Multi-modal flow cooperative control method for comprehensive transportation junction

Abstract

The invention relates to the technical field of energy traffic systems, discloses a multi-modal flow cooperative control method for a comprehensive traffic hub, and aims to solve the problems of information isolated island, response lag and regulation fragmentation caused by factor system decoupling in the prior art. The method comprises the following steps: constructing a four-level system architecture consisting of a multi-source sensing layer, a dynamic modeling layer, a collaborative decision-making layer and an execution feedback layer, and establishing a people flow-traffic flow-information flow nonlinear interaction relationship based on a space-time coupling network model and a multi-agent reinforcement learning framework, and a double-loop control mechanism of feedforward disturbance prediction and feedback regulation compensation is fused. Through adoption of the technical scheme, total factor perception, dynamic deduction, collaborative optimization and closed-loop regulation and control of the multi-modal flow in the hub can be realized, the evacuation efficiency, the resource utilization rate and the service consistency are remarkably improved, and various complex scenes such as normal peak, train delay and sudden public events are effectively coped with.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation systems, specifically to a multimodal flow collaborative control method for integrated transportation hubs. Background Technology

[0002] With the accelerated urbanization and continuous improvement of transportation infrastructure in my country, integrated transportation hubs, as physical intersections and functional integration platforms combining various modes of transportation such as railways, highways, aviation, urban rail transit, public transportation, taxis, and slow-moving systems, are increasingly playing a core role in regional transportation networks. These hubs are not only key nodes for the efficient flow of people and goods, but also important carriers for realizing the concepts of "zero-distance transfer" and "seamless connection," directly impacting urban operational efficiency, public travel experience, and the overall resilience of the transportation system. Against this backdrop, how to implement scientific, coordinated, and dynamically responsive collaborative control of the complex and intertwined multimodal traffic flows within hubs—including passenger flow, vehicle flow, information flow, and service resource flow—has become a highly strategic technical direction in the field of modern intelligent transportation systems research.

[0003] However, as the daily passenger flow of large integrated transportation hubs exceeds one million, and passenger travel demands exhibit a trend towards high personalization, real-time nature, and increased uncertainty, the aforementioned discrete control architecture exposes a series of deep-seated structural contradictions at the system level. The reason for this lies in the fact that this model essentially manages multimodal traffic flows as isolated, independent entities, lacking the ability to model and intervene in the dynamic coupling mechanism between pedestrian, vehicle, and information flows. Furthermore, when a sudden disturbance occurs in a sub-region (such as large-scale passenger congestion caused by train delays), its impact rapidly propagates along spatial and temporal chains to adjacent functional areas. However, due to the lag in information sharing and asynchronous response strategies between subsystems, control actions often exhibit temporal mismatches and spatial conflicts, further exacerbating overall operational instability. In addition, existing methods largely rely on empirical thresholds to trigger control commands, making it difficult to adapt to the continuous optimization needs under non-steady-state conditions, resulting in low resource allocation efficiency and even producing a reverse effect of "local optima leading to global deterioration." This performance shift reflects the inherent limitations of traditional architectures in terms of control granularity, response speed, and system collaboration. They are no longer able to meet the comprehensive requirements for the security, smoothness, and service quality of hub operations under high-density, highly interactive, and rapidly changing conditions.

[0004] Therefore, how to construct a method system that can deeply integrate multi-source heterogeneous sensing data, accurately depict the dynamic evolution of multimodal traffic flow, and on this basis realize cross-domain resource collaborative optimization and feedforward feedback integrated regulation, so as to overcome the fundamental problems of response lag, regulation fragmentation and global efficiency decay caused by system decoupling in the existing technology, has become a key challenge and a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] To achieve the aforementioned objectives, this invention provides a multimodal flow collaborative control method for integrated transportation hubs. This method constructs a four-layer system architecture consisting of a multi-source perception layer, a dynamic modeling layer, a collaborative decision-making layer, and an execution feedback layer. This architecture enables comprehensive fusion perception, multi-scale dynamic extrapolation, cross-domain collaborative optimization, and closed-loop precise control of passenger flow, vehicle flow, information flow, and service resource flow within the hub. Based on a spatiotemporal coupled network model and a multi-agent reinforcement learning framework, this method establishes a nonlinear interaction relationship expression between passenger flow, vehicle flow, and information flow. It also introduces a dual-loop control logic combining a feedforward disturbance prediction mechanism and a feedback adjustment and compensation strategy, fundamentally solving the problems of information silos, response lag, and fragmented control caused by factor system decoupling in existing technologies.

[0006] Furthermore, the multi-source sensing layer consists of a heterogeneous sensor array deployed throughout the entire hub space, including but not limited to 3D LiDAR units, millimeter-wave radar modules, infrared thermal imaging camera groups, Wi-Fi probe devices, Bluetooth sniffing nodes, gate card-swiping recording terminals, video structured analysis servers, and mobile communication base station signaling acquisition devices. Each sensing unit is fixedly installed at key access nodes according to a preset spatial topology layout. The 3D LiDAR units are arranged at 1.5-meter intervals on the side walls of the transfer passage, with the scanning plane perpendicular to the direction of travel, a sampling frequency set to 20Hz, and a point cloud resolution of no less than 0.05m. The millimeter-wave radar modules are embedded in the ceiling structure, covering a circular area with a radius of 8 meters, operating at 77GHz, and possessing target detection performance with a velocity resolution better than 0.1m / s. Infrared thermal imaging camera groups are configured every 12 meters along the longitudinal axis of the waiting hall, with a field of view of 90°×68° and a temperature sensitivity ≤0.08℃, used for non-contact human contour extraction and density estimation. The Wi-Fi probe devices and Bluetooth sniffing nodes... Point-to-point deployments are located above security checkpoints, ticket gates, and commercial service areas, with signal reception sensitivity no less than -95dBm and a MAC address capture period of 3 seconds. After hash encryption, the data is uploaded to the edge computing node. The turnstile card-swiping terminal connects to the local controller via an RS-485 bus, with timestamp accuracy synchronized to the millisecond level. The video structured analysis server adopts a distributed cluster architecture, accessing H.265 encoded video streams from all public area HD cameras, executing pedestrian posture recognition, object classification, and abnormal behavior detection algorithms, and outputting structured metadata. The mobile communication base station signaling acquisition device mirrors S1-MME signaling data from the operator's core network interface, parsing user equipment attachment, handover, and deactivation event sequences. All sensing devices are aggregated to the regional edge gateway via industrial Ethernet, using the IEEE 802.1Q standard for VLAN segmentation to ensure that different types of data streams are isolated on the physical link, with end-to-end communication latency not exceeding 50ms.

[0007] Furthermore, the dynamic modeling layer runs on a high-performance computing cluster in the central data center of the hub, and includes a spatiotemporal coupled network generation module, a multimodal flow dynamic evolution equation solver, and a disturbance propagation path tracing unit. The spatiotemporal coupled network generation module divides the hub's physical space into square grid cells with sides of 1.2 meters, forming a two-dimensional discretized spatial base. Each grid cell is assigned a unique geocode and aligned with the BIM model coordinates. Based on this, the grid type is labeled according to functional zoning attributes, including pedestrian access areas, queuing areas, static waiting areas, and prohibited areas. Furthermore, combined with actual reachable path constraints, a directed graph is constructed. vertex set For all accessible grid center coordinates, edge set Represents the probability of movement between adjacent grid cells, edge weight. Free speed of walking Local density and obstacle repulsive potential energy The decision is made jointly, and the specific expression is as follows: in The value is taken as 1.3 m / s, and the coefficient is... =0.35, =0.6; Calculated based on the distance d from the nearest obstacle, when When <0.8m, Otherwise, the value is zero. The network updates its topology weights every 100ms to reflect real-time congestion changes.

[0008] Furthermore, the multimodal flow dynamic evolution equation solver employs an improved macroscopic continuum model to describe the pedestrian flow density field. The evolutionary process, in which Inside and For spatial location, Represents the coordinates of a point on a two-dimensional plane. This is a local average velocity vector, which describes the average velocity and direction of the pedestrian flow at a corresponding spatial location and time. For source and sink items; Where the fluid dynamic viscosity is Pressure gradient term according to calculate, =0.1, =2.0; Driving force Pointing to the nearest target exit, interaction force The social force model is used to calculate the resultant force of repulsion between individuals and attraction within a group. The equations are solved using the finite volume method in a GPU-accelerated environment, and the boundary conditions are dynamically set based on real-time gate passage data.

