Dynamic vehicle scheduling intelligent decision-making system based on big data

Abstract

The invention discloses a dynamic vehicle scheduling intelligent decision-making system based on big data, and relates to the technical field of intelligent traffic and logistics scheduling, and the system comprises a data collection module, an event analysis module, a knowledge graph construction module, a hypergraph modeling module, a constraint processing engine, an optimization decision-making module, a strategy verification module and an output interaction module. The system has the advantages that multi-source data such as government announcement texts and social media information are acquired in real time through the data acquisition module, event key information is extracted through the event analysis module by utilizing a natural language processing technology, and an event knowledge graph of an incidence relation is constructed through the knowledge graph construction module; the strategy verification module simulates and verifies the strategy effect through the digital twinning technology and iteratively optimizes the strategy effect, the whole process does not need manual intervention to adjust rules, the limitation that a traditional system depends on manual processing is broken through, and the problems that dynamic strategy adjustment is time-consuming, labor-consuming and error-prone in an extreme scene are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation and logistics scheduling, and specifically to a dynamic vehicle scheduling intelligent decision-making system based on big data. Background Art

[0002] With the acceleration of global urbanization and the vigorous development of intelligent transportation technology, vehicle dispatching systems, as the core link of urban transportation and logistics systems, are facing unprecedented challenges and opportunities.

[0003] Existing technologies have certain defects. The first is the lack of ability to respond to extreme scenarios. There is a lack of dynamic adjustment strategies for public emergencies such as epidemic lockdowns and traffic control for large-scale events. Traditional systems can only manually modify scheduling rules temporarily, which is time-consuming, labor-intensive and error-prone. Secondly, there is insufficient modeling of complex scenarios. For multi-modal traffic coordination scenarios such as "bus + shared bicycle" docking scheduling, and multi-constraint scenarios such as route restrictions for hazardous materials transport vehicles and charging time windows for new energy vehicles, existing technologies can only be processed by manually adding "hard rules" and cannot be automatically compatible through algorithms. To this end, we propose a dynamic vehicle scheduling intelligent decision-making system based on big data. Summary of the Invention

[0004] The purpose of the present invention is to provide a dynamic vehicle scheduling intelligent decision-making system based on big data.

[0005] To achieve the above objectives, the present invention provides the following technical solutions: a big data-based dynamic vehicle dispatch intelligent decision-making system, the intelligent decision-making system comprising:

[0006] Data Collection Module: This module collects heterogeneous data from multiple sources in real time, including IoT sensor data, government announcements, social media information, historical order data, and vehicle status data, and transmits the data to the Event Parsing Module and Hypergraph Modeling Module.

[0007] Event parsing module: This module performs natural language processing on unstructured data, extracts key information including event type, impact scope, and time window, generates a structured event description including basic event attributes, spatial impact area, and time validity interval, and transmits it to the knowledge graph construction module.

[0008] Knowledge graph construction module: This module uses events as core nodes and associates entities such as affected road nodes, restricted vehicle type nodes, and alternative route nodes to construct an event knowledge graph containing three types of relationships: "impact," "restriction," and "recommendation." The "impact" relationship represents the effect of an event on road traffic, the "restriction" relationship represents the traffic constraints imposed by an event on vehicle types, and the "recommendation" relationship represents the feasible alternative routes corresponding to the event. The event knowledge graph is then output to the constraint processing engine.

[0009] Hypergraph modeling module: This module defines vehicles, infrastructure, and constraints as nodes, and defines complex multi-entity relationships as hyperedges. This includes complex multi-entity relationships such as the charging matching relationship between vehicles and charging piles, the parking constraint relationship between parking lots and vehicles, and the relationship between prohibited driving rules and road restrictions. It constructs a spatiotemporal hypergraph model consisting of a set of vehicle nodes, a set of infrastructure nodes, and a set of constraint nodes, and transmits it to the constraint processing engine.

[0010] Constraint processing engine: This engine receives event knowledge graphs and spatiotemporal hypergraph models, extracts temporary road bans, time window restrictions, and vehicle type restrictions caused by events, as well as dynamic constraints such as charging time windows and parking lot capacity restrictions in the hypergraph model, generates a constraint set containing spatial constraints, time constraints, and vehicle attribute constraints, and loads it into the optimization decision module.

[0011] Optimization decision module: Based on the dynamic constraint set, it uses a multi-agent reinforcement learning algorithm to generate an initial scheduling strategy for each vehicle. The initial scheduling strategy includes driving routes, task allocation, stop nodes, and time arrangements, and transmits the initial scheduling strategy to the strategy verification module.

[0012] Strategy Verification Module: This module uses digital twin technology to build a virtual model that is isomorphic to the real traffic system. It simulates the execution of the initial scheduling strategy, collects data such as vehicle travel time, idle distance, task completion rate, and user waiting time, and generates strategy evaluation results that include efficiency indicators, cost indicators, and user experience indicators. The results are then fed back to the optimization decision module for strategy iteration.

