A web connection bus active scheduling method and system

Abstract

The application discloses a web connection connecting bus active scheduling method and system, and belongs to the technical field of intelligent traffic and vehicle scheduling. The application obtains web connection connecting bus operation data, visual perception data, and multi-modal environment data such as weather and activity text, generates a tidal passenger flow multi-modal feature matrix through cross-modal space-time alignment; uses the cross-modal attention and causal reasoning ability of a pre-trained multi-modal large language model to predict future potential tidal passenger flow gathering points and scales; constructs a Markov decision process containing a dynamic empty driving penalty factor, generates an active scheduling strategy of empty car pre-deployment and dynamic line adjustment through deep reinforcement learning, and executes under the vehicle networking. The application realizes the scheduling mode change from passive response to active prediction, significantly improves the connection efficiency and transport capacity utilization rate in the tidal passenger flow scene, and is suitable for intelligent scheduling of urban microcirculation buses.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and vehicle scheduling technology, specifically relating to a network-connected shuttle bus active scheduling method and system, which is applicable to the intelligent scheduling of urban micro-circulation buses and subway shuttle buses, and can effectively cope with tidal passenger flow scenarios such as morning and evening peak hours, large event dispersal, and extreme weather. Background Technology

[0002] With the acceleration of urbanization and the continuous improvement of public transportation networks, network-connected shuttle buses (micro-circulation buses) serve as a key link in bridging the "last mile" of urban residents' travel, and their operational efficiency directly affects the overall service level of the urban public transportation system.

[0003] Currently, traditional bus dispatching methods largely rely on fixed, static timetables or intervention solely based on dispatchers' manual experience. However, when facing complex and ever-changing urban traffic environments, especially in tidal passenger flow scenarios (such as morning and evening rush hours, the end of large-scale events or performances, and sudden extreme weather), passenger demand often exhibits strong suddenness, directionality, and spatial concentration. Traditional dispatching mechanisms are typically passive responses, meaning that vehicle allocation only begins when the monitoring system or manual reports indicate a large number of passengers stranded at stations. This lag means that replenishing capacity often requires a long response time, significantly increasing passenger waiting times, reducing the travel experience, and easily causing traffic congestion in local road networks.

[0004] To address this issue, some existing public transport dispatching systems have introduced passenger flow prediction technologies based on computer vision or time-series data. For example, they analyze the current number of people waiting at bus stops using cameras, or utilize traditional deep learning models such as Long Short-Term Memory (LSTM) networks to predict historical passenger flow data over time. However, these methods have significant limitations: they can only process structured numerical data or extract features from a single modality, lacking a comprehensive understanding of complex environmental factors and the ability to apply common-sense logical reasoning. In other words, existing technologies cannot effectively deduce the causal relationship between a sudden downpour and the imminent end of a large gathering, leading to an unusually high passenger flow at a specific station within the next half hour, thus failing to provide true preparedness.

[0005] Furthermore, after obtaining scheduling requirements, how to efficiently allocate multi-vehicle capacity is also a major bottleneck for existing technologies. In tidal passenger flow scenarios, connecting buses often face the dilemma of being fully loaded in one direction and empty in the opposite direction. Existing regional scheduling algorithms mostly adopt heuristic rules or traditional operations research optimization models, which are difficult to achieve optimal coordination of multiple vehicles in highly dynamic and high-dimensional complex road network environments.

[0006] Therefore, how to overcome the limitations of existing technologies, utilize the powerful cross-modal understanding and logical reasoning capabilities of multimodal large language models to proactively predict passenger flow, and combine deep reinforcement learning algorithms to achieve coordinated bus scheduling within the regional road network that balances passenger experience and operating costs, thus realizing a shift from passive response to proactive action, is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to address the technical problems of existing bus dispatching systems, such as prediction lag when facing tidal passenger flow and sudden environmental events, difficulty in handling cross-modal unstructured data, and difficulty in balancing response speed and empty running costs in multi-vehicle collaborative dispatching, by providing a network-connected shuttle bus active dispatching method and system.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] This invention provides a method for proactive scheduling of network-connected shuttle buses, comprising the following steps:

[0010] S1. Acquire multi-source heterogeneous environmental data and vehicle status data: acquire basic operation data and location status data of network-connected shuttle buses in real time, and simultaneously acquire visual perception data of the target service area and multimodal environmental early warning data; wherein, the multimodal environmental early warning data includes at least weather forecast text data and schedule text data of surrounding large-scale social activities;

[0011] S2. Multi-source data spatiotemporal alignment and feature fusion: Features are extracted from the visual perception data and multimodal environmental early warning data respectively to obtain passenger flow visual density features and structured environmental text features. The extracted features and the location status data of the network-connected shuttle buses are mapped to a unified spatiotemporal coordinate system to generate a tidal passenger flow multimodal feature matrix.

[0012] S3. Passenger flow prediction based on multimodal large language model: The tidal passenger flow multimodal feature matrix is ​​input into the pre-trained multimodal large language model. Through the cross-modal attention mechanism and causal logic reasoning of the multimodal large language model, the potential tidal passenger flow gathering points and predicted passenger flow scale of the target service area within a future preset time window are output.

[0013] S4. Generation of Cooperative Scheduling Strategy Based on Deep Reinforcement Learning: Based on the potential tidal passenger flow gathering points and predicted passenger flow scale, a Markov decision process is constructed with the joint objective of minimizing the total waiting time of regional passengers and maximizing the utilization rate of network-connected shuttle buses. The multi-vehicle state and predicted passenger flow state are used as the state space and input into a pre-trained deep reinforcement learning network model for collaborative solution to generate an active cooperative scheduling strategy. The active cooperative scheduling strategy includes an empty vehicle pre-deployment instruction and a dynamic shuttle route generation instruction.

[0014] S5. Vehicle-to-Cloud Collaborative Command Issuance and Execution: Based on the aforementioned proactive collaborative scheduling strategy, standardized vehicle control commands are generated and issued to the corresponding target network-connected shuttle buses via the vehicle-to-everything (V2X) communication protocol to perform proactive scheduling before passenger flow gathers.

[0015] Furthermore, step S2 specifically includes:

[0016] The target service area is divided into multiple discrete spatiotemporal grids;

[0017] Target detection and density clustering are performed on visual perception data to extract real-time waiting number and congestion characteristics within each spatiotemporal grid.

[0018] Key entities in multimodal environmental early warning data are extracted using natural language processing techniques, including activity type, activity time, and sudden weather events, and then transformed into quantified environmental impact feature vectors.

[0019] The location status data, real-time waiting number and congestion characteristics, and environmental impact feature vectors are concatenated according to timestamps and spatiotemporal grid indices to generate the tidal passenger flow multimodal feature matrix.

[0020] Furthermore, step S3 specifically includes:

[0021] The tidal passenger flow multimodal feature matrix is ​​converted into a structured prompt word sequence;

[0022] The structured cue word sequence is input into a pre-trained multimodal large language model, and the implicit causal association weights between environmental mutation events and regional historical passenger flow patterns are calculated through a cross-modal attention mechanism.

[0023] Based on the implicit causal association weights, the probability of a surge in tidal passenger flow and the corresponding predicted passenger flow scale are output for each time slice and each spatiotemporal grid in the future.

[0024] Furthermore, the structured prompt word sequence described in step S3 includes at least the following fields: current time, target area grid number, current number of people waiting for the train, environmental warning information, historical average passenger flow data for the same period, and prediction time window length.

[0025] Furthermore, the construction of the Markov decision process in step S4 includes:

[0026] State space definition: The state space should at least include the predicted passenger flow scale of each spatiotemporal grid, the real-time latitude and longitude coordinates of each grid's connecting buses, the remaining passenger capacity, and the currently assigned routes;

[0027] Action space definition: Action space includes route change instructions for vehicles within the area, departure interval adjustment instructions, and deployment instructions to pre-schedule empty vehicles to potential tidal passenger flow gathering points;

[0028] Reward function definition: The reward function consists of a positive passenger connection efficiency reward and a negative passenger delay penalty and vehicle empty driving distance penalty. Among them, the weight of the empty driving distance penalty is adjusted by introducing a dynamic empty driving penalty factor to achieve a balance between avoiding resource waste caused by excessive front-loading of transport capacity and meeting the extreme demand of tidal passenger flow.

