow data are extracted by using a dilated causal convolution, which adjusts the spacing between the convolution kernels by setting a dilation factor to expand the receptive field and thus enhance the modeling capability for long time series.

[0012] As a further preferred, in step S40, the step of using an attention mechanism for adaptive weighted fusion is as follows: The attention mechanism is introduced, the entity embedding vector of the community geographic knowledge graph is taken as the query vector, the spatial feature embedding and the temporal feature embedding in the spatio-temporal feature embedding are taken as the key and the value respectively, and the adaptive weighted fusion of the spatial and temporal features is realized by the attention weight calculation to generate a unified representation of the fused semantics and spatio-temporal dynamics; The fused representation is sequentially passed through a gating mechanism, a nonlinear activation function and a linear transformation, which are used to refine the feature representation and output the final population flow prediction value.

[0013] In a second aspect, the present application provides an application of the community geographic knowledge graph driven community population flow prediction method as described in any one of the above aspects, which is applied to community planning, dynamic management and public resource optimization configuration scenarios.

[0014] In a third aspect, the present application provides a community geographic knowledge graph driven community population flow prediction system, which comprises: A data acquisition module is configured to acquire historical population flow data and multi-source geographic information data of a target community, wherein the multi-source geographic information data comprises community region boundaries, point of interest (POI) information and building information. A flow tensor construction module is configured to divide the community into a plurality of regular grid units according to the community geographic boundaries based on the historical population flow data, count the in-and-out population flow of each unit within a set time period, and construct an entrance and exit population flow tensor. a knowledge graph construction module, configured to construct a community geographic knowledge graph containing community area entities, POI entities, building entities and relationships therebetween based on multi-source geographic information data, and learn a low-dimensional vector representation corresponding to each entity in the community geographic knowledge graph using a knowledge graph embedding model to obtain an entity embedding vector; a spatio-temporal feature extraction module, configured to extract spatial features and temporal features of community crowd flow by inputting the entrance and exit crowd flow tensors into the spatio-temporal feature extraction module, and generate spatio-temporal feature embeddings; an embedding fusion module, configured to adaptively weight and fuse the spatio-temporal feature embeddings and the entity embedding vectors using an attention mechanism, and output a community crowd flow prediction result by passing the fusion result through a gating mechanism and a linear transformation.

[0015] In a fourth aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the community geographic knowledge graph driven community crowd flow prediction method according to any one of the above aspects when executing the program.

[0016] In a fifth aspect, the present application provides a computer readable storage medium having computer instructions stored thereon, wherein the computer instructions are executable on a processor to implement the community geographic knowledge graph driven community crowd flow prediction method according to any one of the above aspects.

[0017] It can be understood that the beneficial effects of the above-mentioned second, third, fourth and fifth aspects can be referred to the related description in the first aspect, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of the community geographic knowledge graph driven community crowd flow prediction method provided by the present application; Figure 2 is a flowchart of the community geographic knowledge graph driven crowd flow prediction method provided by the present application; Figure 3 is a framework diagram of the community geographic knowledge graph driven crowd flow prediction method provided by the present application; Figure 4 is a schematic diagram of the community knowledge graph constructed based on community geographic information provided by the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0020] To solve the problem that the existing crowd flow prediction method is insufficient in describing the heterogeneity of the micro geographical environment within the community and insufficient in utilizing the spatio-temporal characteristics, resulting in limited prediction accuracy, the present application provides a community geographical knowledge graph driven community crowd flow prediction method. The method fuses various geographical elements (such as regions, interest points, buildings, etc.) within the community with crowd flow data to construct a structured community knowledge graph (CKG). Combined with spatio-temporal feature extraction and attention fusion mechanism, the method realizes fine-grained modeling and accurate prediction of community crowd flow.

[0021] The present application uses knowledge graph embedding technology to uniformly represent various entities and their relationships within the community, and introduces an attention mechanism to fuse semantic, spatial and temporal features, thereby enhancing the model's ability to identify complex relationships between regions and dynamic flow patterns. This provides a more reliable prediction basis for community safety management, resource optimization, and public facility planning.

[0022] As shown in Figure 1 , the community geographical knowledge graph driven community crowd flow prediction method provided by the present application can be applied to community planning, dynamic management, public resource optimization, etc., providing data support and decision-making reference for community governance and sustainable development. The method includes steps S10 to S40, which are described in detail as follows: Step S10, obtaining and according to the historical crowd flow data of the target community, dividing the community into regular grid units according to the community geographical boundary, counting the in-out crowd flow of each unit within a set time period, and constructing an entry and exit crowd flow tensor.