[0013] Output interaction module: used to receive the final scheduling strategy generated by the optimization decision module, generate visual instructions including real-time vehicle route map, constraint condition labeling and abnormal event warning, and send them to the new energy vehicle battery management system, automatic driving controller and dispatcher monitoring interface using standardized interfaces.

[0014] As a further solution of the present invention: in the data acquisition module, the collected IoT sensor data includes vehicle-mounted GPS trajectory data, license plate and speed data recognized by traffic cameras, charging status and remaining power data of charging piles, and the number and location data of parking spaces; the government announcement text includes traffic control notices, weather warning information and public event announcements.

[0015] As a further solution of the present invention: the event parsing module adopts named entity recognition technology and relationship extraction technology to extract "event-impact range-time" triples consisting of event name, affected area geographical coordinate set and valid time interval from unstructured text, wherein the named entity recognition technology is based on a combination model of bidirectional long short-term memory network and conditional random field, and the relationship extraction technology is based on a neural network model of attention mechanism.

[0016] As a further solution of the present invention: in the event knowledge graph constructed by the knowledge graph construction module, the node types include event nodes, road nodes, vehicle type nodes and route nodes, wherein the road nodes include attributes such as road ID, number of lanes, speed limit requirements, etc., the vehicle type nodes include attributes such as vehicle model, load, energy type, etc., the route nodes include attributes such as starting point, end point, and sequence of passing nodes, and the edge types include "influence", "restriction" and "recommendation", wherein the "influence" edge connects the event node with the road node, the "restriction" edge connects the event node with the vehicle type node, and the "recommendation" edge connects the event node with the route node.

[0017] As a further solution of the present invention: in the spatiotemporal hypergraph model defined by the hypergraph modeling module, the node set V is strictly divided as follows:

[0018] V={V vehicle ,V Ifst ,V Ctt};

[0019] Among them, V vehicle Represents a set of vehicle nodes, each of which contains attributes such as vehicle ID, current location, remaining power, load capacity, and vehicle type. Ifst Represents a set of infrastructure nodes, including charging pile nodes, parking lot nodes, and bus station nodes, etc. Ctt Represents a set of constraint nodes, including prohibited road section nodes, time window nodes, and vehicle load limit nodes. Each hyperedge e in the hyperedge set E i Connect at least V vehicle With V Ifst and V Ctt Any node in is used to represent the interaction constraints between vehicles and infrastructure and the adaptation relationship between vehicles and constraints.

[0020] As a further solution of the present invention, the hypergraph modeling module includes a hypergraph embedding algorithm to convert the hypergraph structure into a low-dimensional vector space. The embedding function is expressed as follows:

[0021]

[0022] Among them, h vis the d-dimensional embedding vector of node v, including node attributes and association relationship features, h u is the embedding vector of other nodes u associated with node v in the hyperedge e, E v is the set of all hyperedges containing node v, |e| is the number of nodes contained in hyperedge e, W is a d×d-dimensional trainable weight matrix used for feature transformation, and b is a d-dimensional bias vector. The algorithm is based on weighted aggregation of feature information of all nodes within the hyperedge to achieve vectorized expression of multiple types of constraints, enabling deep neural networks to handle the complex collaborative relationship between vehicles, infrastructure, and constraints.

[0023] As a further solution of the present invention, the constraint processing engine adopts a two-layer optimization architecture including a bottom real-time constraint processing layer and a high-level global optimization layer:

[0024] The underlying real-time constraint processing layer uses an improved Dijkstra algorithm to handle constraints such as prohibited road sections, vehicle battery critical value, order cancellation events, and vehicle fault status, generating a feasible solution space that includes feasible driving routes, effective task allocation, and reasonable docking nodes.

[0025] The high-level global optimization layer is based on a deep neural network combined with a Bayesian optimization algorithm. It searches for a multi-objective optimal solution in the feasible solution space with the goals of shortest scheduling time, lowest operating cost, minimum carbon emissions, and extreme vehicle load balance. The deep neural network input is vehicle status, constraints, and historical scheduling data, and the output is the optimization parameters of each objective weight.

[0026] As a further solution of the present invention: the optimization decision module adopts a strategy generation algorithm combining Monte Carlo tree search and reinforcement learning:

[0027] Based on the event knowledge graph, real-time constraints such as traffic control constraints, weather impact constraints, and passenger flow constraints for large-scale events are dynamically obtained. Using the multi-agent reinforcement learning algorithm, a local optimal driving strategy is generated for each vehicle that takes into account the status of surrounding vehicles, real-time road conditions, and task priorities. Based on the Monte Carlo tree search algorithm, the local strategies of all vehicles are combined and verified, and the global feasibility and goal achievement of the strategy combination are evaluated, realizing the fully automatic generation and iterative optimization of scheduling strategies without human intervention. The reward function of reinforcement learning includes multi-dimensional parameters such as time efficiency, cost control, and user satisfaction.