[0029] Furthermore, the dynamic empty-run penalty factor is adaptively adjusted according to the predicted passenger flow. When the predicted passenger flow exceeds a preset threshold, the dynamic empty-run penalty factor decays exponentially, reducing the weight of the empty-run distance penalty.

[0030] Furthermore, the deep reinforcement learning network model described in step S4 employs either the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm or the Proximal Policy Optimization (PPO) algorithm.

[0031] Furthermore, the vehicle-to-everything (V2X) communication protocol mentioned in step S5 is the 5G-V2X protocol, and the generated standardized vehicle control commands include at least the target road segment, suggested vehicle speed, expected arrival time, and passenger capacity threshold parameters.

[0032] Furthermore, step S5 also includes:

[0033] Real-time tracking of actual passenger flow data; when the deviation between actual passenger flow and predicted passenger flow exceeds a preset threshold, the trajectory data of the Markov decision process for that scheduling is extracted and stored in the experience replay pool, triggering online fine-tuning of the deep reinforcement learning network model.

[0034] A system for implementing the above-mentioned active scheduling method for network-connected shuttle buses includes:

[0035] The multi-source sensing module is used to acquire basic operational data and location status data of the network-connected shuttle buses in real time, and simultaneously acquire visual perception data of the target service area as well as multimodal environmental early warning data such as weather and event schedules;

[0036] The multimodal alignment module is used to extract features from the above multi-source data respectively, and map the extracted features to a unified spatiotemporal coordinate system based on a cross-modal gated fusion network to generate a tidal passenger flow multimodal feature matrix.

[0037] The large model prediction module is used to convert the tidal passenger flow multimodal feature matrix into a structured sequence, input it into a pre-trained multimodal large language model, and output the potential tidal passenger flow gathering points and predicted passenger flow scale of the target service area within a future preset time window through causal logic reasoning.

[0038] The regional collaborative scheduling module is used to receive the predicted passenger flow, construct a Markov decision process with the predicted data in the state space, and use a deep reinforcement learning network model to maximize the reward function containing the dynamic empty run penalty factor to generate an active collaborative scheduling strategy.

[0039] The vehicle-cloud collaborative execution module is used to generate control commands based on proactive collaborative scheduling strategies and send them to the corresponding target network connected shuttle buses for execution via the 5G-V2X vehicle-to-everything (V2X) communication protocol.

[0040] This invention proposes a network-connected shuttle bus active dispatching method and system, which has the following advantages and beneficial effects:

[0041] 1. Breaks through the limitations of single-modal prediction and possesses strong causal logical reasoning ability.

[0042] This invention abandons traditional prediction models that rely solely on time-series digital features or single image features, and innovatively introduces a multimodal large language model. By extracting and quantifying unstructured multimodal data such as meteorological texts and texts from the dispersal of large-scale events in the surrounding area, and utilizing the cross-modal attention mechanism of the large model for common-sense causal reasoning, it can accurately predict nonlinear passenger flow surges caused by the superimposed effects of complex environments, significantly improving the prediction accuracy in extreme tidal passenger flow scenarios.

[0043] 2. Achieve a shift from a passive response model to a proactive, pre-positioned transportation capacity model.

[0044] Existing bus dispatching systems are mostly "passive response" systems that only initiate dispatching after large-scale crowd congestion is detected. This invention uses the advanced spatiotemporal prediction results of a large model as the state input to the Markov decision process, enabling the system to accurately locate future hotspot grids 15-30 minutes before the actual surge in passenger flow, and issue instructions for deploying empty vehicles and dynamically adjusting routes in advance, fundamentally eliminating the phenomenon of long-term passenger congestion and waiting in sudden surges in passenger flow scenarios.