[0009] Furthermore, the disturbance propagation path tracing unit constructs a passenger flow disturbance impact diffusion model based on the Markov transition matrix M, and the matrix elements... Indicates the time from the region Transfer to the region The probability is determined by a combination of historical card-swipe data and real-time video tracking results. When abnormal backlog is detected in a certain area k (defined as the current density exceeding a threshold),... =4.0 people / m 2If the disturbance persists for more than 3 minutes, the perturbation forward propagation mechanism is activated to calculate the expected arrival distribution within the next 15 minutes. ,exist In the state of time Can be derived from Markov transition matrix After multiplying by k times, and... This yields data that reflects the linear evolution of the system state over time and marks the set of affected downstream regions. At the same time, a correlation model for vehicle connection delays will be established to link railway train delays. As input variables, they are mapped to the incremental demand for taxi dispatching through an empirical regression function. ,in =8 vehicles / minute =15 vehicles, used to adjust the departure rhythm of the storage pool in advance.

[0010] Furthermore, the collaborative decision-making layer integrates a multi-agent reinforcement learning optimization engine, and its main architecture includes four functional agents: passenger flow management agent. Transportation connection agency Information service agency With emergency response agents Each agent shares a unified state space S and action space A, where the state vector... It consists of the following dimensions: real-time density of each region Queue length Remaining capacity Vehicle carrying capacity Information dissemination coverage Number of police officers on duty Action vector Includes: Guidance instruction combination G, signal timing scheme Vehicle dispatching orders Broadcast content template B; Police force deployment instructions Each agent independently trains its policy function using the Deep Deterministic Policy Gradient (DDPG) algorithm. Its reward function R is designed as follows: for for for for in, The target comfort density is 1.8 people / m² 2 , Congestion Index (when >3.5 hours count as 1, otherwise 0). Average passenger waiting time Vehicle empty-run rate To delay the release of information, To account for the number of information conflicts from multiple channels, The percentage of the population not covered. To mitigate the impact of unforeseen events on the speed of expansion, For the first intervention response time, Capacity loss due to service disruptions. Weighting coefficient. to The values ​​are taken as [0.3, 0.25, 0.15, 0.3, 0.2, 0.1, 0.4, 0.3, 0.3, 0.5, 0.3, 0.2], and are then normalized before being used in the joint utility calculation.

[0011] Furthermore, the multi-agent system adopts a centralized training-distributed execution (CTDE) model. During the offline phase, it utilizes historical operational data to generate millions of simulation scenario samples, which are then input into a shared experience replay buffer. During online operation, each agent bases its data on local observations... Independent selection of actions And through the Nash equilibrium solver, conflicting actions are coordinated, ultimately outputting a consistent set of control instructions. For example, when It is recommended to close the entrance to a certain transfer passage to alleviate downstream pressure. When it is required that the passage be kept clear to allow shuttle buses to pass, the coordination module calculates the utility loss ratio for both parties. Loss less If the gain is positive, then adopt it. The proposal triggers an alternative route recommendation mechanism.

[0012] Furthermore, the execution feedback layer consists of a distributed execution terminal network, including a variable message sign control system, an automatic gate group control unit, a traffic signal priority controller, a wireless broadcast push engine, and a mobile terminal message service platform. The variable message sign control system is connected to LED display arrays deployed at the intersections of main passageways on each floor, with a display size of no less than 1.8m × 1.2m and a brightness ≥ 5000 cd / m². 2The system features a 60Hz refresh rate, receives guidance path vector instructions from the collaborative decision-making layer, and dynamically displays the optimal evacuation direction using the ISO 20511 standard graphic symbol system. The automatic gate group control unit connects to all ticket checking equipment via the Modbus TCP protocol, supports remote batch setting of access permissions, and can simultaneously lock or open 20 gates in a single area within 3 seconds. The traffic signal priority controller connects to the city's traffic signal platform, implementing green wave adjustments at key intersections around the hub. When a large number of departing passengers need to be quickly evacuated, the green light duration is extended by up to 45 seconds towards the bus lane. The wireless broadcast push engine connects to speaker nodes distributed across various functional areas via an IP network, playing differentiated voice prompts grouped by area number, with the volume automatically adjusted to a background noise level of +10dB. The mobile terminal message service platform connects to mainstream map application SDKs via an HTTPS interface, pushing personalized travel suggestions to registered users' mobile apps, including estimated walking time, recommended elevator numbers, the departure time of the next shuttle bus, and seat availability information.

[0013] Furthermore, the method establishes a dual-loop closed-loop control mechanism, with the outer loop being a feedforward prediction loop and the inner loop being a real-time feedback loop. The feedforward prediction loop operates every 5 minutes, acquiring external input disturbances based on train timetables, flight dynamic databases, and weather forecast APIs. It then uses a spatiotemporal coupled network model to predict passenger flow load in various regions within the next 30 minutes, generating pre-control plans in advance and sending them to the execution terminal for pre-loading. The real-time feedback loop samples every 500 milliseconds, collecting the latest data streams from each sensing node. After aggregation by edge computing nodes, the data is sent to the dynamic modeling layer to recalculate the current state. If the actual state deviates from the predicted trajectory by more than the tolerance threshold δ=15%, an emergency re-optimization process is immediately triggered, calling the collaborative decision-making layer to recalculate the optimal strategy and overwrite the original instructions.

[0014] Furthermore, the system is configured with an adaptive degradation module for abnormal operating conditions. When some sensing devices fail or communication links are interrupted, a backup inference rule base is activated. For example, in the event of video surveillance interruption, passenger flow distribution is analyzed solely based on Wi-Fi probe and card swipe data, and a Kalman filter is used to fuse historical trends and observations from neighboring areas to estimate missing data. When an edge computing node crashes, the system switches to a local cache model to continue executing the previous effective control strategy and reduces the instruction update frequency to once every 2 minutes until the main system recovers.

[0015] Furthermore, the method defines five typical control scenarios and their corresponding strategies during implementation: The first scenario is peak-hour congestion management, in which a multi-path diversion mechanism is activated, using variable message signs to guide passenger flow evenly to multiple parallel transfer corridors to avoid overloading of a single channel; the second scenario is large-scale train delays, in which an emergency response plan is activated, coordinating with the railway department to issue delay notices, simultaneously increasing the number of temporary shuttle buses, adjusting the priority of taxi dispatching in underground parking garages, and initiating tiered early warning broadcasts in the waiting area; the third scenario is sudden public events (such as fire alarms), in which rigid [measures] are implemented. The isolation strategy includes automatically locking nearby turnstiles, shutting down air conditioning and ventilation systems to prevent smoke from spreading, illuminating emergency directional signs, and guiding personnel to evacuate along predetermined safe routes. The fourth category is the impact of severe weather. When meteorological monitoring data shows that a red rainstorm warning is in effect, indoor temporary shelter areas will be opened in advance, anti-slip mats and rain gear distribution points will be mobilized, and the operating time of the last subway train will be extended by 30 minutes. The fifth category is the guarantee for major events. For foreseeable large passenger flows such as after concerts, tidal channels will be set up in advance, the layout of railings will be changed to form a one-way circulation flow, and mobile patrol robots will be deployed to strengthen order maintenance.

[0016] Furthermore, the system is equipped with a 3D visualization monitoring interface, running on the multi-screen splicing display system of the command center. It uses WebGL technology to render a real-world digital twin model of the hub, overlaying real-time heat maps, vector streamline diagrams, and resource distribution icons. It supports mouse-through querying of detailed parameters for any area and allows zooming and panning of the viewpoint via touchscreen gestures. The underlying data of the monitoring interface is directly output from the dynamic modeling layer, with refresh latency strictly controlled within 800ms to ensure the timeliness of situational awareness.