[0023] In step S10, the community is divided into regular grid units, and an entry and exit crowd flow tensor is constructed. Specifically, the method can be as follows: Step S11, based on the latitude and longitude information of the community boundary, the target community is divided into several discrete community sub-regions according to a predetermined grid size; Step S12, for each community sub-region, counting the personnel in-out flow and in a given time window T to obtain the crowd in-out flow values of each sub-region under different time periods T, and constructing a multi-dimensional tensor based on the time dimension and spatial region, which is used to represent the change of in-out crowd flow of the community in the spatio-temporal dimension.

[0024] In step S20, the multi-source geographic information data of the target community is acquired, including community area boundaries, point of interest (POI) information, and building information, a community geographic knowledge graph containing community area entities, POI entities, building entities, and relationships therebetween is constructed, and a knowledge graph embedding model is used to learn a low-dimensional vector representation corresponding to each entity in the community geographic knowledge graph to obtain entity embedding vectors.

[0025] In step S20, the steps of constructing the community geographic knowledge graph and extracting the entity embedding vectors can be specifically as follows: In step S21, the community geographic knowledge graph is constructed, first, entity types are defined, including community areas, points of interest (POI), buildings, POI categories, etc.; then, relationship types are defined, such as “area-contains-POI”, “POI-belong-to-category”, “area-contains-building”, “building-adjacent-building”, etc., and a multi-element heterogeneous knowledge graph is constructed accordingly; In step S22, the knowledge graph embedding model is used to train the graph, and a low-dimensional dense vector representation of each community area entity is learned through graph embedding to capture the semantic and spatial features of the area in the graph.

[0026] In step S30, the spatial and temporal features of community population flow are extracted from the entrance and exit population flow tensors to generate spatio-temporal feature embeddings.

[0027] In step S30, the steps of extracting the spatial and temporal features of community population flow can be specifically as follows: In step S31, the spatial feature embedding of community population flow data is extracted using a relational graph convolutional network (RGCN); according to the different relationship types predefined in the graph, the neighbor information of each entity node is aggregated and fused according to the relationship categories, so as to obtain the spatial structure feature representation of population flow under different semantic relationships; ; In step S32, the temporal feature of population flow is extracted using dilated causal convolution (Dilated Causal Convolution); by setting the dilation factor to control the convolution kernel spacing, the receptive field is gradually expanded to enhance the modeling ability of long-term temporal dependence; with the help of multi-layer dilated convolution structure of different scales, the dynamic pattern of community population flow under multiple time granularities is captured .

[0028] In step S40, the spatio-temporal feature embeddings and the entity embedding vectors are adaptively weighted and fused using an attention mechanism, and the fusion result is output through a gating mechanism and linear transformation to output the community population flow prediction result.

[0029] In step S40, the step of adaptive weighted fusion using attention mechanism can be specifically as follows: In step S41, the attention mechanism is introduced, the entity embedding vector in the knowledge graph obtained in step S20 is taken as a query vector (Query), the spatial embedding and the temporal embedding extracted in step S30 are taken as a key (Key) and a value (Value) respectively, and adaptive weighted fusion of the spatial and temporal features is realized through attention weight calculation to generate a unified representation of fused semantics and spatiotemporal dynamics. .

[0030] In step S42, the above fused representation is input into a post-processing module, which includes a gating unit, a nonlinear activation function and a linear transformation layer, for refining the feature representation and outputting a final crowd flow prediction value. .

[0031] The community geographic knowledge graph driven crowd flow prediction method provided by the present application has the beneficial effects that: by fusing community geographic information and crowd flow data, a community geographic knowledge graph is constructed to realize the structured representation of the environment; meanwhile, the spatiotemporal features are extracted from the historical flow data, and the effective fusion of geographic semantic features and spatiotemporal dynamic features is realized with the aid of the attention mechanism. This method helps to alleviate the problems of insufficient description of micro community environment heterogeneity and insufficient utilization of semantic information in existing crowd flow prediction models, thereby improving the prediction accuracy and generalization performance.

[0032] In one embodiment, the technical solution for achieving the above-mentioned purpose can be specifically as follows: Figure 2 and 3 As shown in the above two figures, the present embodiment provides a community geographic knowledge graph driven crowd flow prediction method, which specifically includes the following steps: Step 1: Obtain crowd flow data and construct a regional flow tensor, obtain the historical crowd flow data of the community to be predicted, divide the community into community regions, and construct the in-out crowd flow tensor of each region according to the division result. and ; the specific method is as follows: Step 1-1: Based on the geographic boundary information of the community, regular grid division is performed according to the latitude and longitude. Assuming that the maximum latitude of the community coverage area is , the minimum latitude is , the maximum longitude is , the minimum longitude is , and the latitude and longitude direction span of the current grid region unit is , the community is divided into regions in the latitude and longitude directions. I, J The calculation formula is as follows: (1) (2) Step 1-2: According to the region division result in step 1-1, the i-th region The inflow and outflow of the t-th time slice are defined as follows: (3) (4) where P represents a set of trajectories in the t-th time interval, is a trajectory in P, is the m-th geographic space coordinate within the trajectory.