[0045] 3. A perfect balance between active shuttle efficiency and system empty-running costs based on deep reinforcement learning.

[0046] To address the operational research challenges that "proactive pre-emption" of large passenger flows can easily lead to empty vehicle operation and resource waste in certain areas, this invention innovatively designs a reinforcement learning multi-objective joint reward function that includes a dynamic empty-run penalty adjustment factor. This mechanism can adaptively adjust the penalty weight based on the passenger flow intensity predicted by the large model, guiding the multi-agent network to learn to appropriately release empty vehicle scheduling authority when dealing with extreme tidal passenger flows, and tighten the empty-run threshold during off-peak periods, thereby achieving globally optimal collaborative allocation of transport capacity resources across complex road networks and spatiotemporal dimensions. Attached Figure Description

[0047] Figure 1 This is an overall flowchart of the network-connected shuttle bus active dispatching method described in this invention;

[0048] Figure 2 This is an architecture block diagram of the network-connected shuttle bus active dispatching system described in this invention. Detailed Implementation

[0049] The present invention will be further described below with reference to the embodiments. It should be noted that these are merely examples and descriptions of the inventive concept. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in the claims, they should all be considered to fall within the protection scope of the present invention.

[0050] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Those skilled in the art will readily understand other advantages and effects of the present invention from the content disclosed in this specification.

[0051] Example 1

[0052] refer to Figure 1 This invention discloses a specific method for active scheduling of network-connected shuttle buses, comprising the following steps:

[0053] S1, acquire multi-source heterogeneous environmental data and vehicle status data;

[0054] In this embodiment, the system retrieves multi-source data from the target service area in real time using a preset baseline time granularity. Specifically, this includes:

[0055] (1) Visual perception data: real-time video stream images are obtained by retrieving surveillance cameras from the road network and bus stops;

[0056] (2) Multimodal environmental early warning data: Real-time capture of unstructured weather forecast texts issued by the meteorological bureau through API interface, as well as text data of schedules of large-scale social events (such as concerts, sports events, large-scale exhibitions, etc.) captured from local ticketing platforms or social media;

[0057] (3) Vehicle status data: The real-time GPS location, speed, current passenger load and remaining capacity of all network-connected shuttle buses in the area are obtained through the 5G-V2X vehicle network protocol.

[0058] S2, Spatiotemporal alignment of multi-source data and structuring of tidal passenger flow characteristics;

[0059] Because the acquired data differs in temporal frequency and spatial scale, this invention first divides the target service area into... A discrete spatiotemporal grid, each grid denoted as . The current time slice is denoted as .

[0060] For any spacetime grid The system constructs a multimodal feature matrix. The specific process is as follows:

[0061] S21, using a convolutional neural network to perform object detection and density clustering on visual perception data, extracting the density of waiting crowds, and representing it as a visual feature vector after normalization. A pre-trained text encoder is used to extract semantic entities from environmental warning texts, converting them into environmental text feature vectors. Simultaneously, the vehicle state vectors within the grid are aggregated. .

[0062] S22, a cross-modal gated fusion network is used to unify the dimensions of the above features and concatenate them to generate a structured tidal passenger flow multimodal feature matrix:

[0063]

[0064] in, Here are the learnable weight matrices for each modality. This indicates element-wise addition. A vector representing the spatio-temporal positional encoding to preserve the spatio-temporal topology of the data.

[0065] S3, Passenger Flow Causal Reasoning and Proactive Prediction Based on Multimodal Large Language Model;

[0066] S31. Traditional time series models cannot handle complex sudden events. This invention utilizes the powerful cross-modal commonsense reasoning ability of multimodal large language models (MLLM) to assess the nonlinear surge in traffic demand caused by environmental mutations.

[0067] S32, the system will use the multimodal feature matrix Based on the system's preset rules, it is transformed into a structured prompt sequence.

