A business district passenger flow prediction method based on a graph neural network
By constructing a dynamic physical network of business districts, integrating multi-source data, and utilizing graph neural networks and time-series modeling, the problem of dynamically depicting customer flow propagation and providing real-time decision support in business district customer flow prediction was solved. This enabled dynamic prediction and intelligent scheduling of business district customer flow, improving prediction accuracy and scheduling efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 灯火星球(北京)国际数字科技有限公司
- Filing Date
- 2025-12-15
- Publication Date
- 2026-06-26
AI Technical Summary
Existing commercial district customer flow prediction technologies cannot dynamically depict customer flow propagation, lack real-time fusion of multi-source data, and the models cannot adapt to real-time control and lack closed-loop decision support, resulting in failure in emergency dispatch scenarios.
By constructing a dynamically adjustable physical network of business districts, integrating multi-source real-time data, extracting the correlation features of node behavior, and combining graph neural networks and temporal modeling, the entire link optimization from customer flow perception and dynamic prediction to intelligent scheduling is achieved. Through graph convolution and self-attention mechanism, the propagation weights between nodes are dynamically learned, and the network customer flow impact of different traffic diversion schemes is simulated to form a decision-making closed loop of 'perception-prediction-simulation-intervention'.
It possesses dynamic adaptability and real-time response capabilities, supports adjustments to the graph structure connections based on actual control measures and emergencies, enables proactive early warning and pre-event intervention, improves forecast accuracy and scheduling efficiency, and forms a closed-loop decision support system from forecasting to scheduling.
Smart Images
Figure CN121660739B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology, and mainly relates to a method for predicting customer traffic in business districts based on graph neural networks. Background Technology
[0002] The method of predicting customer traffic in business districts originated from the rapid development of business districts and commercial areas, as well as the increasing demand for customer traffic prediction from managers. As business districts expand, the customer traffic within them becomes more complex, and traditional management methods can no longer effectively cope with the increasing forecasting needs.
[0003] In Chinese patent application CN202510384845.3, a method, system, and apparatus for predicting customer flow in commercial districts based on graph neural networks are disclosed. This invention relates to the field of commercial district customer flow prediction technology. The method mainly includes: constructing a feature space for predicting customer flow in commercial districts based on traffic zone sets and commercial district sets; embedding traffic zone features into a graph neural network to obtain traffic zone embedding vectors; embedding commercial district features into a first multilayer perceptron to obtain commercial district embedding vectors; connecting the embedding vectors with corresponding OD flow features to obtain OD flow comprehensive representation vectors; training and validating a second multilayer perceptron; and using the OD flow comprehensive representation vector that performs best on the validation set as input to a fully connected neural network to predict commercial district customer flow. This solution incorporates the geographic semantic information of land parcels into the model for feature embedding, constructing a multi-graph structure combining two spatial dependencies. By using graph neural networks and graph embedding techniques to predict OD flow in commercial districts, the prediction accuracy is higher.
[0004] The aforementioned patent has provided a relatively complete method for predicting passenger flow in commercial districts. Existing commercial district passenger flow prediction technologies mostly focus on macroscopic and static origin-destination flow prediction, which essentially treats traffic zones and commercial districts as independent units for total volume analysis. This method cannot depict the dynamic propagation and real-time distribution of passenger flow in the complex network within large transportation hubs. Moreover, it lacks deep integration of high-real-time cross-domain operational data, making it difficult for the model to perceive and respond to the impact of sudden events. In particular, the traditional model structure is rigid and cannot dynamically adjust network connections according to real-time control instructions to simulate passenger flow detour patterns, causing it to fail in emergency dispatch scenarios. Ultimately, the existing solutions are limited to passive early warning and lack the decision support capability to conduct virtual simulation, comparison, and optimization of prediction results and dispatch plans, failing to form a closed loop from "perception" to "intervention".
[0005] To address the aforementioned issues, this invention proposes a method that constructs a dynamically adjustable physical network of commercial districts, integrates multi-source real-time data, extracts node behavioral correlation features, combines graph neural networks and temporal modeling techniques, and introduces closed-loop simulation and decision support mechanisms to achieve end-to-end optimization from passenger flow perception and dynamic prediction to intelligent scheduling. The method abstracts transportation hubs and commercial spaces into dynamically adjustable graph structures, with nodes covering various key physical locations and edge attributes including traffic direction and capacity parameters. Furthermore, it integrates four types of real-time data: passenger flow, traffic, commerce, and environment, constructing a feature vector for each node that integrates temporal, event, and environmental states. Using graph neural networks combined with temporal dependency modeling, it dynamically learns the propagation weights between nodes using graph convolution and self-attention mechanisms, and adjusts connection relationships based on real-time event encoding to predict passenger flow changes under sudden events. Finally, based on the prediction, it evaluates the impact of different diversion schemes on network passenger flow through simulation, selects the optimal scheduling strategy based on congestion identification and effect comparison, and forms a complete decision-making closed loop of "perception—prediction—simulation—intervention." Summary of the Invention
[0006] This invention provides a method for predicting customer flow in commercial districts based on graph neural networks, aiming to solve the problems of existing technologies that cannot dynamically depict customer flow propagation, lack real-time fusion of multi-source data, cannot adapt to real-time control, and lack closed-loop decision support.