[0017] Furthermore, the method adheres to Level 3 security requirements at the data security level. All sensor data is encrypted using the AES-256 algorithm before transmission and stored in a privacy-protected database conforming to the GB / T 35273-2020 standard. Personal identification information is immediately anonymized, retaining only anonymized trajectory data used for statistical analysis. System access employs an RBAC (Restricted Access Control) model, granting different operators access to corresponding functional modules, and requiring dual verification and authentication for critical commands.

[0018] Furthermore, the collaborative decision-making layer optimization engine regularly iterates and updates its strategies. Each month, it automatically extracts the previous cycle's operational logs to generate a training set, which includes three types of scenario samples: normal operating conditions, disturbance operating conditions, and sudden events. After data cleaning and labeling, the samples are input into the offline training pipeline. The new strategy model undergoes no less than 1,000 hours of stress testing in a simulation environment to verify its stability and robustness under extreme loads. Only after meeting the requirements is it allowed to be deployed and go online to replace the old version.

[0019] Furthermore, the system supports a cross-hub collaborative mode. When the distance between two adjacent integrated transportation hubs is less than 5 kilometers, core status parameters can be exchanged through a dedicated fiber optic link to establish a joint scheduling and coordination mechanism. For example, if a failure at Hub A causes a surge in outbound passenger flow while Hub B currently has a low load, a cross-hub diversion plan is automatically activated. This plan guides some passengers to Hub B via a regional navigation system to transfer to rail transit, alleviating local pressure.

[0020] Furthermore, the method adopts a modular design concept in hardware deployment, with each functional unit packaged as a standardized rack-mount device, conforming to the 19-inch rack specification. Power input supports AC 220V±10% 50Hz, with total power consumption not exceeding 800W. Cooling is achieved through forced air cooling, and noise levels are below 55dB(A). All equipment enclosures have a protection rating of at least IP30, and critical servers are configured with dual power supply redundancy and RAID5 disk arrays to ensure uninterrupted 24 / 7 operation.

[0021] Furthermore, the software system is developed based on a microservice architecture and uses Docker containerization to encapsulate various functional modules, including perception data access service, spatiotemporal modeling and computing service, multi-agent decision service, instruction distribution service and log auditing service. The services communicate efficiently with each other through the gRPC protocol. Service registration and discovery are managed by the Consul cluster. Configuration files are centrally stored in the ETCD key-value database, supporting hot updates without restarting the process.

[0022] Furthermore, the method employs the IEEE 1588 Precision Time Protocol (PTP) for time synchronization, deploying a master clock source in the network core switch and embedding PTP slave clock chips in all edge nodes, achieving a system-wide time error of less than ±10 microseconds. This ensures precise alignment of multi-source data on the time axis, providing a reliable foundation for subsequent correlation analysis.

[0023] Furthermore, the system includes an online performance evaluation module that calculates three core KPIs in real time: average evacuation efficiency. ,in For the first The actual walking distance of each passenger. Its travel time, where n is the total number of statistical samples; resource utilization rate This refers to the number of shuttle vehicles already in use. Total available quantity; service consistency The number of cases where information dissemination and on-site guidance are consistent. This represents the total number of cases. The above metrics are calculated every 10 minutes, and a trend curve is plotted for management review.

[0024] Furthermore, the method performs a full system calibration process during the initialization phase, which includes three main steps: The first step is spatial calibration, which uses a total station to measure the installation coordinates of all sensors and performs ICP registration with the BIM model to ensure that the deviation between the physical coordinate system and the digital model is less than 3 cm; the second step is parameter tuning, which selects typical passenger flow scenarios to collect benchmark data and optimizes the unknown parameters in the social force model through a genetic algorithm to minimize the Hausdorff distance between its simulation output and the measured trajectory; the third step is strategy pre-playing, which imports historical large passenger flow event data to verify whether the system can correctly identify bottlenecks and generate reasonable intervention plans. If the preset effect threshold is reached in three consecutive tests, the system is deemed to be ready for online operation.

[0025] Furthermore, the variable message sign control system in the execution feedback layer has a built-in font rendering engine that supports the GB18030 encoding of simplified Chinese characters, with a character height of no less than 200 pixels to ensure clear visibility within a 15-meter distance. The displayed content uses a red, yellow, and green three-color tiered warning system: green indicates normal traffic flow, yellow indicates mild congestion and suggests detouring, and red indicates severe congestion and prohibits entry. The color transition thresholds are as follows: =2.5 people / m 2 and =4.0 people / m 2 .

[0026] Furthermore, the automatic gate group control unit is equipped with a heartbeat detection mechanism, which sends a status message to the central server every 2 seconds, including the device number, current mode, passage count and fault code; once no response is received for 3 consecutive times, it is determined that the communication is interrupted, automatically switches to the local cache strategy mode, continues to execute the last valid instruction, and reports alarm information to the operation and maintenance platform.

[0027] Furthermore, a dedicated secure channel is established between the traffic signal priority controller and the city traffic management center, and the control commands are encrypted using the SM4 national cryptographic algorithm to prevent illegal tampering. Signal adjustment requests must be accompanied by a digital signature and timestamp, and can only be executed after the receiving end verifies them. A confirmation receipt is returned after each adjustment, forming a complete operation audit chain.

[0028] Furthermore, the wireless broadcast push engine has a built-in speech synthesis module that uses a neural network model based on the Tacotron 2 architecture to convert text commands into natural and fluent Mandarin speech in real time. The speaker's voiceprint features are fixed as a standard broadcast voice of a middle-aged man, the speech rate is set to 260 words per minute, and the pause intervals conform to the semantic rhythm rules of Chinese, ensuring that the information is clearly and easily understood.

[0029] Furthermore, the mobile terminal messaging service platform uses the OAuth 2.0 protocol to authenticate user identities and only provides personalized push notifications to users who have completed real-name authentication and authorized location services; the message push frequency is limited to no more than 3 messages per hour to avoid information harassment, and users can unsubscribe from all notifications with one click through the APP interface.

[0030] Furthermore, the three-dimensional visualization monitoring interface provides a historical playback function, which supports selecting any past time point to reconstruct the passenger flow distribution at that time. The playback speed can be adjusted between 0.5 times and 8 times, which is convenient for post-event analysis of the evolution of events and identification of management weaknesses.

[0031] Furthermore, the abnormal operating condition adaptive degradation module has a built-in health assessment unit that continuously monitors the operating status of each subsystem. When it detects that the CPU load has been continuously exceeding 85% for more than 5 minutes, or the memory usage rate is higher than 90%, it automatically triggers the resource cleanup program, suspends non-critical background tasks, and ensures that the core control logic is executed with priority.

[0032] Furthermore, the method establishes a delay level response mechanism when dealing with train delay scenarios: Level 1 response is applicable to delays of 5-15 minutes, and information broadcast reminders are initiated; Level 2 response is applicable to delays of 16-30 minutes, and connecting transport capacity is increased by 15%; Level 3 response is applicable to delays of 31-60 minutes, and backup waiting areas are opened and drinking water is provided; Level 4 response is applicable to delays of more than 60 minutes, and public transportation is coordinated to extend its operation, and a large-scale evacuation plan is initiated.

[0033] Furthermore, before formal commissioning, the system must undergo a trial operation phase of no less than 30 days. During this period, all control commands and system response data will be recorded throughout the process. An independent performance evaluation will be conducted by a third-party organization, focusing on verifying the effectiveness, stability, and security of the multimodal flow collaborative control. Only after the evaluation report is deemed satisfactory can the system be transferred to normal operation mode. Compared with the prior art, the beneficial effects achieved by the present invention are: By constructing a four-layer system architecture consisting of a multi-source perception layer, a dynamic modeling layer, a collaborative decision-making layer, and an execution feedback layer, and based on a spatiotemporal coupled network model and a multi-agent reinforcement learning framework, a nonlinear interaction relationship between pedestrian flow, vehicle flow, and information flow is established, and a dual-loop control mechanism integrating feedforward disturbance prediction and feedback adjustment compensation is integrated. By adopting the above technical solution, this application can achieve full-element perception, dynamic simulation, collaborative optimization, and closed-loop control of multimodal flows within the hub, significantly improving evacuation efficiency, resource utilization, and service consistency, and effectively coping with various complex scenarios such as normal peak hours, train delays, and sudden public events. Attached Figure Description

[0034] Figure 1This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the hub space discretized grid and directed graph topology constructed by the spatiotemporal coupled network generation module in this invention; Figure 3 This is a schematic diagram of the agent interaction and decision-making process of the multi-agent reinforcement learning collaborative decision-making layer of the present invention; Figure 4 This is a flowchart illustrating the timing coordination relationship between the feedforward prediction loop and the real-time feedback loop in the dual-loop closed-loop control mechanism of this invention. Figure 5 This is a schematic diagram showing the digital twin model and multi-dimensional situation overlay display of the three-dimensional visualization monitoring interface of the present invention.