[0068] S33, within the large language model, uses a multi-head cross-modal attention mechanism to treat text features as query vectors. Visual and vehicle state features are used as key vectors Sum value vector Perform implicit causal relationship calculation:

[0069]

[0070] in, The dimension mask factor of the key vector. This is used to mask information from future time steps. The model ultimately outputs structured JSON predictions, including the probability of a surge in passenger flow in each grid cell within a specified future time window, and the predicted expected passenger flow size. This step successfully transformed incalculable social and natural mutations into quantifiable potential passenger flow demand.

[0071] S4, Generation of Regional Public Transport Cooperative Scheduling Strategies Based on Deep Reinforcement Learning;

[0072] S41, after obtaining the predicted passenger flow scale, this invention models the active collaborative scheduling problem of regional connected buses as a Markov decision process, and uses a deep reinforcement learning algorithm to solve the optimal multi-vehicle scheduling strategy in the tidal passenger flow scenario.

[0073] S42, Definition of state space: at time step global state It not only includes the current physical state of the system, but also directly integrates the future state predicted by the large model, and is defined as:

[0074]

[0075] in, Indicates vehicle Real-time coordinates This indicates the remaining passenger capacity.

[0076] S43, Definition of Action Space: Scheduling Actions This includes the set of instructions issued to all vehicles within the area, represented as... Among them, a single action . This indicates that the original route within the section will be maintained. This indicates that they will remain in place and await further instructions. This indicates the core proactive intervention instruction: assigning vehicles, even when empty, to target grids with potential surges in passenger flow. Perform pre-deployment.

[0077] S44, Definition of the reward function: In order to achieve the optimal balance between meeting the sudden demand of tidal passenger flow and controlling the cost of vehicle idling, the system designs a multi-objective joint reward function that includes a dynamic penalty mechanism. :

[0078]

[0079] In the formula, The number of successfully connected passengers (positive reward). Penalty for total waiting time of stranded passengers. Penalty for unattended driving distance incurred when a vehicle performs proactive deployment. For dynamic no-run penalty adjustment factor: When a large model predicts that a region is about to experience an extremely high volume of tidal passenger flow, This will decrease exponentially, thereby significantly reducing the "empty-run cost weight" of vehicles traveling to the area in the reinforcement learning network, encouraging agents to sacrifice short-term empty-run costs in exchange for maximizing the overall macro-connectivity efficiency of the system.

[0080] S45, Strategy Training and Solution: A large number of historical tidal passenger flow events are injected into an offline microscopic traffic simulation environment for reinforcement learning network training. In online applications, the model outputs the current optimal proactive cooperative scheduling strategy that maximizes the expected cumulative discount reward with a sub-second response.

[0081] S5, vehicle-cloud collaborative command issuance and execution;

[0082] S51: After generating an active strategy, the regional collaborative scheduling module converts it into standardized vehicle control messages, including the target road segment, suggested speed, and expected arrival time. These messages are then sent to the onboard OBU (On-Board Unit) connected to the target network via the 5G vehicle-to-everything (V2X) protocol. Upon receiving the instructions, the vehicles pre-position their transport capacity before the actual passenger flow gathers.

[0083] S52, the system will track the actual passenger flow in real time. If the actual passenger flow deviates from the predicted scale of the large model by more than a preset threshold, the system will extract the MDP trajectory data of this run and store it in the experience replay pool to trigger the online dynamic fine-tuning of the reinforcement learning model, so as to realize the closed-loop adaptive evolution of the algorithm.

[0084] S53, based on the same inventive concept as the above method, this embodiment of the invention also provides a network-connected shuttle bus active dispatching system based on a multimodal large language model. The system includes:

[0085] (1) Multi-source sensing module, used to acquire basic operation data and location status data of network-connected shuttle buses in real time, and simultaneously acquire visual perception data and multi-modal environmental early warning data such as weather and event schedules;

[0086] (2) Multimodal alignment module, used to extract features from the perceived data respectively, and generate a unified dimension tidal passenger flow multimodal feature matrix using a cross-modal gating fusion network;

[0087] (3) Large model prediction module, used to convert the multimodal feature matrix into a structured sequence input into a pre-trained multimodal large language model, and output the potential tidal passenger flow gathering points and predicted passenger flow scale of the target service area within a future preset time window through a cross-modal attention mechanism;

[0088] (4) Regional collaborative scheduling module, which is used to construct a Markov decision process with predictive data in the state space, and to maximize the reward function containing dynamic empty driving penalty factor by using a deep reinforcement learning network model to generate an active collaborative scheduling strategy.