[0007] To solve the above problems, the present invention employs the following technology:
[0008] A method for predicting customer traffic in commercial districts based on graph neural networks:
[0009] The method is based on an independent system, including a business district network construction module, a data acquisition module, a feature generation and fusion module, a graph neural network prediction module, and a result application and early warning module; Specific implementation steps
[0010] Step S1: The business district network construction module abstracts the physical structure of transportation hubs and commercial spaces into a dynamic network, and establishes connection relationships based on actual travel paths to form a business district network diagram that represents the physical spatial structure.
[0011] Step S2: The data acquisition module systematically collects four types of real-time data: passenger flow count, traffic operation, business activities, and external environment, and constructs a unified spatiotemporal data sequence to support prediction;
[0012] Step S3: The feature generation and fusion module fuses temporal, operational, and environmental features for each node, and mines behavioral relationships between nodes based on historical passenger flow collaboration patterns to form enhanced network feature inputs, specifically including:
[0013] For each node, time slices are formed based on data acquisition intervals, and all real-time data is aggregated and allocated to the corresponding time slices;
[0014] Next, a feature vector is generated for each node in real time. This vector is concatenated with the node's name, the node's current passenger flow, the simple difference from the previous moment, the names of all its connecting edges, and the real-time events and environmental state expressed by one-hot encoding, thus forming a node feature vector that only reflects the current instantaneous state.
[0015] By analyzing the feature vector of each node, a stable correspondence can be found between its different feature values;
[0016] Identify and quantify the co-occurrence and co-variation patterns between event encoding and passenger flow values within feature vectors; extract the correlation between nodes based on the correlation of changes, and generate a behavioral correlation strength vector for each node, which quantifies the similarity of the behavioral patterns of the node with all other nodes in the network at the current moment.
[0017] This vector is used as a new feature dimension and concatenated to the original feature vector of the node to form the final enhanced feature input;
[0018] Step S4: The graph neural network prediction module constructs a neural network model that integrates graph structure and temporal dependence. By dynamically adjusting the connection weights, it learns the passenger flow propagation and detour patterns under control measures, and completes the model training.
[0019] Step S5: The results application and early warning module obtains future passenger flow forecasts based on real-time data; formulates diversion plans and simulates and compares the network passenger flow impact of different plans.
[0020] As a preferred embodiment, the independent system specifically includes:
[0021] Business district network construction module; based on the geographical conditions of business districts, it abstracts and constructs a virtual business district network;
[0022] Data acquisition module: Integrates multiple data acquisition systems to collect multi-source data, including passenger flow data, traffic operation data, business district operation data, and external environment data;
[0023] Feature generation and fusion module: Fusion of multi-source data at sub-nodes, calculation of data correlation, and formation of feature input that takes into account both spatial relationships and behavioral connections;
[0024] Graph Neural Network Prediction Module: Based on the virtual business district network and feature input, a graph neural network model is constructed; deep learning algorithms are used to learn the behavioral patterns of the crowd.
[0025] Results Application and Early Warning Module: Predicts the operation of the business district based on the behavior patterns of the crowd; formulates traffic diversion plans based on the prediction results; simulates the operation results based on the traffic diversion plans and compares the effects of multiple plans.
[0026] As a preferred embodiment, the formation of the business district network map representing the physical spatial structure specifically includes:
[0027] Within a transportation hub-type business district, all key physical locations that have a significant impact on the gathering and dispersal of people are selected as network nodes, specifically including: transportation nodes, commercial nodes, and connecting nodes.
[0028] Based on the actual building structure and pedestrian feasible paths, establish connection relationships between nodes: label the connection edges with attributes; and model special access restrictions.
[0029] The obtained nodes and connections are stored in a graph data structure to form a static basic network; and an entry point for modifying node connections is reserved to support dynamic updates of connection attributes based on temporary control measures.