[0035] The system comprises: 1. Multi-source sensing layer; 2. Dynamic modeling layer; 3. Collaborative decision-making layer; 4. Execution feedback layer; 5. Spatiotemporal coupled network generation module; 6. Multimodal flow dynamic evolution equation solver; 7. Disturbance propagation path tracing unit; 8. Passenger flow management agent. 9. Transportation shuttle agency 10. Information service agency 11. Emergency Response Agency ; 12. Feedforward prediction loop; 13. Real-time feedback loop; 14. Variable message sign control system; 15. Automatic gate group control unit; 16. Traffic signal priority controller; 17. Wireless broadcast push engine; 18. Mobile terminal message service platform; 19. Three-dimensional visualization monitoring interface; 20. Dual-loop closed-loop control mechanism. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] refer to Figures 1-5A multimodal flow collaborative control method for integrated transportation hubs is proposed. This method constructs a four-layer system architecture consisting of a multi-source perception layer 1, a dynamic modeling layer 2, a collaborative decision-making layer 3, and an execution feedback layer 4. This architecture enables comprehensive fusion perception, multi-scale dynamic extrapolation, cross-domain collaborative optimization, and closed-loop precise control of passenger flow, vehicle flow, information flow, and service resource flow within the hub. Based on a spatiotemporal coupled network model and a multi-agent reinforcement learning framework, this method establishes a nonlinear interaction relationship expression between passenger flow, vehicle flow, and information flow. It also introduces a dual-loop control logic combining a feedforward disturbance prediction mechanism and a feedback adjustment and compensation strategy to support real-time collaborative management of dynamically coupled multimodal flows in complex hub environments.

[0038] Furthermore, the multi-source sensing layer 1 consists of a heterogeneous sensor array deployed throughout the entire hub space, including a 3D LiDAR unit, a millimeter-wave radar module, an infrared thermal imaging camera group, a Wi-Fi probe device, a Bluetooth sniffing node, a gate card swiping recording terminal, a video structured analysis server, and a mobile communication base station signaling acquisition device. Each sensing unit is fixedly installed at key access nodes according to a preset spatial topology layout. The 3D LiDAR unit is deployed at 1.5-meter intervals on the side walls of the transfer channel, with the scanning plane perpendicular to the direction of travel, a sampling frequency set to 20Hz, and a point cloud resolution of not less than 0.05m. Its output data includes spatial point coordinates (x, y, z) and reflection intensity values, used to construct a local 3D point cloud map and identify pedestrian contours. The millimeter-wave radar module is embedded in the ceiling structure, covering a circular area with a radius of 8 meters, operating at a frequency of 77GHz, and has a target detection performance with a velocity resolution better than 0.1m / s, capable of simultaneously tracking the position and radial velocity of up to 64 moving targets. Infrared thermal imaging cameras are deployed every 12 meters along the longitudinal axis of the waiting hall, with a field of view of 90°×68° and a temperature sensitivity of ≤0.08℃. They extract heat source areas from the human body by segmenting through temperature difference thresholds and combine this with background modeling algorithms to achieve non-contact density estimation. Wi-Fi probe devices and Bluetooth sniffing nodes are co-located above the security checkpoint, ticket gates, and commercial service area. Their signal reception sensitivity is no less than -95dBm, and their MAC address capture period is 3 seconds. After SHA-256 hash encryption, the data is uploaded to the edge computing node, retaining the device's presence signal but not storing the original identifier. The turnstile card-swiping terminals are connected to the local controller via an RS-485 bus, recording entry / exit timestamps, ticket type, and device number. Timestamp accuracy is synchronized to the millisecond level, and aggregated data packets are uploaded in batches every 500 milliseconds. The video structured analysis server adopts a distributed cluster architecture, accessing H.265 encoded video streams from all public area HD cameras. It deploys a YOLOv7-based pedestrian detection model and an ST-GCN spatiotemporal graph convolutional network for pose recognition, outputting structured metadata packets containing target ID, location coordinates, direction of movement, type of carried items (e.g., suitcase, stroller), and abnormal behavior tags (e.g., falls, wrong-way walking). The mobile communication base station signaling acquisition device mirrors S1-MME signaling data from the operator's core network interface, parses user equipment attachment, handover, and deactivation event sequences, and associates them with the base station location database using a spatiotemporal matching algorithm to generate anonymized mobile phone signaling trajectory sequences. All sensing devices are aggregated to the regional edge gateway via an industrial Ethernet network. VLANs are partitioned using the IEEE 802.1Q standard: video streams are assigned to VLAN 101, radar data to VLAN 102, and signaling data to VLAN 103, ensuring that different types of data streams are physically isolated and that end-to-end communication latency does not exceed 50ms.

[0039] Furthermore, the dynamic modeling layer 2 runs on a high-performance computing cluster in the central data center of the hub, and includes a spatiotemporal coupled network generation module 5, a multimodal flow dynamic evolution equation solver 6, and a disturbance propagation path tracing unit 7. The spatiotemporal coupled network generation module 5 divides the hub's physical space into square grid cells with sides of 1.2 meters, forming a two-dimensional discretized spatial base. Each grid cell is assigned a unique geocode and aligned with the BIM model coordinates. The encoding format is "". ",in Indicates the floor number. The index is an integer. Based on this, the grid type is labeled according to the functional zone attributes: pedestrian access area (code 01) allows free movement; queuing area (code 02) has a maximum dwell time threshold of 180 seconds; static lingering area (code 03) is used for waiting or resting; and prohibited area (code 04) is marked as an obstacle or restricted area. Further, combined with actual reachable path constraints, a directed graph is constructed. Here, the vertex set V corresponds to the coordinates of the center of all walkable grid cells, and the edge set E represents the possibility of movement between adjacent grid cells, including connections in the four directions: up, down, left, and right. Edge weights. Free speed of walking Local density and obstacle repulsive potential energy The decision is made jointly, and the specific expression is as follows: in, The value is taken as 1.3 m / s, and the coefficient is... =0.35, =0.6; Based on the distance to the nearest obstacle Calculate, when When <0.8m, Otherwise, the value is zero. The network updates the topology weights every 100ms. The weight update process is triggered by the edge computing nodes, which initiate the recalculation process by subscribing to the density update messages published by the perception layer. The new weights are pushed to the dynamic evolution equation solver through the gRPC interface.

[0040] Furthermore, the multimodal flow dynamic evolution equation solver (6) uses an improved macroscopic continuum model to describe the population density field. The evolutionary process, in which Inside and For spatial location, Represents the coordinates of a point on a two-dimensional plane. This is a local average velocity vector, which describes the average velocity and direction of the pedestrian flow at a corresponding spatial location and time. Source and sink terms; representing the rate of passenger flow inflow or outflow at station entrances and exits (unit: people / m²). 2 / s). Velocity field Driven by generalized Navier-Stokes-like equations: The hydrodynamic viscosity is Pressure gradient term according to calculate, =0.1, =2.0; Driving force Pointing to the nearest target exit, interaction force The combined force of repulsive forces between individuals and attractive forces within a group is calculated using a social force model. This system of equations is solved using the finite volume method in a GPU-accelerated environment, with boundary conditions dynamically set based on real-time gate access data: the entrance mesh is set as a Dirichlet boundary. The exit mesh is set to Neumann boundary. The solver is deployed on an NVIDIA A100 GPU server cluster, and a single node can process flow field calculations in parallel for up to 8 independent regions, with an average single-step solution time of 87 milliseconds.