[0028] As a further solution of the present invention: the digital twin model constructed by the strategy verification module includes:

[0029] Road network virtual mapping sub-model: used to synchronize road parameters such as traffic flow speed, signal light timing scheme, road congestion level and temporary construction areas in real time;

[0030] Vehicle dynamic behavior simulation sub-model: used to simulate vehicle acceleration, deceleration, lane change, parking, charging and other driving behaviors and energy consumption processes;

[0031] User demand prediction sub-model: This model is used to predict order volume, order type, and user expected waiting time in different time periods and regions based on spatiotemporal data, historical order data, and real-time event data. The digital twin model uses quantitative indicators such as vehicle delay time, empty driving rate, user waiting time, and energy consumption to verify the execution effect of the scheduling strategy and generate an evaluation report.

[0032] As a further solution of the present invention: the visualization instructions generated by the output interaction module include:

[0033] A real-time route map of the vehicle with highlighted prohibited roads, charging station locations, and restricted time reminders;

[0034] The vehicle terminal interface supports communication adaptation with the new energy vehicle battery management system, the CAN protocol, Ethernet / IP protocol and MQTT protocol of the autonomous driving controller, ensuring that the dispatch instructions are accurately transmitted to the on-board terminal.

[0035] By adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are:

[0036] 1. The present invention uses a data acquisition module to acquire multi-source data such as government announcement texts and social media information in real time. The event parsing module uses natural language processing technology to extract key information about the event. The knowledge graph construction module constructs an event knowledge graph containing the association between the event and roads, vehicle types, alternative routes, etc. The constraint processing engine loads dynamic constraints such as temporary road bans and time window restrictions caused by the event in real time. The optimization decision module automatically generates scheduling strategies based on multi-agent reinforcement learning. The strategy verification module simulates and verifies the strategy effects through digital twin technology and iteratively optimizes. The entire process does not require manual intervention to adjust the rules, breaking through the limitations of traditional systems that rely on manual processing. It solves the problem that dynamic strategy adjustment in extreme scenarios is time-consuming, labor-intensive, and error-prone, and achieves a second-level response to public emergencies and fully automatic generation of scheduling strategies, greatly improving the timeliness and accuracy of the system in dealing with extreme scenarios.

[0037] 2. The present invention defines vehicles, charging piles, no-entity rules, etc. as nodes through the hypergraph modeling module, and uses hyperedges to represent complex multi-entity associations such as charging matching between vehicles and infrastructure, and restrictive relationships between constraints and roads. It constructs a spatiotemporal hypergraph model and uses a hypergraph embedding algorithm to convert the hypergraph structure into a low-dimensional vector space, so that the deep neural network can handle the vectorized expression of multiple types of constraints. The constraint processing engine uses a two-layer optimization architecture. The bottom layer processes hard constraints in real time to generate a feasible solution space, and the high layer searches for multi-objective optimal solutions. The optimization decision module automatically generates strategies based on dynamic constraints. The entire solution breaks through the limitations of traditional manual addition of "hard rules", solves the problem that algorithms cannot be automatically compatible in scenarios of multi-modal traffic coordination and multiple constraints, and realizes fully automatic modeling and strategy optimization of constraints in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of a system flow in an embodiment of the present invention;

[0039] Figure 2 Schematic diagram of the core working process of the system in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0041] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0042] Please see the attached Figure 1 -Attached Figure 2 The present invention provides a dynamic vehicle dispatch intelligent decision-making system based on big data, and the intelligent decision-making system includes:

[0043] Data acquisition module: used to collect multi-source heterogeneous data in real time, including IoT sensor data, government announcement texts, social media information, historical order data, and vehicle status data, and transmit the data to the event analysis module;

[0044] Event parsing module: This module performs natural language processing on unstructured data, extracts key information including event type, impact scope, and time window, generates a structured event description including basic event attributes, spatial impact area, and time validity interval, and transmits it to the knowledge graph construction module.

[0045] Knowledge graph construction module: This module uses events as core nodes and associates entities such as affected road nodes, restricted vehicle type nodes, and alternative route nodes to construct an event knowledge graph containing three types of relationships: "impact," "restriction," and "recommendation." The "impact" relationship represents the effect of an event on road traffic, the "restriction" relationship represents the traffic constraints imposed by an event on vehicle types, and the "recommendation" relationship represents the feasible alternative routes corresponding to the event. The event knowledge graph is then output to the constraint processing engine.

[0046] Hypergraph modeling module: This module defines vehicles, infrastructure, and constraints as nodes, and defines complex multi-entity relationships as hyperedges. This includes complex multi-entity relationships such as the charging matching relationship between vehicles and charging piles, the parking constraint relationship between parking lots and vehicles, and the relationship between prohibited driving rules and road restrictions. It constructs a spatiotemporal hypergraph model consisting of a set of vehicle nodes, a set of infrastructure nodes, and a set of constraint nodes, and transmits it to the constraint processing engine.