[0030] As a preferred implementation, the data acquisition module systematically collects four types of real-time data: passenger flow counts, traffic operations, commercial activities, and the external environment, constructing a unified spatiotemporal data sequence to support prediction. Specifically, this includes:
[0031] The hardware components of the data acquisition module include: installing binocular cameras on nodes and connecting channels in the graph structure; adding IoT collectors to adjustable devices to obtain real-time device operating status; interfacing with the public transportation operation system to obtain real-time train arrival and departure times and scheduling plans; and associating with relevant systems to obtain data affecting passenger flow.
[0032] The central processing unit collects four types of data streams at a preset frequency, including: passenger flow count data, traffic operation data, business activity data, and external environment data; and preprocesses the collected data.
[0033] As a preferred embodiment, the preprocessing of the collected data specifically includes:
[0034] Preprocessing steps include: time alignment, spatial mapping, and format normalization;
[0035] Subsequently, outliers and missing values in the collected data are processed and filled using interpolation of adjacent time slices. Finally, the standardized data is serialized and output: structured data packets are generated in time slice order, and each data packet contains the feature vectors of all nodes at the current time, forming a continuous and complete spatiotemporal data sequence.
[0036] As a preferred implementation, the graph neural network prediction module constructs a neural network model that integrates graph structure and temporal dependence. By dynamically adjusting connection weights, it learns the passenger flow propagation and detour patterns under traffic control measures. The model training specifically includes:
[0037] Specifically, a neural network model that integrates graph structure and temporal dependency is constructed. Graph neural networks (GNNs) are used to combine graph structure and temporal features to learn the propagation rules between nodes. Through graph convolutional layers, node features are transmitted between their neighboring nodes, thereby capturing the relationship with neighboring nodes.
[0038] During the model learning process, the connection weights are dynamically adjusted through the self-attention mechanism of the graph neural network. The graph neural network dynamically adjusts the connection weights between nodes through training. Based on the learning results, the model identifies the connection relationships that have a greater impact on changes in customer traffic, and adjusts the weights of each node according to these impacts. The model is also trained by combining the one-hot encoding of events in actual business district operations.
[0039] During the learning process, the focus is on analyzing the relationship between passenger flow direction and various characteristics. Based on a large amount of data training, the flow direction of passenger flow is learned when certain characteristics exist at different nodes. Individual pattern calculations are performed for each node, and finally, the model is constructed.
[0040] As a preferred implementation, the result application and early warning module obtains future passenger flow predictions based on real-time data, specifically including:
[0041] The data obtained from the current node is bound to the node, and the obtained data is input into the graph neural network model for prediction. Based on the patterns obtained from long-term training of the model, the changes in passenger flow in the future are predicted from different time and event codes.
[0042] During the simulation, all passenger flows follow the preset passenger flow direction. Combined with the learned passenger flow patterns, the changes in passenger flow at different nodes are analyzed and predicted in the entire model to obtain the changes in passenger flow over a future period.
[0043] When events occur at certain nodes, the attraction of that node to other nodes is assessed based on the historical event patterns of that node. Based on the network's entry point, changes in passenger flow are predicted, and the impact of the event on the business district's operation is simulated. Different passenger flow carrying capacities are set for all nodes in the simulation.
[0044] Passenger capacity is set based on the actual space of the node. The specific setting method is calculated by the model based on the maximum passenger flow recorded by the historical node data, using 80% as the limit.
[0045] When the passenger flow at a certain node exceeds the passenger flow capacity set for that node during the simulation, it is recorded as a congestion event at the current node. The number of congestion events is recorded, and the operation results are fully simulated according to the preset time. The total number of passenger flows generated at each moment during the process is counted.
[0046] The simulation results are used to assess the degree of congestion. A comprehensive assessment of the congestion level is conducted based on the number of times nodes are congested and the passenger flow capacity of nodes in the simulation. When the number of times nodes in the simulated business district network map are congested exceeds a preset threshold, it is determined that the business district is severely congested under the current operating mode. The simulation results are then sent to the operation system for staff reference.
[0047] As a preferred implementation method, the process of formulating traffic diversion schemes and simulating and comparing the network passenger flow impact of different schemes specifically includes:
[0048] When severe congestion events are predicted in the module's results, multiple diversion schemes are developed and simulated to select the optimal scheme. The specific diversion scheme development method and content are not fixed, and the development method is based on the node connection characteristics of the business district and the actual business district management scheme.
[0049] After developing several traffic diversion plans, the parameters in the business district network diagram are modified, and the operational results are simulated. From the simulation results, the plan with the fewest congestion events is selected, and the plan and prediction results are sent to the operation management system for staff reference.