[0041] Furthermore, the disturbance propagation path tracking unit 7 constructs a passenger flow disturbance impact diffusion model based on the Markov transition matrix M. In the Markov transition matrix M, the matrix elements... Indicates the time from the region The probability of migrating to region j is represented by an N×N matrix, where N is the total number of passable regions. The initial value is jointly determined by historical card-swiping data and video tracking results, and smoothed using Laplace to prevent zero probability. During online operation, the matrix is ​​incrementally updated every 5 minutes using real-time data from the last 10 minutes, with an update weight of 0.1. When a certain region is detected... Abnormal backlog occurs (defined as the current density exceeding a threshold) =4.0 people / m 2 If the disturbance lasts for more than 3 minutes, the perturbation forward propagation mechanism is activated to calculate the expected arrival distribution within the next T=15 minutes. ,in Minutes. Affected downstream area set. Areas marked as potential risks are assigned an alert level: If A yellow alert is issued for values ​​in the range [0.3, 0.6), an orange alert for values ​​in the range [0.3, 0.6), and a red alert for values ​​≥0.6. Simultaneously, a correlation model for vehicle connection delays is established to link railway train delays. As an input variable, it is mapped to the taxi dispatch demand increment Δ through an empirical regression function. ,in =8 vehicles / minute =15 vehicles. The model parameters were obtained by fitting the data using the least squares method based on the operating data of the past year, with a coefficient of determination of 15. This is used to adjust the departure rhythm of vehicles from the storage pool in advance.

[0042] Furthermore, the collaborative decision-making layer 3 integrates a multi-agent reinforcement learning optimization engine, and its main architecture includes four functional agents: passenger flow management agent. Transportation connection agency Information service agency With emergency response agents Each agent shares a unified state space S and action space A, where the state vector... It consists of the following dimensions: real-time density of each region (N values ​​in total), queue length (M ticket gates), remaining throughput capacity (Calculated based on lane width and speed limit), vehicle carrying capacity (Statistics for taxis, buses, and subways are separate), information dissemination coverage rate (Percentage of users reached), Number of police officers on duty (Based on patrol groups). Motion vectors Includes: Guidance instruction combination G (variable information display content encoding), signal timing scheme (Green light duration vector), vehicle dispatching command (Departure batch and destination), Broadcast content template B (preset voice ID), Police deployment instructions (Reinforcement area number). Each agent independently trains its policy function using the Deep Deterministic Policy Gradient (DDPG) algorithm. The network structure consists of three fully connected layers (512-256-128), with ReLU activation function and tanh normalized output layer. Its reward function R is designed as follows: for for for for in, The target comfort density is 1.8 people / m² 2 , Congestion Index (when >3.5 hours count as 1, otherwise 0). Average passenger waiting time (in seconds). Vehicle empty-run rate (empty-run mileage / total mileage). Information release delay duration (unit: seconds). This refers to the number of times information conflicts occur across multiple channels (such as inconsistencies between information boards and app notifications). The percentage of the population not covered (estimated based on signaling blind spots). The rate of spread of the impact of a sudden event (unit: m / min). The initial intervention response time (in seconds). Capacity loss due to service interruption (unit: passenger trips). Weighting coefficient. to The values ​​are taken as [0.3, 0.25, 0.15, 0.3, 0.2, 0.1, 0.4, 0.3, 0.3, 0.5, 0.3, 0.2], and are then normalized before being used in the joint utility calculation.

[0043] Furthermore, the multi-agent system adopts a centralized training-distributed execution (CTDE) model. During the offline phase, it utilizes historical operational data to generate millions of simulation scenario samples, which are then input into a shared experience replay buffer. During online operation, each agent bases its data on local observations... Independent selection of actions The system coordinates conflicting actions using a Nash equilibrium solver, ultimately outputting a consistent set of control instructions. The coordination process employs an iterative optimal response algorithm. The initial action set consists of independent decisions made by each agent. In each iteration, each agent re-optimizes its own strategy while keeping the actions of other agents fixed, until convergence or the maximum number of iterations (10) is reached. For example, when... It is recommended to close the entrance to a certain transfer passage to alleviate downstream pressure. When it is required that the passage be kept clear to allow shuttle buses to pass, the coordination module calculates the utility loss ratio for both parties. Loss less If the gain is positive, then adopt it. The proposal triggers an alternative route recommendation mechanism, which is then... Generate detour directions and push them to nearby information boards and user apps.

[0044] Furthermore, the execution feedback layer 4 consists of a distributed execution terminal network, including a variable message sign control system 14, an automatic gate group control unit 15, a traffic signal priority controller 16, a wireless broadcast push engine 17, and a mobile terminal message service platform 18. The variable message sign control system 14 is connected to an LED display array deployed at the intersection of main passageways on each floor, with a display size of not less than 1.8m × 1.2m and a brightness ≥ 5000 cd / m². 2The system has a refresh rate of 60Hz and receives guidance path vector map instructions from the collaborative decision-making layer. It dynamically displays the optimal evacuation direction using the ISO 20511 standard graphic symbol system. The instructions include the path start point, end point, recommendation level (level 1-3), and estimated travel time. The rendering engine automatically generates arrow streamlines and color-coded backgrounds based on the current viewpoint. The automatic gate group control unit 15 connects to all ticket checking equipment via the Modbus TCP protocol, supports remote batch setting of access permissions, and can simultaneously lock or open 20 gates in a single area within 3 seconds. After the instruction is executed, a confirmation code and the actual status are returned. An alarm is triggered when the difference exceeds 5%. The traffic signal priority controller 16 connects to the city traffic signal platform and implements green wave adjustments for key intersections around the hub. When a large number of departing passengers need to be quickly evacuated, the green light duration is extended to the bus lane for up to 45 seconds. The adjustment request includes the intersection number, direction, duration, and priority level, and is executed after verification by the traffic management center. The wireless broadcast push engine 17 connects to speaker nodes distributed across various functional areas via an IP network, playing differentiated voice prompts grouped by area number. The volume is automatically adjusted to the background noise level +10dB, with an adjustment cycle of once every 10 seconds, and the output gain is dynamically adjusted based on feedback values ​​from the noise sensor. The mobile terminal message service platform 18 connects to mainstream map application SDKs via an HTTPS interface, pushing personalized travel suggestions to registered users' mobile apps. The content includes estimated walking time, recommended elevator number, the departure time of the next shuttle bus, and seat availability information. The push range is limited to active devices within the hub's geofence and whose user location updates more than once per minute.

[0045] Furthermore, the method establishes a dual-loop closed-loop control mechanism, with the outer loop being a feedforward prediction loop and the inner loop being a real-time feedback loop. The feedforward prediction loop operates every 5 minutes, acquiring external input disturbances based on train timetables, flight dynamic databases, and weather forecast APIs. It then uses a spatiotemporal coupled network model to predict passenger flow load in various regions within the next 30 minutes. The prediction process employs Monte Carlo simulation to generate 1000 possible trajectory samples, taking the expected value as the prediction result. After the pre-control scheme is generated, it is sent to the execution terminal for pre-loading but not yet activated. The real-time feedback loop samples every 500 milliseconds, collecting the latest data stream from each sensing node. After aggregation by edge computing nodes, the data is sent to the dynamic modeling layer to recalculate the current state. If the actual state deviates from the predicted trajectory by more than a tolerance threshold, the loop will initiate a real-time feedback loop. If the deviation reaches 15%, an emergency re-optimization process is immediately triggered, invoking the collaborative decision-making layer to recalculate the optimal strategy and override the original instruction. The deviation calculation method is as follows: ,in This is the density vector for the entire region.

[0046] Furthermore, the system is configured with an adaptive degradation module for abnormal operating conditions. When some sensing devices fail or communication links are interrupted, a backup inference rule base is activated. For example, in the event of video surveillance interruption, passenger flow distribution is analyzed solely based on Wi-Fi probe and card swipe data. A Kalman filter is used to fuse historical trends and observations from neighboring areas to estimate missing data. The state transition matrix is ​​set as a first-order autoregressive model, and the observation matrix is ​​constructed based on spatial adjacency relationships. When an edge computing node crashes, the system switches to a local caching model to continue executing the previous effective control strategy, reduces the instruction update frequency to once every 2 minutes, and simultaneously activates a heartbeat reconnection mechanism, attempting to rebuild the connection every 10 seconds. After 5 consecutive failures, the system reports to the operations and maintenance center.