[0047] Constraint processing engine: This engine receives event knowledge graphs and spatiotemporal hypergraph models, extracts temporary road bans, time window restrictions, and vehicle type restrictions caused by events, as well as dynamic constraints such as charging time windows and parking lot capacity restrictions in the hypergraph model, generates a constraint set containing spatial constraints, time constraints, and vehicle attribute constraints, and loads it into the optimization decision module.

[0048] Optimization decision module: Based on the dynamic constraint set, it uses a multi-agent reinforcement learning algorithm to generate an initial scheduling strategy for each vehicle. The initial scheduling strategy includes driving routes, task allocation, docking nodes, and time arrangements, and transmits the initial scheduling strategy to the strategy verification module.

[0049] Strategy Verification Module: This module uses digital twin technology to build a virtual model that is isomorphic to the real traffic system. It simulates the execution of the initial scheduling strategy, collects data such as vehicle travel time, idle distance, task completion rate, and user waiting time, and generates strategy evaluation results that include efficiency indicators, cost indicators, and user experience indicators. The results are then fed back to the optimization decision module for strategy iteration.

[0050] Output interaction module: used to receive the final scheduling strategy generated by the optimization decision module, generate visual instructions including real-time vehicle route map, constraint condition labeling and abnormal event warning, and send them to the new energy vehicle battery management system, automatic driving controller and dispatcher monitoring interface using standardized interfaces.

[0051] In one embodiment of the present invention: in the data acquisition module, the collected IoT sensor data includes vehicle-mounted GPS trajectory data, license plate and speed data recognized by traffic cameras, charging status and remaining power data of charging piles, and the number and location data of parking lots; and government announcement texts include traffic control notices, weather warning information, and public event announcements.

[0052] In one embodiment of the present invention: the event parsing module adopts named entity recognition technology and relationship extraction technology to extract the "event-impact range-time" triple consisting of the event name, the geographical coordinate set of the affected area and the valid time interval from the unstructured text, wherein the named entity recognition technology is based on a combination model of a bidirectional long short-term memory network and a conditional random field, and the relationship extraction technology is based on a neural network model of an attention mechanism.

[0053] In one embodiment of the present invention: in the event knowledge graph constructed by the knowledge graph construction module, the node types include event nodes, road nodes, vehicle type nodes and route nodes, wherein the road nodes include attributes such as road ID, number of lanes, speed limit requirements, etc., the vehicle type nodes include attributes such as vehicle model, load, energy type, etc., the route nodes include attributes such as starting point, end point, and sequence of passing nodes, and the edge types include "influence", "restriction" and "recommendation", wherein the "influence" edge connects the event node with the road node, the "restriction" edge connects the event node with the vehicle type node, and the "recommendation" edge connects the event node with the route node.

[0054] In one embodiment of the present invention, in the spatiotemporal hypergraph model defined by the hypergraph modeling module, the node set V is strictly divided as follows:

[0055] V={V vehicle ,V Ifst ,V Ctt};

[0056] Among them, V vehicle Represents a set of vehicle nodes, each of which contains attributes such as vehicle ID, current location, remaining power, load capacity, and vehicle type. Ifst Represents a set of infrastructure nodes, including charging pile nodes, parking lot nodes, and bus station nodes, etc. Ctt Represents a set of constraint nodes, including prohibited road section nodes, time window nodes, and vehicle load limit nodes. Each hyperedge e in the hyperedge set E i Connect at least V vehicle With V Ifst and V Ctt Any node in is used to represent the interaction constraints between vehicles and infrastructure and the adaptation relationship between vehicles and constraints.

[0057] In one embodiment of the present invention, the hypergraph modeling module includes a hypergraph embedding algorithm to transform the hypergraph structure into a low-dimensional vector space. The embedding function is expressed as follows:

[0058]

[0059] Among them, h v is the d-dimensional embedding vector of node v, including node attributes and association relationship features, h u is the embedding vector of other nodes u associated with node v in the hyperedge e, E v is the set of all hyperedges containing node v, |e| is the number of nodes contained in hyperedge e, W is a d×d-dimensional trainable weight matrix used for feature transformation, and b is a d-dimensional bias vector. The algorithm is based on weighted aggregation of feature information of all nodes within the hyperedge to achieve vectorized expression of multiple types of constraints, enabling deep neural networks to handle the complex collaborative relationship between vehicles, infrastructure, and constraints.

[0060] In one embodiment of the present invention, the constraint processing engine adopts a two-layer optimization architecture including a low-level real-time constraint processing layer and a high-level global optimization layer:

[0061] The underlying real-time constraint processing layer uses an improved Dijkstra algorithm to handle constraints such as prohibited road sections, vehicle battery critical value, order cancellation events, and vehicle fault status, generating a feasible solution space that includes feasible driving routes, effective task allocation, and reasonable docking nodes.