[0050] The beneficial effects of this invention are:
[0051] 1. It has dynamic adaptation and real-time response capabilities, supports dynamic adjustment of graph structure connection relationships and weights based on real-time information such as actual control measures and emergencies, and can respond promptly to changes in passenger flow in scenarios such as traffic closures and event holding;
[0052] 2. Achieve closed-loop decision support from prediction to scheduling. Based on prediction, the system can simulate and compare the effects of various diversion schemes to assist operators in selecting the optimal scheduling strategy, forming an integrated decision-making closed loop of "perception-prediction-simulation-intervention".
[0053] 3. Enhance proactive early warning and pre-event intervention capabilities, enabling early warnings to be issued before passenger congestion occurs, and evaluating the effectiveness of intervention measures through simulation and analysis, helping managers to implement pre-event resource allocation and traffic flow optimization, transforming passive response into proactive management. Attached Figure Description
[0054] Figure 1 This is a flowchart of the method of the present invention;
[0055] Figure 2 This is a comparison diagram of the effects of the present invention compared to the traditional method. Detailed Implementation
[0056] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.
[0057] Example 1 Figure 1 As shown in the flowchart of the present invention, this embodiment provides a method for predicting customer traffic in a business district based on a graph neural network, specifically including the following:
[0058] This method is based on an independent system architecture, whose modules include:
[0059] Business district network construction module; based on the geographical conditions of business districts, it abstracts and constructs a virtual business district network;
[0060] Data acquisition module: Integrates multiple data acquisition systems to collect multi-source data, including passenger flow data, traffic operation data, business district operation data, and external environment data;
[0061] Feature generation and fusion module: Fusion of multi-source data at sub-nodes, calculation of data correlation, and formation of feature input that takes into account both spatial relationships and behavioral connections;
[0062] Graph Neural Network Prediction Module: Based on the virtual business district network and feature input, a graph neural network model is constructed; deep learning algorithms are used to learn the behavioral patterns of the crowd.
[0063] Results Application and Early Warning Module: Predicts the operation of the business district based on crowd behavior patterns; develops traffic diversion plans based on the prediction results; simulates the operation results based on the traffic diversion plans and compares the effects of multiple plans;
[0064] Based on the above modules, this embodiment provides a method for predicting customer traffic in a business district based on a graph neural network. The specific implementation steps are as follows:
[0065] Step S1: The business district network construction module abstracts the physical structure of transportation hubs and commercial spaces into a dynamic network, and establishes connection relationships based on actual travel paths to form a business district network diagram that represents the physical spatial structure.
[0066] Step S2: The data acquisition module systematically collects four types of real-time data: passenger flow count, traffic operation, business activities, and external environment, and constructs a unified spatiotemporal data sequence to support prediction;
[0067] Step S3: The feature generation and fusion module fuses temporal, operational, and environmental features for each node, and mines the behavioral relationships between nodes based on historical passenger flow collaboration patterns to form enhanced network feature inputs;
[0068] Step S4: The graph neural network prediction module constructs a neural network model that integrates graph structure and temporal dependence. By dynamically adjusting the connection weights, it learns the passenger flow propagation and detour patterns under control measures, and completes the model training.
[0069] Step S5: The results application and early warning module obtains future passenger flow forecasts based on real-time data; formulates diversion plans and simulates and compares the network passenger flow impact of different plans.
[0070] Example 2: A business district customer flow prediction method based on graph neural networks as described in Example 1, the specific steps of which are as follows:
[0071] Step S1: The business district network construction module abstracts the physical structure of transportation hubs and commercial spaces into a dynamic network, and establishes connection relationships based on actual travel paths to form a business district network diagram that represents the physical spatial structure.
[0072] Specifically, within the transportation hub-type business district, all key physical locations that have a significant impact on the gathering and distribution of people are selected as network nodes, including: transportation nodes (such as the concourse, platform, entrances and exits, and transfer passages of subway / high-speed rail stations), commercial nodes (the main entrances of shopping malls, the core entrances of each floor, and the underground commercial street nodes that are directly connected to the transportation hub), and connecting nodes (the turning points of passages connecting transportation and commerce, the overpass plaza, and important pedestrian intersections in the surrounding area).
[0073] Based on the actual building structure and walkable paths, establish connections between nodes: when there is a direct path between two nodes (such as passageways, stairs, or corridors), establish a connecting edge.
[0074] Add attribute annotations to the connecting edges, including key parameters such as traffic direction (two-way / one-way), physical length, walking reference time, effective passage width, and maximum design throughput (people / hour);
[0075] Model special access restrictions: such as the fixed running direction of escalators, the normally open / normally closed status of turnstiles, and the daily opening and closing status of fire doors;
[0076] The obtained nodes and connections are stored in a graph data structure to form a static basic network; an entry point for modifying node connections is reserved to support dynamic updates of connection attributes based on temporary control measures (such as activity barriers or construction closures); finally, network connectivity is manually verified to ensure that the abstract paths match the actual main pedestrian channels.