[0047] Furthermore, the method defines five typical control scenarios and their corresponding strategies during implementation. The first scenario is peak-hour traffic management, where a multi-path diversion mechanism is activated. Variable message signs guide passenger flow to be evenly distributed across multiple parallel transfer corridors, avoiding overload on any single corridor. The diversion ratio is dynamically allocated based on the remaining capacity of each path. The second scenario is widespread train delays, where an emergency response plan is activated. The railway department issues delay notices, and temporary shuttle bus services are increased (one bus every 10 minutes). The priority of taxi dispatch in the underground parking garage is adjusted (departure interval reduced from 5 minutes to 2 minutes), and tiered warning broadcasts are initiated in the waiting area (broadcast every 15 minutes). The third scenario is sudden public events (such as fire alarms), where a rigid isolation strategy is implemented. Nearby turnstiles (within a 50-meter radius) are automatically locked, air conditioning and ventilation systems are shut down to prevent smoke spread, emergency directional signs are illuminated, and personnel are guided to evacuate along predetermined safe routes, avoiding the area within 200 meters upstream and downstream of the alarm point. The fourth category is the impact of severe weather. When meteorological monitoring data shows that a red rainstorm warning is in effect, temporary indoor waiting areas (converted from former commercial areas) will be opened in advance, anti-slip mats and rain gear distribution points will be deployed (one for every 500 people), and the operating time of the last subway train will be extended by 30 minutes. The fifth category is the guarantee for major events. For foreseeable large passenger flows such as after concerts, tidal channels will be set up in advance, the layout of railings will be changed to form a one-way circulation line, and mobile patrol robots will be deployed to strengthen order maintenance. The robots will patrol in a spiral pattern and complete a full scan every 3 minutes.

[0048] Furthermore, the system is equipped with a 3D visualization monitoring interface 19, running on the multi-screen splicing display system of the command center. It uses WebGL technology to render a digital twin model of the hub, achieving a LOD400 level of accuracy, and includes all building components, equipment locations, and pipeline routes. The interface overlays a real-time heat map (color mapping density 0-5 people / m²). 2The interface features vector streamline diagrams (arrow length indicates speed) and resource distribution icons (vehicles, police force, service points), allowing users to hover their mouse over any area to view detailed parameters, including current density, average speed, equipment status, and recent command records. The underlying data of the monitoring interface is directly output from the dynamic modeling layer and pushed via the WebSocket protocol, with refresh latency strictly controlled within 800ms to ensure timely situational awareness. The interface provides a historical playback function, supporting the reconstruction of passenger flow distribution at any past time point. Playback speed can be adjusted between 0.5x and 8x, with a timeline accuracy of 1 second, facilitating post-event analysis of event evolution.

[0049] Furthermore, the method adheres to Level 3 security requirements at the data security level. All sensor data is encrypted using the AES-256 algorithm before transmission. The key is generated by the Hardware Security Module (HSM) and rotated periodically (every 7 days), stored in a privacy-protected database conforming to GB / T 35273-2020 standards. Personal identity information is immediately anonymized, retaining only anonymized trajectory data used for statistical analysis. System access implements an RBAC (Restricted Access Control) model, defining three roles: operator, administrator, and auditor. Different roles are granted corresponding functional module access permissions. Execution of critical instructions requires dual verification and authentication. Operation logs record the complete instruction content, execution time, operator ID, and reviewer ID, and are retained for no less than 180 days.

[0050] Furthermore, the collaborative decision-making layer 2 optimization engine regularly iterates and updates its strategies. Monthly, it automatically extracts operational logs from the previous cycle to generate a training set, including samples from three scenarios: normal operating conditions (60%), disturbance operating conditions (30%), and sudden events (10%). After data cleaning and labeling, these samples are input into the offline training pipeline. The new strategy model undergoes at least 1000 hours of stress testing in a simulation environment. Test scenarios cover extreme situations such as maximum passenger flow exceeding the design value by 20%, frequent multi-agent conflicts, and sudden increases in communication latency, verifying its stability and robustness under extreme loads. The criteria for success are: control success rate ≥ 98%, average response latency ≤ 1.2 seconds, and resource waste rate ≤ 5%. Only after meeting these criteria is deployment allowed to replace the old version. The switchover process uses a blue-green deployment strategy, with the old and new systems running in parallel for 48 hours. Once the output consistency reaches 99.5%, traffic to the old system is cut off.

[0051] Furthermore, the system supports a cross-hub collaborative mode. When the distance between two adjacent integrated transportation hubs is less than 5 kilometers, core status parameters, including total passenger flow, key channel density, connecting transport capacity status, and early warning levels, can be exchanged via a dedicated fiber optic link to establish a joint scheduling and coordination mechanism. The coordination protocol employs a distributed optimization algorithm based on utility maximization, with the objective function being... ,in The weighting factor is 1.5, and the solution is obtained using the Alternating Direction Multiplier Method (ADMM). For example, when a malfunction at hub A causes a surge in outbound passenger flow (density > 4.5 people / m²), the solution is obtained. 2 Hub B currently has a low load (density <1.5 people / m²). 2 When the time is right, the cross-hub diversion plan will be automatically activated, guiding some passengers to Hub B to transfer to rail transit through the regional navigation system. The diversion ratio is determined by the utility gain ratio of both parties, with an upper limit of 15% of the total passenger flow.

[0052] Furthermore, the method adopts a modular design concept in hardware deployment. Each functional unit is packaged as a standardized rack-mount device, conforming to the 19-inch rack specification (600mm×800mm×2000mm). Power input supports AC 220V±10% 50Hz, with total power consumption not exceeding 800W. Cooling is achieved through forced air cooling, equipped with six 80mm PWM fans, achieving a noise level below 55dB(A). All equipment enclosures have a protection rating of at least IP30. Critical servers are configured with dual power redundancy and RAID5 disk arrays, with at least four hard drives, each with a 4TB capacity, ensuring 24 / 7 uninterrupted operation. Inter-device connections utilize LC-LC single-mode fiber optic cables with a transmission rate of 10Gbps and a maximum transmission distance of 10km.

[0053] Furthermore, the software system is developed based on a microservice architecture, using Docker containerization to encapsulate various functional modules, including a perception data access service, a spatiotemporal modeling and computation service, a multi-agent decision-making service, an instruction distribution service, and a log auditing service. Images are stored in a private Harbor repository, and version numbers follow semantic naming rules. Services communicate efficiently via the gRPC protocol, with message serialization using Protocol Buffers. Service registration and discovery are managed by a Consul cluster, and configuration files are centrally stored in an ETCD key-value database, supporting hot updates without restarting processes. Service deployment uses Kubernetes orchestration, setting resource requests (2 CPU cores, 4GB memory) and limits (4 CPU cores, 8GB memory). The automatic scaling threshold is set when the average CPU utilization is >75% for 5 consecutive minutes.

[0054] Furthermore, the method employs the IEEE 1588 Precision Time Protocol (PTP) for time synchronization. A master clock source is deployed in the core network switch, using GPS satellite time synchronization with a accuracy of ±10 ns. All edge nodes have built-in PTP slave clock chips, achieving nanosecond-level synchronization through hardware timestamps. The overall system time error is less than ±10 microseconds. Time synchronization status is reported every minute. When the deviation exceeds ±20 microseconds, a calibration process is triggered, with a calibration cycle of 10 seconds, continuing until the error recovers to within the threshold.

[0055] Furthermore, the system includes an online performance evaluation module that calculates three core KPIs in real time: average evacuation efficiency. , where d_i is the actual walking distance of the i-th passenger (obtained by Wi-Fi trajectory integration). The time it takes to pass through (the time difference between entering and exiting the station). The total number of samples (valid samples within a 10-minute sliding window); resource utilization rate. V_used represents the number of shuttle vehicles already in use. Total available capacity (including spare vehicles); service consistency The number of cases where information push and on-site guidance are consistent (by comparing the APP path and the information board display). This represents the total number of cases. The above metrics are calculated every 10 minutes, and trend curves are plotted for management review. The data is saved to the time-series database InfluxDB and retained for 365 days.