[0062] The high-level global optimization layer is based on a deep neural network combined with a Bayesian optimization algorithm. It searches for a multi-objective optimal solution in the feasible solution space with the goals of shortest scheduling time, lowest operating cost, minimum carbon emissions, and extreme vehicle load balance. The deep neural network input is vehicle status, constraints, and historical scheduling data, and the output is the optimization parameters of each objective weight.

[0063] In one embodiment of the present invention, the optimization decision module adopts a strategy generation algorithm combining Monte Carlo tree search and reinforcement learning:

[0064] Based on the event knowledge graph, real-time constraints such as traffic control constraints, weather impact constraints, and passenger flow constraints for large-scale events are dynamically obtained. Using the multi-agent reinforcement learning algorithm, a local optimal driving strategy is generated for each vehicle that takes into account the status of surrounding vehicles, real-time road conditions, and task priorities. Based on the Monte Carlo tree search algorithm, the local strategies of all vehicles are combined and verified, and the global feasibility and goal achievement of the strategy combination are evaluated, realizing the fully automatic generation and iterative optimization of scheduling strategies without human intervention. The reward function of reinforcement learning includes multi-dimensional parameters such as time efficiency, cost control, and user satisfaction.

[0065] In one embodiment of the present invention, the digital twin model constructed by the strategy verification module includes:

[0066] Road network virtual mapping sub-model: used to synchronize road parameters such as traffic flow speed, signal light timing scheme, road congestion level and temporary construction areas in real time;

[0067] Vehicle dynamic behavior simulation sub-model: used to simulate vehicle acceleration, deceleration, lane change, parking, charging and other driving behaviors and energy consumption processes;

[0068] User demand prediction sub-model: used to predict order volume, order type and user expected waiting time in different time periods and regions based on spatiotemporal data, historical order data and real-time event data. The digital twin model uses quantitative indicators such as calculated vehicle delay time, empty driving rate, user waiting time and energy consumption value to verify the execution effect of the scheduling strategy and generate an evaluation report.

[0069] In one embodiment of the present invention, the output interaction module generates a visualization instruction including:

[0070] A real-time route map of the vehicle with highlighted prohibited roads, charging station locations, and restricted time reminders;

[0071] The vehicle terminal interface supports communication adaptation with the new energy vehicle battery management system, the CAN protocol, Ethernet / IP protocol and MQTT protocol of the autonomous driving controller, ensuring that the dispatch instructions are accurately transmitted to the on-board terminal.

[0072] In one embodiment of the present invention:

[0073] Example 1, please refer to the attached Figure 1 -Attached Figure 2 , Dynamic bus dispatching under marathon event traffic control:

[0074] From 8:00 AM to 12:00 PM on October 15, 2025, traffic control will be implemented between Guanghua Road and Jianguo Road in Chaoyang District, Beijing. Public buses will be prohibited from passing through. A detour strategy must be generated within 10 seconds for 20 affected buses to avoid entering the restricted area.

[0075] Data acquisition module, real-time data:

[0076] The boundary coordinates of the controlled road section (starting point: 39.9042°N, 116.4074°E, end point: 39.9142°N, 116.4274°E) were obtained through traffic camera AI recognition. The road ID is 123-456.

[0077] The onboard GPS reports the locations of 20 buses (e.g., vehicle A: 39.9050°N, 116.4100°E, current route: Jianguo Road straight ahead) and their vehicle types (all 12-meter gasoline buses) in real time.

[0078] The government announcement text is analyzed to obtain the control time window (2025-10-15T08:00:00 to 2025-10-15T12:00:00) and the affected road grades (urban main roads).

[0079] Historical data:

[0080] The historical congestion period in this area is the morning peak period of 7:30-9:00, with an average speed of 20km / h;

[0081] Historical data on alternative routes: The average travel time via Chaoyang Road (Road ID789) is 15 minutes longer than the original route;

[0082] Event parsing module

[0083] Use BiLSTM-CRF model to extract triples:

[0084] Event Name: "Traffic Control for Marathon Event";

[0085] Impact range: polygonal area (vertex coordinate set) containing road IDs 123-456;

[0086] Time window: 2025-10-15T08:00:00 to 2025-10-15T12:00:00.

[0087] Generate a structured description: {Event ID: 1001, Type: Traffic Control, Spatial Range: [Roads 123-456], Time: [8:00-12:00], Restricted Vehicle Type: [Bus]}.

[0088] Knowledge graph building blocks

[0089] Node creation:

[0090] Event node: Event 1001 (Type = Traffic Control, Time = 8:00-12:00);

[0091] Road node: Road 123 (ID=123, Lane=4, Speed ​​limit 60km / h), Road 789 (ID=789, lane=3, speed limit 50 km / h);

[0092] Vehicle type node: Bus 12m (Type = 12m bus, Energy = fuel);

[0093] Route node: Route Oriminal (Starting point = Guanghua Road, Ending point = Jianguo Road, Passing by = Road 123 ), Route Alternative (Starting point = Guanghua Road, Ending point = Jianguo Road, Passing by = Road 789 ).