[0077] Step S2: The data acquisition module systematically collects four types of real-time data: passenger flow count, traffic operation, business activities, and external environment, and constructs a unified spatiotemporal data sequence to support prediction;
[0078] Specifically, the hardware components of the data acquisition module include: installing binocular cameras (with passenger flow statistics function) at all node locations in the graph structure, as well as at passages, entrances / exits, escalator entrances, etc., to collect real-time data on the number of people passing through, density, etc.; installing IoT data collectors on turnstiles, escalators, electronic guidance screens, etc., to obtain real-time equipment operating status (on / off, direction, speed); connecting with the subway / bus operation system to obtain real-time train arrival and departure times and scheduling plans; connecting with the commercial MIS / POS system to obtain promotional activity periods and the operating status of key stores; and connecting with the meteorological bureau API to obtain weather data and connecting with the holiday calendar to obtain holiday markers.
[0079] The central processing unit collects four types of data streams at a preset frequency (e.g., every 1-5 minutes), including: passenger flow counting data: number of people entering and exiting at each node, two-way traffic flow in the passage, and real-time density value of the area; traffic operation data: train arrival and departure times, entrance and exit opening and closing status, escalator / turnstile operation mode, and temporary flow restriction announcements; commercial activity data: start and end times of large-scale promotions, location of the main entrance to the event, and information on temporary store closures; and external environment data: weather type (sunny / rainy / snowy), temperature, holiday signs, and schedules of large-scale city-level events.
[0080] Next, the collected data is preprocessed. The specific preprocessing operations include: time alignment: aligning the timestamps of each data stream to the standard time axis and slicing and aggregating them according to fixed time slices; spatial mapping: binding each data record to a corresponding network node number, which corresponds to the defined network structure nodes; format standardization: converting heterogeneous data into a unified format, such as encoding "light rain" as a weather level parameter and "escalator up" as a direction code.
[0081] Outliers and missing values in the collected data are processed: instantaneous spikes or zero values caused by equipment failure and data from brief communication interruptions are identified and corrected by interpolation of adjacent time slices; finally, the standardized data is serialized and output: structured data packets are generated in time slice order, each data packet contains the feature vectors of all nodes at the current time, forming a continuous and complete spatiotemporal data sequence.
[0082] Step S3: The feature generation and fusion module fuses temporal, operational, and environmental features for each node, and mines the behavioral relationships between nodes based on historical passenger flow collaboration patterns to form enhanced network feature inputs;
[0083] Specifically, for each node, a time slice is formed with a data collection interval (5 minutes). All real-time data is aggregated and allocated to the corresponding time slice. For example, the train arrival information at 8:02 and its corresponding passenger flow information are allocated to the time slice at 8:05.
[0084] Next, a feature vector is generated for each node in real time. This vector is concatenated with the node's name, the node's current passenger flow, the simple difference from the previous moment, the names of all its connecting edges, and the real-time events and environmental states (such as train arrival, weather, and activities) expressed through one-hot encoding, thus forming a node feature vector that only reflects the current instantaneous state.
[0085] Furthermore, within the feature vectors of nodes in the same time slice, inherent correlation patterns between different feature values are mined. Specifically, by analyzing the feature vector of each node itself, stable correspondences can be found between its different feature values. The practical significance of this relationship is as follows: when the encoding of the event feature "train arrival" is activated, the "passenger flow" feature value of that node is usually at a high level; similarly, when the "commercial district activity" feature is activated, the "passenger flow" feature value also shows a high level. By identifying and quantifying the co-occurrence and co-variation patterns between this event encoding and passenger flow values within the feature vector, it can be abstracted into a measurable "behavioral pattern." If multiple nodes exhibit highly similar "behavioral patterns" within the same or consecutive time slices... If the nodes exhibit highly consistent passenger flow response patterns to the same events (such as train arrivals or event organization), the system determines that there is a behavioral association between these nodes. This association directly stems from the real-time analysis of the intrinsic structure of the feature vectors, realizing the mapping from the node's own state to the behavioral similarity between nodes. Based on this, dynamic behavioral association connections are added between nodes. A behavioral association strength vector is generated for each node, which quantifies the similarity of the node's behavioral pattern with all other nodes in the network at the current moment. This vector, as a new feature dimension, is concatenated after the node's original feature vector to form the final enhanced feature input.
[0086] Step S4: The graph neural network prediction module constructs a neural network model that integrates graph structure and temporal dependence. By dynamically adjusting the connection weights, it learns the passenger flow propagation and detour patterns under control measures, and completes the model training.