[0033] Step 2: Obtain multi-source geographic information data of the target community, including community area boundary, point of interest (POI) information, building information, etc., construct a community geographic knowledge graph based on the data, and extract structured representation vectors of each region using a graph embedding model. Specifically, the following sub-steps are included: Step 2-1: Define and construct a community geographic knowledge graph. Refer to the entity-relation modeling method in the database, abstract the entity types and relation types in the graph structure from the multi-source heterogeneous geographic data of the community. As shown in the following table, specifically: Figure 4 First, extract information related to the community environment from public map platforms (such as Baidu Map) to construct the entity set in the graph. The entity types include but are not limited to: POI entity , community area entity , building entity in the community , POI category entity . Second, based on the above entity recognition results, extract typical relation types in the graph to describe the semantic and spatial correlation between different entities. The relations include but are not limited to: boundary relation of area entity , spatial relation between building entity and area entity , spatial relation between POI entity and area entity , relation between building entity and POI entity . Based on the above entities and relations, the community geographic knowledge graph is constructed, providing a structured representation basis for subsequent embedding learning. Step 2-2: Based on the constructed community knowledge graph, use the graph embedding model to learn the embedding vectors of the entities. Specifically, this embodiment uses the TuckER model (Tucker Embedding for Knowledge Graph Completion) to evaluate the triples in CKG

[0034] ​rationality to obtain the low-dimensional vector representation of the community area in the graph. The TuckER model takes triplets as input and scores the rationality by tensor decomposition. The scoring function is defined as follows: (5) where, is a learnable core tensor, denotes the tensor product along the nth dimension, are the embedding vectors of the head entity , tail entity and relation , respectively.

[0035] After training the model, the structured embedding representation of each community area entity in the knowledge graph can be learned, which is used to guide the fusion of subsequent spatio-temporal features and crowd flow prediction.

[0036] Step 3: Input the crowd flow tensor constructed in step 1 into the spatio-temporal feature extraction module to extract spatial and temporal features respectively, and obtain the crowd flow spatio-temporal embedding representation for fusion. Specifically, the following sub-steps are included: Step 3-1: Use the spatial prediction module to model the spatial dependence of inter-regional crowd flow. Specifically, use a layer of Relational Graph Convolutional Network (R-GCN) to perform embedding calculation on the crowd flow features at the regional level. For the lllth community area, the spatial feature calculation formula is as follows: (6) where, is the input feature of region l, denotes the set of regions connected to region l through relation r, and are learnable weight matrices, and σ is a nonlinear activation function.

[0037] Step 3-2: Use the time prediction module to model the time trend of community crowd flow. Specifically, use Dilated Causal Convolution to build a time convolution layer to enhance the model's ability to perceive long-time dependencies. For region i, the dilated convolution is defined as: (7) where, d is the dilation factor, is the flow data of node i, is a one-dimensional convolution filter with size k.

[0038] To improve the time modeling ability of the model, multiple convolution filters of different sizes are used in the initial layer to extract features at different time scales in parallel. The outputs of the two groups of dilated convolution are denoted as: (8) Finally, the two convolution branches are fused through a gating mechanism to obtain the time feature embedding: (9) Step 4: In this step, by introducing a multi-head attention mechanism, the spatial and temporal features extracted in step 3 are fused with the regional knowledge graph embedding extracted in step 2, thereby enhancing the model's joint perception ability of geographic semantics and spatio-temporal changes, and finally realizing the prediction of crowd flow. Specifically, it includes the following sub-steps: Step 4-1: For each region l, let and denote the representations after the spatial and temporal modules, denote the knowledge graph embedding of the region.

[0039] First, project these vectors to obtain query vector , key vector and value vector , use the attention mechanism to calculate the importance of the time and space representations and fuse them: (10) (11) Step 4-2: Input the spatio-temporal representation obtained by weighted fusion into the embedding post-processing module. First, output the multi-head attention layer after fusion to the gating unit: (12) Then, transform through the activation layer and linear layer into the final prediction result .