[0094] Relationship Building:

[0095] Event 1001 →[Influence]→Road 123 ;

[0096] Event 1001 →[Restrictions]→Bus 12m ;

[0097] Event 1001 →[Recommended]→Route Alternative .

[0098] Constraint Processing Engine

[0099] Dynamic constraint generation:

[0100] Space constraints: Bus prohibited 12m Type of vehicle entering the Road 123 ;

[0101] Time constraints: effective from 8:00 to 12:00;

[0102] Alternative route constraints: Prioritize Route Alternative (Distance to original route +3 km, estimated time +15 minutes).

[0103] Two-layer optimized architecture:

[0104] Bottom layer: Improved Dijkstra algorithm excludes Road 123 , generate a feasible route for each vehicle (e.g. vehicle A turns from its current location to Chaoyang Road, passing Road 789 ), dynamically update edge weights based on real-time traffic conditions;

[0105] High-level: The deep neural network inputs vehicle location, control time, and alternative route duration, outputs target weights (time weight 0.6, cost weight 0.4), and generates the optimal detour sequence.

[0106] Optimize decision-making module

[0107] Multi-agent reinforcement learning generates a policy for each vehicle:

[0108] Vehicle A: Exit from current location at 8:05 and take Chaoyang Road (Road 789 ) Detour, estimated to arrive at the destination at 12:10, reducing the empty trip rate by 8%;

[0109] The Monte Carlo tree (by simulating the global traffic status under the strategy combination, evaluating the congestion index and task completion rate) searches and verifies the strategy combination of 20 vehicles to avoid congestion on the detour route (such as dispersing the departure time interval by 30 seconds).

[0110] Policy Validation Module

[0111] Digital Twin Simulation:

[0112] Road network virtual mapping shows Road 123 Closed, Road 789 Traffic volume increased by 20%, with average speeds dropping to 18 km / h;

[0113] Vehicle dynamic simulation shows that the average delay time after detour is 12 minutes, and the user waiting time increases by 5 minutes, both of which are within an acceptable range.

[0114] Output interaction module

[0115] Send instructions to the vehicle terminal:

[0116] Real-time road map marking Road 789 Green and feasible route, Road 123 Red prohibition;

[0117] The dispatcher interface displays the estimated arrival time of all vehicles on the new routes, and the abnormal warning threshold is set to a delay of >20 minutes.

[0118] Implementation Effect

[0119] It takes 8 seconds from event detection to policy issuance, which is 92% more efficient than manual scheduling.

[0120] None of the 20 buses mistakenly entered the controlled area, the compliance rate of the detour plan was 100%, and the number of user complaints decreased by 85%.

[0121] Example 2, please refer to the attached Figure 1 -Attached Figure 2 , multi-modal traffic coordinated scheduling under sudden large passenger flow in subway during morning peak hour:

[0122] During the morning rush hour (7:30-9:00) on December 5, 2025, People's Square Station on Shanghai Metro Line 2 was temporarily closed due to equipment failure, stranding an estimated 5,000 passengers. Within 15 minutes, 20 bus shuttles, 30 cruising taxis, 50 ride-hailing vehicles, and 800 shared bikes needed to be coordinated and dispatched to achieve a seamless connection from "metro exit - bus line - taxi / ride-hailing connection - shared bike last mile" while also meeting the following requirements:

[0123] Public transport restrictions: Shuttle buses must avoid Nanjing East Road Pedestrian Street (Road ID: 4001, large vehicles are prohibited from passing between 7:00 AM and 10:00 PM);

[0124] Taxis / online ride-hailing vehicles must wait for passengers at designated pick-up points (area ID: T01) within 200 meters of a subway station, and may not carry more than four passengers at a time.

[0125] Shared bicycle restrictions: They must be placed in parking spots (P02-P05) in areas with dense office buildings (within an 800-meter radius), with a parking capacity of ≤150 bicycles per spot.

[0126] Time constraints: The total time from passengers leaving the station to arriving at the destination is ≤ 20 minutes, and the empty trip rate is ≤ 30%;

[0127] 1. Data Acquisition Module

[0128] 1. Real-time data collection

[0129] Subway passenger flow data: The number of stranded passengers (5,000 people) and their destinations (60% for the Bund business district, 30% for Nanjing East Road office buildings, and 10% for surrounding communities) are collected through sensors at subway station gates.

[0130] Vehicle status data:

[0131] Public shuttle bus (20 vehicles): Currently located in the parking lot (coordinates: N31.2300°, E121.4800°), with a maximum capacity of 50 people per vehicle and a range of ≥ 200 kilometers;

[0132] Cruising taxis (30 vehicles): 40% empty, average distance from People's Square Station 3-5 km;

[0133] Online car-hailing (50 vehicles): The platform's real-time order list shows 200 available vehicles within a 10-kilometer radius;

[0134] Shared bicycles (800): 500 in stock at the warehouse, and 300 available at nearby parking spots P02-P05;

[0135] Government announcement text: Large vehicles are prohibited from passing through Nanjing East Road Pedestrian Street (ID: 4001) all day (the prohibition rules are extracted through the event analysis module).