[0087] Specifically, a neural network model that integrates graph structure and temporal dependencies is constructed. After extracting the constructed business district network graph structure and binding the corresponding data of the nodes, a graph neural network (GNN) is used to combine the graph structure with temporal features to learn the propagation rules between nodes. Through the graph convolutional layer (GCN layer), node features are transmitted between its neighboring nodes, thereby capturing the relationship with neighboring nodes.
[0088] During model learning, the connection weights are dynamically adjusted through the self-attention mechanism of the graph neural network. The graph neural network dynamically adjusts the connection weights between nodes through training. Based on the learning results, the model identifies the connections that have a greater impact on changes in customer flow, and then adjusts the weights of each node according to these impacts. In addition, in actual business district operations, some control measures (such as traffic closures, event barriers, etc.) may affect customer flow, which is reflected in changes in one-hot encoding in the feature input. The graph neural network can adjust the network connection weights according to the input of these control measures, so that the model can simulate and predict changes in customer flow under different conditions.
[0089] During the learning process, the focus is on analyzing the relationship between passenger flow direction and various characteristics. Based on extensive data training, the system learns the direction of passenger flow when certain characteristics exist at different nodes. For example, when an event is held at a certain node, the passenger flow at other nodes shows a pattern of large inflows towards that event node and relatively small outflows. Similarly, the system analyzes the patterns of passenger flow relationships in different events. In addition, individual pattern calculations are performed for each node. For example, even if an event is held at a certain node, the passenger flow will not be large.
[0090] Once the model is built, based on the propagation relationship between multiple nodes, the model can analyze the changes in passenger flow at the nodes and, by combining various event information in the feature input, analyze the patterns of passenger flow changes between nodes.
[0091] Step S5; The results application and early warning module obtains future passenger flow forecasts based on real-time data; formulates diversion plans and simulates and compares the network passenger flow impact of different plans;
[0092] Specifically, the data acquired at the current node is bound to that node, and the resulting data is input into a graph neural network model for prediction. Based on the patterns learned from long-term training, the model predicts future changes in passenger flow from different times and event codes. During the simulation, all passenger flow follows a preset direction, and the changes in passenger flow at different nodes are analyzed in conjunction with the learned passenger flow patterns. For example, if there is a bidirectional elevator between two nodes, the passenger flow changes are predicted based on the current time and passenger flow patterns learned from multiple event codes. This prediction is then applied to the entire model to obtain changes in passenger flow over a future period. When events occur at certain nodes, the attraction of that node to other nodes is assessed based on its historical event patterns. Using the network's entry point as a basis, changes in passenger flow are predicted, simulating the impact of the event on the business district's operations. In the simulation, different passenger flow capacity is set for all nodes, based on the actual space of the node. The specific setting method can be determined by the model based on historical data. Historical node data records the maximum number of customers, calculated as 80%. When the customer flow at a node exceeds its set capacity during the simulation, it is recorded as a congestion event, and the number of congestions is recorded. The simulation fully simulates the operational results over a preset time period. First, the total customer flow at each moment is counted. Then, the congestion level is assessed by combining the number of congestion events and the node's capacity. If a node in the simulated business district network has a congestion event count exceeding 60% of the total data collected, and its capacity ranks within the top 10% of all nodes (the specific threshold is adjusted adaptively based on the actual size of the business district), it is considered severely congested under the current operating mode. Similarly, if at least five nodes have congestion events exceeding 60% of the total data collected, it is also considered severely congested under the current operating mode. The simulation results are then sent to the operations system for staff reference.
[0093] When severe congestion events are predicted in the module's results, multiple traffic diversion schemes are developed and simulated, and the optimal scheme is selected. The specific methods and content for developing these traffic diversion schemes are not fixed; they are developed based on the node connection characteristics of the business district and the actual business district management plan. Here are a few examples of traffic diversion schemes:
[0094] Simulation Scheme 1: Prioritize selecting the most congested node from the simulation results (the node with the most congestion events and the largest passenger flow capacity). Centered on this node, analyze it based on the graph neural network model, traverse the connection paths of this node, and find multiple bidirectional connection edges that can completely change the flow direction. Modify the flow direction of these multiple bidirectional connection edges into multiple unidirectional connection edges, while ensuring the connection relationship between nodes to prevent the formation of closed islands in the business district network graph. In practice, this means modifying the direction of the passage between multiple layers of space in the business district to form a basic unidirectional flow connection path between business district nodes, thereby improving traffic efficiency.