[0089] (5) Vehicle-cloud collaborative execution module, used to generate control commands based on the active collaborative scheduling strategy and send them to the corresponding target network connected shuttle buses for execution through the vehicle network communication protocol.

[0090] Embodiments of the present invention also provide an electronic device, which is typically configured as a cloud scheduling server or an edge computing node. Specifically, the electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores computer instructions executable by the at least one processor, which, when executed, enable the at least one processor to perform the method steps S1 to S5 described in detail in the foregoing embodiments. Physically, the electronic device may further include a communication interface and a bus for enabling data interaction with external devices.

[0091] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the present invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A web-connected bus connection active scheduling method, characterized in that, Includes the following steps: S1. Acquire multi-source heterogeneous environmental data and vehicle status data: acquire basic operation data and location status data of network-connected shuttle buses in real time, and simultaneously acquire visual perception data of the target service area and multimodal environmental early warning data; wherein, the multimodal environmental early warning data includes at least weather forecast text data and schedule text data of surrounding large-scale social activities; S2. Multi-source data spatiotemporal alignment and feature fusion: Features are extracted from the visual perception data and multimodal environmental early warning data respectively to obtain passenger flow visual density features and structured environmental text features. The extracted features and the location status data of the network-connected shuttle buses are mapped to a unified spatiotemporal coordinate system to generate a tidal passenger flow multimodal feature matrix. S3. Passenger flow prediction based on multimodal large language model: The tidal passenger flow multimodal feature matrix is ​​input into the pre-trained multimodal large language model. Through the cross-modal attention mechanism and causal logic reasoning of the multimodal large language model, the potential tidal passenger flow gathering points and predicted passenger flow scale of the target service area within a future preset time window are output. S4. Generation of Cooperative Scheduling Strategy Based on Deep Reinforcement Learning: Based on the potential tidal passenger flow gathering points and predicted passenger flow scale, a Markov decision process is constructed with the joint objective of minimizing the total waiting time of regional passengers and maximizing the utilization rate of network-connected shuttle buses. The multi-vehicle state and predicted passenger flow state are used as the state space and input into a pre-trained deep reinforcement learning network model for collaborative solution to generate an active cooperative scheduling strategy. The active cooperative scheduling strategy includes an empty vehicle pre-deployment instruction and a dynamic shuttle route generation instruction. S5. Vehicle-to-Cloud Collaborative Command Issuance and Execution: Based on the aforementioned proactive collaborative scheduling strategy, standardized vehicle control commands are generated and issued to the corresponding target network-connected shuttle buses via the vehicle-to-everything (V2X) communication protocol to perform proactive scheduling before passenger flow gathers.

2. The web-connected shuttle bus active scheduling method of claim 1, wherein, Step S2 specifically includes: The target service area is divided into multiple discrete spatiotemporal grids; Target detection and density clustering are performed on visual perception data to extract real-time waiting number and congestion characteristics within each spatiotemporal grid. Key entities in multimodal environmental early warning data are extracted using natural language processing techniques, including activity type, activity time, and sudden weather events, and then transformed into quantified environmental impact feature vectors. The location status data, real-time waiting number and congestion characteristics, and environmental impact feature vectors are concatenated according to timestamps and spatiotemporal grid indices to generate the tidal passenger flow multimodal feature matrix.