[0136] Historical data reference

[0137] The Bund business district's morning peak congestion index was 8.5 (severe congestion), with an average vehicle speed of 15 km / h;

[0138] The average ride-hailing service from People's Square to the Bund takes 12 minutes, with a 25% empty trip rate.

[0139] The average turnover rate of shared bicycles in areas with dense office buildings is 3 times per hour.

[0140] 2. Event Analysis and Knowledge Graph Construction

[0141] Event parsing module

[0142] Extract triples:

[0143] Event name: "Metro Line 2 People's Square Station is temporarily closed";

[0144] Impact area: 1 km around People's Square subway station (geographic coordinates set);

[0145] Time window: 2025-12-05T07:30:00 to T09:00:00;

[0146] Generate a structured description: {Event ID: 3001, Type: Public transportation failure, Affected population: 5,000 people, Recommended connection method: Bus + Taxi / Online car-hailing + Shared bike}.

[0147] 2. Knowledge graph construction (knowledge graph construction module)

[0148] Node extension:

[0149] Vehicle type node: Added Taxi 巡游出租车 (Type = Compact, 4-passenger capacity), RideHailling 网约车 (Type = medium, maximum capacity 5 people);

[0150] Infrastructure nodes: boarding point T01 (capacity = 20 vehicles), parking points P02-P05 (total capacity = 600 vehicles);

[0151] Demand node: Demand 02 (Stranded passengers at People's Square Station, heading to The Bund / Nanjing East Road);

[0152] Event 3001 →[Influence]→Road 4001 (Public bus shuttles are prohibited);

[0153] Event 3001 →[Recommended]→RouteA 01(Public bus detour via Beijing East Road, ID=4002);

[0154] Demand 02 →[Match]→Bus 接驳车 (Carrying ≥50 people / vehicle), Taxi 巡游出租车 (carrying ≤ 4 people / vehicle), Bike 共享单车 (coverage ≤ 1 km)

[0155] 3. Hypergraph Modeling and Constraint Processing

[0156] Hypergraph model construction (hypergraph modeling module)

[0157] Node collection:

[0158] V vehicle : Includes bus shuttles, cruising taxis, online ride-hailing vehicles, and shared bicycle nodes, with attributes such as location, passenger status, and battery life;

[0159] V Ifst : Pick-up point T01, parking spots P02-P05, charging piles (for new energy online ride-hailing vehicles);

[0160] V Ctt :Road 4001 ), time window (7:30-9:00), passenger capacity limit (taxi ≤ 4 people);

[0161] Hyperedge definition:

[0162] Super Edge E5: Connecting bus shuttles and roads 4002 , time window, indicating "buses will detour via Beijing East Road from 7:30 to 9:00";

[0163] Superedge E6: connects online ride-hailing vehicles, pick-up point T01, and passenger capacity limit, indicating that "online ride-hailing vehicles must wait at T01 and can only carry ≤ 4 passengers per trip";

[0164] Superedge E7: connects shared bicycles, parking spots P02-P05, and the spatial constraint, indicating that "shared bicycles must be parked at P02-P05, with ≤150 bicycles per spot."

[0165] Constraint Processing Engine

[0166] Dynamic constraint collection:

[0167] Spatial constraints: Buses are prohibited from entering the Road 4001 , taxis / online ride-hailing vehicles are only allowed to pick up passengers at T01;

[0168] Time constraint: The first batch of 200 shared bikes must be deployed before 7:45;

[0169] Resource constraints: Pickup point T01 can accommodate a maximum of 20 taxis / online ride-hailing vehicles at a time;

[0170] Two-layer optimized architecture:

[0171] Low-level real-time constraint processing:

[0172] Bus shuttle route: Parking lot → Beijing East Road (ID=4002) → The Bund (avoiding restricted areas), estimated journey time: 18 minutes;

[0173] Taxi / online ride-hailing routes: Use an improved Dijkstra algorithm to plan the shortest route to avoid congestion (such as bypassing the Yan'an Elevated Road);

[0174] High-level global optimization:

[0175] Objective function: minimize total evacuation time (weight 0.5) + minimize empty driving rate (weight 0.3) + maximize shared bicycle utilization rate (weight 0.2);

[0176] Deep neural network output: Buses depart every 5 minutes, taxi / ride-hailing orders are prioritized by distance, and shared bikes are replenished by 50 every 10 minutes to P02-P05.