[0095] Simulation Scheme 2: Identify the source node with the largest passenger flow in the simulation results. For each node, calculate the difference between its total outflow and total inflow passenger flow throughout the entire simulation period. For nodes with consistently positive and large outflow values, temporarily reduce the passenger flow release rate of that node. In the model, this is represented by directly modifying the passage capacity parameters between nodes. In practical terms, it means setting up security personnel at the node to restrict the speed at which passengers enter the business district.
[0096] After developing several traffic diversion plans, the parameters in the business district network diagram are modified, and the operational results are simulated again. From the various simulation results, the plan with the fewest congestion events is selected, and the plan and prediction results are sent to the operation management system for staff reference.
[0097] like Figure 2 The comparison chart of the effects of this invention and traditional methods is shown in the figure. The black bars in the figure represent the effects of this invention, and the gray bars represent the effects of traditional methods. Integrating multi-source data effectively improves the accuracy of prediction, and the running parameters of the model can be dynamically modified, which improves the accuracy of the method in responding to sudden events. By simulating in the system, the efficiency of resource allocation is greatly improved.
[0098] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for predicting customer traffic in a business district based on graph neural networks, characterized in that: The method is based on an independent system, including a business district network construction module, a data acquisition module, a feature generation and fusion module, a graph neural network prediction module, and a result application and early warning module; The specific implementation steps are as follows: Step S1: The business district network construction module abstracts the physical structure of transportation hubs and commercial spaces into a dynamic network, and establishes connection relationships based on actual travel paths to form a business district network diagram that represents the physical spatial structure. Step S2: The data acquisition module systematically collects four types of real-time data: passenger flow count, traffic operation, business activities, and external environment, and constructs a unified spatiotemporal data sequence to support prediction; Step S3: The feature generation and fusion module fuses temporal, operational, and environmental features for each node, and mines behavioral relationships between nodes based on historical passenger flow collaboration patterns to form enhanced network feature inputs, specifically including: For each node, time slices are formed based on data acquisition intervals, and all real-time data is aggregated and allocated to the corresponding time slices; Next, a feature vector is generated for each node in real time. This vector is concatenated with the node's name, the node's current passenger flow, the simple difference from the previous moment, the names of all its connecting edges, and the real-time events and environmental state expressed by one-hot encoding, thus forming a node feature vector that only reflects the current instantaneous state. By analyzing the feature vector of each node, a stable correspondence can be found between its different feature values; Identify and quantify the co-occurrence and co-variation patterns between event encoding and passenger flow values within feature vectors; extract the correlation between nodes based on the correlation of changes, and generate a behavioral correlation strength vector for each node, which quantifies the similarity of the behavioral patterns of the node with all other nodes in the network at the current moment. This vector is used as a new feature dimension and concatenated to the original feature vector of the node to form the final enhanced feature input; Step S4: The graph neural network prediction module constructs a neural network model that integrates graph structure and temporal dependence. By dynamically adjusting the connection weights, it learns the passenger flow propagation and detour patterns under control measures, and completes the model training. Step S5: The results application and early warning module obtains future passenger flow forecasts based on real-time data; formulates diversion plans and simulates and compares the network passenger flow impact of different plans.
2. The method for predicting customer traffic in a business district based on a graph neural network according to claim 1, characterized in that: The independent system specifically includes: Business district network construction module; based on the geographical conditions of business districts, it abstracts and constructs a virtual business district network; Data acquisition module: Integrates multiple data acquisition systems to collect multi-source data, including passenger flow data, traffic operation data, business district operation data, and external environment data; Feature generation and fusion module: Fusion of multi-source data at sub-nodes, calculation of data correlation, and formation of feature input that takes into account both spatial relationships and behavioral connections; Graph Neural Network Prediction Module: Based on the virtual business district network and feature input, a graph neural network model is constructed; deep learning algorithms are used to learn the behavioral patterns of the crowd. Results Application and Early Warning Module: Predicts the operation of the business district based on the behavior patterns of the crowd; formulates traffic diversion plans based on the prediction results; simulates the operation results based on the traffic diversion plans and compares the effects of multiple plans.
3. The method for predicting customer flow in a business district based on a graph neural network according to claim 1, characterized in that: The business district network map that forms the physical spatial structure specifically includes: Within a transportation hub-type business district, all key physical locations that have a significant impact on the gathering and dispersal of people are selected as network nodes, specifically including: transportation nodes, commercial nodes, and connecting nodes. Based on the actual building structure and pedestrian feasible paths, establish connection relationships between nodes: label the connection edges with attributes; and model special access restrictions. The obtained nodes and connections are stored in a graph data structure to form a static basic network; and an entry point for modifying node connections is reserved to support dynamic updates of connection attributes based on temporary control measures.