3. The web-connected shuttle bus active scheduling method of claim 1, wherein, Step S3 specifically includes: The tidal passenger flow multimodal feature matrix is ​​converted into a structured prompt word sequence; The structured cue word sequence is input into a pre-trained multimodal large language model, and the implicit causal association weights between environmental mutation events and regional historical passenger flow patterns are calculated through a cross-modal attention mechanism. Based on the implicit causal association weights, the probability of a surge in tidal passenger flow and the corresponding predicted passenger flow scale are output for each time slice and each spatiotemporal grid in the future.

4. The web-connected shuttle bus active scheduling method of claim 1, wherein, The structured prompt word sequence described in step S3 includes at least the following fields: current time, target area grid number, current number of people waiting for the bus, environmental warning information, historical average passenger flow data for the same period, and prediction time window length.

5. The web-connected shuttle bus active scheduling method of claim 1, wherein, The construction of the Markov decision process in step S4 includes: State space definition: The state space should at least include the predicted passenger flow scale of each spatiotemporal grid, the real-time latitude and longitude coordinates of each grid's connecting buses, the remaining passenger capacity, and the currently assigned routes; Action space definition: Action space includes route change instructions for vehicles within the area, departure interval adjustment instructions, and deployment instructions to pre-schedule empty vehicles to potential tidal passenger flow gathering points; Reward function definition: The reward function consists of a positive passenger connection efficiency reward and a negative passenger delay penalty and vehicle empty driving distance penalty. Among them, the weight of the empty driving distance penalty is adjusted by introducing a dynamic empty driving penalty factor to achieve a balance between avoiding resource waste caused by excessive front-loading of transport capacity and meeting the extreme demand of tidal passenger flow.

6. The web-connected shuttle bus active scheduling method of claim 5, wherein, The dynamic empty-run penalty factor is adaptively adjusted according to the predicted passenger flow. When the predicted passenger flow exceeds the preset threshold, the dynamic empty-run penalty factor decays exponentially, reducing the weight of the empty-run distance penalty.

7. The web-connected shuttle bus active scheduling method of claim 1, wherein, The deep reinforcement learning network model described in step S4 employs a multi-agent deep deterministic policy gradient algorithm or a proximal policy optimization algorithm. 8.The web-connected bus connection active scheduling method according to claim 1, wherein, The vehicle-to-everything (V2X) communication protocol mentioned in step S5 is the 5G-V2X protocol, and the generated standardized vehicle control commands include at least the target road segment, suggested vehicle speed, expected arrival time, and passenger capacity threshold parameters.

9. The web-connected shuttle bus active scheduling method of claim 1, wherein, Step S5 also includes: Real-time tracking of actual passenger flow data; when the deviation between actual passenger flow and predicted passenger flow exceeds a preset threshold, the trajectory data of the Markov decision process for that scheduling is extracted and stored in the experience replay pool, triggering online fine-tuning of the deep reinforcement learning network model.

10. A system for implementing the network-connected bus platoon active scheduling method of any one of claims 1-9, characterized in that, include: The multi-source sensing module is used to acquire basic operational data and location status data of the network-connected shuttle buses in real time, and simultaneously acquire visual perception data of the target service area as well as multimodal environmental early warning data such as weather and event schedules; The multimodal alignment module is used to extract features from the above multi-source data respectively, and map the extracted features to a unified spatiotemporal coordinate system based on a cross-modal gated fusion network to generate a tidal passenger flow multimodal feature matrix. The large model prediction module is used to convert the tidal passenger flow multimodal feature matrix into a structured sequence, input it into a pre-trained multimodal large language model, and output the potential tidal passenger flow gathering points and predicted passenger flow scale of the target service area within a future preset time window through causal logic reasoning. The regional collaborative scheduling module is used to receive the predicted passenger flow, construct a Markov decision process with the predicted data in the state space, and use a deep reinforcement learning network model to maximize the reward function containing the dynamic empty run penalty factor to generate an active collaborative scheduling strategy. The vehicle-cloud collaborative execution module is used to generate control commands based on proactive collaborative scheduling strategies and send them to the corresponding target network connected shuttle buses for execution via the 5G-V2X vehicle-to-everything (V2X) communication protocol.

Images