[0040] Those skilled in the art will readily understand that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A community population flow prediction method driven by a community geographic knowledge graph, characterized in that, Includes the following steps: S10, Obtain and divide the target community into sections based on historical population flow data and geographical boundaries. Each grid cell is a regular unit. The number of people entering and exiting each cell within a set time period is counted to construct the entrance and exit crowd flow tensors. S20. Acquire and construct a community geographic knowledge graph containing community area entities, POI entities, building entities and their relationships based on multi-source geographic information data of the target community, including community area boundaries, points of interest (POI) information and building information. Then, use a knowledge graph embedding model to learn the low-dimensional vector representation of each entity in the community geographic knowledge graph to obtain the entity embedding vector. S30: Extract the spatial and temporal features of community population flow from the entrance and exit crowd flow tensor to generate spatiotemporal feature embeddings; S40, using an attention mechanism, the spatiotemporal feature embedding and entity embedding vector are adaptively weighted and fused, and the fusion result is output as a community crowd flow prediction result through a gating mechanism and linear transformation.

2. The community population flow prediction method driven by community geographic knowledge graph as described in claim 1, characterized in that, Step S10, the step of dividing the community into a regular grid, is as follows: Based on the latitude and longitude information of the community's geographical boundaries, the target community is systematically divided according to a preset grid size, thus dividing the entire community into... Each community sub-area; For each sub-area of ​​the community, the flow of people entering and leaving the area within a given time window T is statistically analyzed. and We obtain the population flow data of each sub-region at different time periods T, and construct a multi-dimensional tensor based on the time dimension and spatial region to represent the changes in population flow in and out of the community in the spatiotemporal dimension.

3. The community population flow prediction method driven by community geographic knowledge graph as described in claim 1, characterized in that, Step S20, the step of constructing the community geographic knowledge graph and extracting entity embedding vectors, is as follows: Constructing a community geographic knowledge graph includes defining entity types and relationship types. The entity types in the knowledge graph include community regions, points of interest (POIs), buildings, and POI category entities. The relationship types include the inclusion relationship between regions and POIs, the subordinate relationship between POIs and buildings, the spatial adjacency relationship between regions and buildings, and the affiliation relationship between POIs and categories. The constructed community geographic knowledge graph is trained using a knowledge graph embedding model, and the graph embedding method is used to generate corresponding embedding vector representations for each entity in the graph.

4. The community population flow prediction method driven by community geographic knowledge graph as described in claim 1, characterized in that, Step S30, the step of extracting the spatial and temporal features of community population movement, specifically includes: A relational graph convolutional network is used to extract spatial features of community population flow data. The relational graph convolutional network performs weighted aggregation of information of entity neighbors in the graph based on different types of relationships to obtain and fuse the spatial association features of the region. Dilated causal convolution is used to extract the temporal features of community population flow data. The dilated causal convolution adjusts the spacing between convolution kernels by setting a dilation factor to expand the receptive field, thereby enhancing the ability to model long-term series.

5. The community population flow prediction method driven by community geographic knowledge graph as described in claim 1, characterized in that, In step S40, the step of adaptive weighted fusion using the attention mechanism is as follows: An attention mechanism is introduced, using the entity embedding vector of the community geographic knowledge graph as the query vector, and the spatial feature embedding and temporal feature embedding in the spatiotemporal feature embedding as the key and value respectively. Through attention weight calculation, the spatial and temporal features are adaptively weighted and fused to generate a unified representation of fused semantics and spatiotemporal dynamics. The fused representation is sequentially passed through a gating mechanism, a nonlinear activation function, and a linear transformation to refine the feature representation and output the final predicted value of the crowd flow.

6. An application of a community population flow prediction method driven by a community geographic knowledge graph as described in any one of claims 1 to 5, characterized in that, It can be applied to community planning, dynamic management, and optimization of public resource allocation.

7. A community population flow prediction system driven by a community geographic knowledge graph, characterized in that, include: The data acquisition module is used to acquire historical population flow data and multi-source geographic information data of the target community. The multi-source geographic information data includes community area boundaries, point of interest (POI) information, and building information. The traffic tensor construction module is used to divide the community into sections based on historical population traffic data and the community's geographical boundaries. Each grid cell is a regular unit. The number of people entering and exiting each cell within a set time period is counted to construct the entrance and exit crowd flow tensors. The knowledge graph construction module is used to construct a community geographic knowledge graph based on multi-source geographic information data, which includes community area entities, POI entities, building entities and their relationships. It also uses a knowledge graph embedding model to learn the low-dimensional vector representation of each entity in the community geographic knowledge graph to obtain the entity embedding vector. The spatiotemporal feature extraction module is used to extract the spatial and temporal features of community population flow from the entrance and exit crowd flow tensors, and generate spatiotemporal feature embeddings. The embedding fusion module is used to adaptively weight and fuse the spatiotemporal feature embeddings with the entity embedding vectors using an attention mechanism, and output the community crowd flow prediction result through a gating mechanism and linear transformation.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the community population flow prediction method driven by the community geographic knowledge graph as described in any one of claims 1 to 5.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement the community population flow prediction method driven by the community geographic knowledge graph as described in any one of claims 1 to 5.

Images