4. The method for predicting customer flow in a business district based on a graph neural network according to claim 1, characterized in that: The data acquisition module systematically collects four types of real-time data: passenger flow count, traffic operation, commercial activities, and external environment, constructing a unified spatiotemporal data sequence to support prediction. Specifically, this includes: The hardware components of the data acquisition module include: installing binocular cameras on nodes and connecting channels in the graph structure; adding IoT collectors to adjustable devices to obtain real-time device operating status; interfacing with the public transportation operation system to obtain real-time train arrival and departure times and scheduling plans; and associating with relevant systems to obtain data affecting passenger flow. The central processing unit collects four types of data streams at a preset frequency, including: passenger flow count data, traffic operation data, business activity data, and external environment data; and preprocesses the collected data.
5. The method for predicting customer traffic in a business district based on a graph neural network according to claim 4, characterized in that: The preprocessing of the collected data specifically includes: Preprocessing steps include: time alignment, spatial mapping, and format normalization; Subsequently, outliers and missing values in the collected data are processed and filled using interpolation of adjacent time slices. Finally, the standardized data is serialized and output: structured data packets are generated in time slice order, and each data packet contains the feature vectors of all nodes at the current time, forming a continuous and complete spatiotemporal data sequence.
6. The method for predicting customer traffic in a business district based on a graph neural network according to claim 1, characterized in that: The graph neural network prediction module constructs a neural network model that integrates graph structure and temporal dependence. It learns the passenger flow propagation and detour patterns under traffic control measures by dynamically adjusting connection weights. The model training specifically includes: Specifically, a neural network model that integrates graph structure and temporal dependency is constructed. Graph neural networks (GNNs) are used to combine graph structure and temporal features to learn the propagation rules between nodes. Through graph convolutional layers, node features are transmitted between their neighboring nodes, thereby capturing the relationship with neighboring nodes. During the model learning process, the connection weights are dynamically adjusted through the self-attention mechanism of the graph neural network. The graph neural network dynamically adjusts the connection weights between nodes through training. Based on the learning results, the model identifies the connection relationships that have a greater impact on changes in customer traffic, and adjusts the weights of each node according to these impacts. The model is also trained by combining the one-hot encoding of events in actual business district operations. During the learning process, the focus is on analyzing the relationship between passenger flow direction and various characteristics. Based on a large amount of data training, the flow direction of passenger flow is learned when certain characteristics exist at different nodes. Individual pattern calculations are performed for each node, and finally, the model is constructed.
7. The method for predicting customer traffic in a business district based on a graph neural network according to claim 1, characterized in that: The results application and early warning module obtains future passenger flow predictions based on real-time data, specifically including: The data obtained from the current node is bound to the node, and the obtained data is input into the graph neural network model for prediction. Based on the patterns obtained from long-term training of the model, the changes in passenger flow in the future are predicted from different time and event codes. During the simulation, all passenger flows follow the preset passenger flow direction. Combined with the learned passenger flow patterns, the changes in passenger flow at different nodes are analyzed and predicted in the entire model to obtain the changes in passenger flow over a future period. When events occur at certain nodes, the attraction of that node to other nodes is assessed based on the historical event patterns of that node. Based on the network's entry point, changes in passenger flow are predicted, and the impact of the event on the business district's operation is simulated. Different passenger flow carrying capacities are set for all nodes in the simulation. Passenger capacity is set based on the actual space of the node. The specific setting method is calculated by the model based on the maximum passenger flow recorded by the historical node data, using 80% as the limit. When the passenger flow at a certain node exceeds the passenger flow capacity set for that node during the simulation, it is recorded as a congestion event at the current node. The number of congestion events is recorded, and the operation results are fully simulated according to the preset time. The total number of passenger flows generated at each moment during the process is counted. The simulation results are used to assess the degree of congestion. A comprehensive assessment of the congestion level is conducted based on the number of times nodes are congested and the passenger flow capacity of nodes in the simulation. When the number of times nodes in the simulated business district network map are congested exceeds a preset threshold, it is determined that the business district is severely congested under the current operating mode. The simulation results are then sent to the operation system for staff reference.
8. The method for predicting customer flow in a business district based on a graph neural network according to claim 1, characterized in that: The specific steps involved in developing traffic diversion plans and simulating and comparing the impact of different plans on network passenger flow include: When severe congestion events are predicted in the module's results, multiple diversion schemes are developed and simulated to select the optimal scheme. The specific diversion scheme development method and content are not fixed, and the development method is based on the node connection characteristics of the business district and the actual business district management scheme. After developing several traffic diversion plans, the parameters in the business district network diagram are modified, and the operational results are simulated. From the simulation results, the plan with the fewest congestion events is selected, and the plan and prediction results are sent to the operation management system for staff reference.