[0056] Furthermore, the method performs a full system calibration process during the initialization phase, comprising three main steps. The first step is spatial calibration, which involves measuring the installation coordinates of all sensors using a total station with a measurement accuracy of ±2mm, and performing ICP registration with the BIM model. The iterative convergence condition is a mean square error <0.5cm. 2 The first step is to ensure that the deviation between the physical coordinate system and the digital model is less than 3 centimeters. The second step is parameter tuning, which involves collecting baseline data from a typical passenger flow scenario (morning peak arrival) for 72 hours, recording 100,000 complete trajectories, and then optimizing the unknown parameters in the social force model using a genetic algorithm. The population size is 100, with 50 iterations. The fitness function is the minimization of the Hausdorff distance between the simulated trajectory and the measured trajectory, and the convergence threshold is a distance decrease rate of <1%. The third step is strategy pre-playing, which involves importing historical data on large passenger flow events (such as a single-day passenger flow of 420,000 during the Spring Festival travel rush) to test whether the system can correctly identify bottlenecks (such as congestion in the North Transfer Channel) and generate reasonable intervention plans (diverting passengers to the South Channel). If the preset effect threshold is achieved in three consecutive tests (density peak reduction ≥20%, evacuation time shortened ≥15%), then the system is deemed ready for online operation.

[0057] Furthermore, the variable message sign control system in the execution feedback layer 4 has a built-in font rendering engine that supports the GB18030 encoding of simplified Chinese characters, with a character height of no less than 200 pixels to ensure clear visibility within a 15-meter distance; the displayed content adopts a red, yellow, and green three-color hierarchical warning system, with green indicating normal passage (ρ≤2.5 people / m). 2 Yellow indicates mild congestion and suggests taking an alternate route (2.5 < ρ ≤ 4.0 people / m²). 2 Red indicates severe congestion and entry is prohibited (ρ>4.0 people / m²). 2The color conversion is triggered by a unified command from the central server, with a delay of no more than 1 second.

[0058] Furthermore, the automatic gate group control unit 15 is configured with a heartbeat detection mechanism, sending a status message to the central server every 2 seconds. The message includes the device number, current mode (normal / emergency / maintenance), passage count (daily cumulative), and fault code (0 for normal, 1 for ticket blockage, 2 for network outage, and 3 for motor failure). If no response is received for 3 consecutive times, it is determined that the communication is interrupted, and the system automatically switches to the local caching strategy mode to continue executing the last valid instruction and report alarm information to the operation and maintenance platform. The alarm information includes the device ID, interruption time, last instruction content, and local log summary.

[0059] Furthermore, a dedicated secure channel is established between the traffic signal priority controller 16 and the urban traffic management center. The control commands are encrypted using the SM4 national cryptographic algorithm, with a key length of 128 bits and an encryption mode of CBC. The initialization vector is randomly generated. Signal adjustment requests must be accompanied by a digital signature and a timestamp. The signature algorithm is SM2. The signal can only be executed after the receiving end verifies the request. After each adjustment, a confirmation receipt is returned. The receipt includes the execution status, the actual adjustment value, and the signature, forming a complete operation audit chain. The records are kept for no less than 180 days.

[0060] Furthermore, the wireless broadcast push engine 17 incorporates a speech synthesis module, employing a neural network model based on the Tacotron 2 architecture. The acoustic model training data includes 100 hours of standard Mandarin recordings. The spectrum prediction network comprises 5 convolutional layers and 2 LSTM layers. The vocoder uses a WaveNet structure with an output sampling rate of 16kHz and an Opus encoding format. Text commands are converted into natural and fluent Mandarin speech in real time. The speaker's voiceprint is fixed as a standard broadcasting tone for a middle-aged male, with a speech rate set at 260 words per minute. Pause intervals conform to the semantic rhythm rules of Chinese, ensuring clear and understandable information delivery.

[0061] Furthermore, the mobile terminal messaging service platform 18 uses the OAuth 2.0 protocol to authenticate user identity, and the authorization scope is limited to location data and device tokens. Personalized push notifications are only provided to users who have completed real-name authentication and authorized location services. The message push frequency is limited to no more than 3 messages per hour. When the threshold is exceeded, the system automatically discards low-priority messages (such as advertisements). Users can unsubscribe from all notifications with one click through the APP interface. The unsubscribe request is immediately synchronized to the message queue to ensure that subsequent messages are not sent.

[0062] In one specific embodiment, a large high-speed rail hub has an average daily passenger flow of 350,000, and the system has entered routine operation after deployment. During the morning peak hours (7:00-9:00), the passenger flow entering the station reaches 12,000 people per hour, and the dynamic modeling layer detects that the density at the east entrance has increased to 3.8 people per square meter.2 The flow exceeded the comfort threshold. The collaborative decision-making level activated the diversion plan, and passenger flow management was implemented. Instruction generated: Direct 30% of passenger flow to the west entrance. This instruction was confirmed and executed after coordination among multiple agents. The variable message sign control system updated its display, a red arrow pointed to the west passage, the public address system played guiding messages, and the mobile app pushed detour suggestions. Five minutes later, the density at the east entrance dropped to 2.9 people / m². 2 The population density at the western entrance increased from 1.6 to 2.4 people / m². 2 The overall distribution tends to be more balanced. Resource utilization rate during the same period... The service consistency η_c increased from 78% to 86%, and reached 97.3%.

[0063] In another embodiment, severe weather caused widespread train delays, with delays lasting up to 45 minutes. The system automatically activated a three-level response mechanism, and the information service agent... Generate delay notices and push them to all channels; transportation connection agent Calculations show that additional taxi capacity is needed. =8 × 45 + 15 = 375 vehicles. The dispatch instruction was issued to the parking lot management system. The automatic gate group control unit opened the backup ticket checking channel, and the wireless broadcast announced waiting times every 10 minutes. The 3D monitoring interface showed that the peak density in the waiting area reached 4.2 people / m². 2 The system automatically marked it as a red alert zone, and the emergency response agent... It was recommended to open a temporary waiting area, and to implement it after approval. Within one hour, the evacuation rate of stranded passengers was 91.5%, and the average waiting time was extended by 22 minutes, which is lower than the historical average of 38 minutes for similar events.

[0064] Furthermore, the abnormal operating condition adaptive degradation module incorporates a health assessment unit to continuously monitor the operating status of each subsystem, including indicators such as CPU load, memory usage, disk I / O, and network throughput. When CPU load is detected to exceed 85% for more than 5 minutes, or memory usage exceeds 90%, a resource cleanup procedure is automatically triggered, suspending non-critical background tasks (such as log archiving and report generation) and releasing at least 20% of computing resources to ensure the priority execution of core control logic. After resource release, monitoring continues for 5 minutes. If the load drops below 70%, normal scheduling resumes; otherwise, a system overload alarm is issued.

[0065] Furthermore, the method establishes a delay level response mechanism when dealing with train delays: Level 1 response is applicable to delays of 5-15 minutes, initiating information broadcast reminders every 10 minutes; Level 2 response is applicable to delays of 16-30 minutes, increasing connecting transport capacity by 15%, and issuing dispatch instructions 10 minutes in advance; Level 3 response is applicable to delays of 31-60 minutes, opening backup waiting areas and providing drinking water supplies, increasing the number of service points from 2 to 6; Level 4 response is applicable to delays exceeding 60 minutes, coordinating public transportation to extend operations, activating large-scale evacuation plans, and launching 5 emergency connecting lines, increasing transport capacity by 300%.

[0066] Furthermore, before formal commissioning, the system must undergo a trial operation phase of no less than 30 days, during which all control commands and system response data are recorded throughout the process, with an average daily data volume exceeding 2TB, stored in the distributed file system Ceph. An independent performance evaluation will be conducted by a third-party organization, assessing the effectiveness of multimodal flow collaborative control (through A / B testing compared to manual scheduling), stability (MTBF ≥ 720 hours of continuous operation), and security (≤ 2 high-risk penetration tests). Only after the evaluation report is deemed satisfactory can the system transition to normalized operation.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal flow collaborative control method for integrated transportation hubs, characterized in that, The method is based on a four-layer system architecture consisting of a multi-source perception layer (1), a dynamic modeling layer (2), a collaborative decision-making layer (3), and an execution feedback layer (4); The method expresses the nonlinear interaction relationship between pedestrian flow, vehicle flow, and information flow by constructing a spatiotemporal coupled network model and a multi-agent reinforcement learning framework, and forms a dual-loop control logic by combining a feedforward disturbance prediction mechanism and a feedback adjustment and compensation strategy. The dynamic modeling layer (2) runs in the central data center and includes a spatiotemporal coupled network generation module (5), a multimodal flow dynamic evolution equation solver (6), and a disturbance propagation path tracing unit (7), which constructs a directed graph. edge weight Free speed of walking Local density and obstacle repulsive potential energy Joint decision; The collaborative decision-making layer (3) integrates passenger flow management agents. Transportation connection agency Information service agency With emergency response agents Each agent shares a unified state space S and action space A, independently trains its policy function using a deep deterministic policy gradient algorithm, and coordinates conflicting action outputs through a Nash equilibrium solver to create a consistent set of control instructions.

2. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The spatiotemporal coupled network generation module (5) constructs a directed graph. At that time, vertex set For all accessible grid center coordinates, edge set Represents the probability of movement between adjacent grid cells, edge weight. The calculation expression is: , where the coefficient =0.35, =0.6; Based on the distance to the nearest obstacle Calculate, when When <0.8m, Otherwise, it is zero; grid cells are assigned a unique geocode and aligned with the BIM model coordinates, with the coding format being " Furthermore, based on the functional zoning attributes, these areas are marked as pedestrian access areas, queuing areas, static lingering areas, or prohibited areas.

3. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The multimodal flow dynamic evolution equation solver (6) uses an improved macroscopic continuous medium model to describe the population density field. The evolutionary process, in which Inside and For spatial location, Represents the coordinates of a point on a two-dimensional plane. This is a local average velocity vector, which describes the average velocity and direction of the pedestrian flow at a corresponding spatial location and time. For source and sink items; velocity field Driven by generalized Navier-Stokes-like equations: The hydrodynamic viscosity is Pressure gradient term according to calculate, =0.1, =2.0; Driving force Pointing to the nearest target exit, interaction force The social force model is used to calculate the resultant force of repulsion between individuals and attraction within a group. The equations are solved using the finite volume method in a GPU-accelerated environment, and the boundary conditions are dynamically set based on real-time gate passage data.

4. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The disturbance propagation path tracking unit (7) constructs a passenger flow disturbance impact diffusion model based on the Markov transition matrix M. In the Markov transition matrix M, the matrix elements... Indicates the time from the region The probability of moving to region j is represented by a matrix with dimensions N×N, where N is the total number of passable regions. The initial value is jointly calibrated by historical card swipe data and video tracking results and then smoothed using Laplace. When running online, it updates incrementally every 5 minutes; when a certain area Density exceeding 4.0 people / m² 2 If the disturbance persists for more than 3 minutes, the perturbation forward propagation mechanism is activated to calculate the expected arrival distribution within the next 15 minutes. ,exist In the state of time Can be derived from Markov transition matrix After multiplying by k times, and... This yields data that reflects the linear evolution of the system state over time and marks the set of affected downstream regions. ;in To measure the key ratio, and simultaneously establish a correlation model for vehicle connection delays, train delays were considered. Mapped to incremental taxi dispatch demand ,in =8 vehicles / minute =15 vehicles.

5. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The state vectors of each agent in the collaborative decision-making layer (3) Includes real-time density of each region Queue length Remaining capacity Vehicle carrying capacity Information dissemination coverage Number of police officers on duty e; Action vectors of each agent Including boot instruction combinations Signal timing scheme Vehicle dispatching orders Broadcast content template B and police force deployment instructions ; The reward functions for each agent are as follows: In the formula, , , and These are integrated passenger flow management agents Transportation connection agency Information service agency With emergency response agents The reward function, Refers to specific densities such as critical density; It is the optimal density; The average queue length, Represents the degree of congestion or the effectiveness of congestion; It is speed-related flow. This refers to the target output and target flow rate. in, exist >3.5 Hours 1, This represents the average waiting time. Empty running rate For information delay duration, For the number of information conflicts, This represents the percentage of areas not covered. For event expansion speed, For response time, For capacity loss; weighting coefficient to After normalization, the values ​​are [0.3, 0.25, 0.15, 0.3, 0.2, 0.1, 0.4, 0.3, 0.3, 0.5, 0.3, 0.2].

6. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The multi-agent system adopts a centralized training-distributed execution model, with each agent operating online based on local observations. Independent selection of actions And Nash equilibrium coordination is achieved through an iterative optimal response algorithm; When passenger flow management agent It is recommended to close the entrance to a certain transfer passage and the transportation connection agent. When maintaining smooth communication is required, the coordination module calculates the utility loss ratio for both parties. The loss is less than The gain is adopted. The proposal triggers an alternative route recommendation mechanism, which is then handled by the information service agent. Generate detour directions and push them to nearby variable message signs and user mobile terminals.

7. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The dual-loop closed-loop control mechanism (20) based on the dual-loop control logic includes an outer loop feedforward prediction loop (12) and an inner loop real-time feedback loop (13). The feedforward prediction loop (12) runs once every 5 minutes. It obtains external disturbances based on the train operation diagram, flight dynamic database and weather forecast API, and predicts the passenger flow load of each area in the next 30 minutes by combining the spatiotemporal coupling network model. It uses Monte Carlo simulation to generate 1,000 trajectory samples and takes the expected value as the prediction result. The pre-control scheme is sent to the execution terminal for pre-loading but not activated for the time being. The real-time feedback loop (13) samples once every 500 milliseconds. If the relative error of the L2 norm of the actual state deviating from the predicted trajectory exceeds 15%, the emergency re-optimization process is immediately triggered, and the collaborative decision layer (3) is called again to calculate the optimal strategy and overwrite the original instruction.

8. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The variable message sign control system (14) in the execution feedback layer (4) is connected to the LED display array deployed at the intersection of the main channels, dynamically displaying the optimal evacuation direction. The display content includes the starting point, ending point, recommended level and estimated passage time of the route. The automatic gate group control unit (15) is connected to all ticket checking equipment through the Modbus TCP protocol, supporting the simultaneous locking or opening of 20 gates in a single area within 3 seconds. After the instruction is executed, the confirmation code and the actual status are returned. If the difference exceeds 5%, an alarm is triggered. The traffic signal priority controller (16) is connected to the urban traffic signal platform, which extends the green light time to the bus lane direction at key intersections by up to 45 seconds.

9. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The system is configured with an adaptive degradation module for abnormal operating conditions. When video monitoring is interrupted, a Kalman filter is used to fuse Wi-Fi probe and card swipe data to estimate passenger flow distribution. The state transition matrix is ​​a first-order autoregressive model, and the observation matrix is ​​constructed based on spatial adjacency relationships. When an edge computing node crashes, the system switches to a local caching model to execute the previous effective control strategy. The health assessment unit continuously monitors CPU load and memory usage. When the CPU load exceeds 85% for 5 minutes or the memory usage is higher than 90%, non-critical background tasks are automatically paused to release at least 20% of computing resources.

10. The multimodal flow collaborative control method for integrated transportation hubs according to claim 1, characterized in that, The method defines five typical control scenarios and their corresponding strategies: during normal peak hours, a multi-path diversion mechanism is activated, and variable message signs are used to guide the even distribution of passenger flow; a Level 1 response is initiated when trains are delayed by 5-15 minutes, with information broadcast every 10 minutes; a Level 2 response is initiated when trains are delayed by 16-30 minutes, and connecting transport capacity is increased by 15%; a Level 3 response is initiated when trains are delayed by 31-60 minutes, with backup waiting areas opened and drinking water supplies provided; a Level 4 response is initiated when trains are delayed by more than 60 minutes, coordinating the extension of public transport operations and activating emergency connecting lines; and a rigid isolation strategy is implemented during public emergencies, automatically locking turnstiles within a 50-meter radius, shutting down the air conditioning and ventilation system, and guiding personnel to evacuate along a safe path 200 meters upstream and downstream of the alarm point.

Images