[0177] 4. Optimization Decision-Making and Strategy Verification

[0178] Optimize decision-making module

[0179] Multi-agent reinforcement learning:

[0180] Public transportation strategy: The first bus will depart at 7:40 a.m., with buses departing every 5 minutes for a total of 10 trips, carrying 500 people to the Bund Bus Hub;

[0181] Taxi / Ride-hailing Strategy:

[0182] Cruising taxis: 10 will go to T01 to pick up passengers, with each taxi picking up 4 people per trip, with 3 round trips (estimated to evacuate 120 people);

[0183] Ride-hailing services: 30 vehicles were dispatched to the vicinity of T01, with priority given to orders from the Nanjing East Road office building (estimated evacuation capacity: 150 people).

[0184] Bike-sharing strategy: 50 bikes will be deployed from the warehouse to P02-P05 at 7:45 am to meet the last-mile demand (expected to cover 200 people).

[0185] Monte Carlo Tree Search Verification:

[0186] Simulate the coordination of various modes: public transportation takes 50% of the passenger flow, taxis / online ride-hailing services take 30%, and shared bicycles take 20%;

[0187] Global feasibility: There is no congestion at the T01 pick-up point (average waiting time for vehicles ≤ 3 minutes), and the utilization rate of shared bicycle parking spots is 80% (≤ 150 bikes / spot).

[0188] Strategy Verification Module (Digital Twin Simulation)

[0189] Virtual scene parameters:

[0190] Road network: Traffic volume on Beijing East Road increased by 40%, with an average speed of 18 km / h;

[0191] Passenger behavior: 70% choose public transportation + shared bicycles, 30% choose taxis / online ride-hailing services;

[0192] Simulation results:

[0193] Average evacuation time: 17 minutes (≤20 minutes target);

[0194] Empty ride rate: Buses 0% (full), taxis / ride-hailing 28% (≤30% target), shared bike turnover rate 2.8 times / hour (close to the historical average of 3 times);

[0195] Abnormal warning: None (all vehicles did not violate the restriction rules).

[0196] 5. Output Interaction and Implementation Effect

[0197] Output interaction module

[0198] Vehicle terminal instructions:

[0199] Bus shuttle: Navigation shows "Beijing East Road → Bund Bus Hub", prohibited road section Road 4001 Red highlight;

[0200] Taxi / online ride-hailing service: The app pushes a notification that "T01 pick-up point takes priority" and displays the queue length in real time;

[0201] Shared bicycle dispatch system: Sends instructions to the warehouse: "Add 50 bicycles to P02-P05 at 7:45, 8:00, and 8:15";

[0202] Dispatcher interface:

[0203] Real-time monitoring: bus locations (the first bus arrived at the Bund at 7:48), taxi / online ride-hailing waiting list (15 currently in T01), and shared bicycle parking numbers (450 in total in P02-P05);

[0204] Visual report: Passenger evacuation progress is updated every 10 minutes (2,800 people have been evacuated at 7:50, 56% completed).

[0205] According to the embodiments, the present invention realizes fully automatic generation of scheduling strategies and second-level response in extreme scenarios through multi-module collaboration, breaking through the limitations of traditional manual processing, and using hypergraph modeling, hypergraph embedding algorithm and two-layer optimization architecture to break through the limitations of manually adding "hard rules". The entire set of technologies solves the problems of extreme scenario response and complex scenario modeling through multi-module collaboration and innovative algorithms, improves the system's ability to adapt to dynamic constraints and multi-mode collaboration efficiency, and achieves intelligent, efficient and accurate scheduling effects.

[0206] index The present invention Traditional manual scheduling CN112487126A solution Response time ≤10 seconds ≥5 minutes 30 seconds Complex constraint processing Automatically compatible with 15 categories Only supports 3 types Support 8 categories Scheduling compliance rate 98.7％ 82％ 95％

[0207] Table 1

[0208] Table 1 shows the effects of this system (e.g., “response in 8 seconds”), which is not compared with traditional methods or existing patented solutions.

[0209] Although the present invention is disclosed above with reference to preferred embodiments, this is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent variations, and modifications made to the above embodiments in accordance with the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection defined by the claims of the present invention.

Claims

1. A dynamic vehicle dispatch intelligent decision-making system based on big data, characterized by: The intelligent decision-making system includes: Event parsing module: used to perform natural language processing on unstructured data, extract key information including event type, impact scope and time window, and generate structured event descriptions; Knowledge graph construction module: This module uses events as core nodes, associates affected road nodes, restricted vehicle type nodes, and alternative route node entities, and constructs an event knowledge graph containing three types of relationship edges: "impact," "restriction," and "recommendation." Hypergraph modeling module: This module defines vehicles, infrastructure, and constraints as nodes, defines complex multi-entity relationships as hyperedges, and constructs a spatiotemporal hypergraph model consisting of a set of vehicle nodes, a set of infrastructure nodes, and a set of constraint nodes. Constraint processing engine: used to receive event knowledge graphs and spatiotemporal hypergraph models, extract temporary road bans, time window restrictions, and vehicle type restrictions caused by events, and generate a constraint set containing spatial constraints, time constraints, and vehicle attribute constraints; Optimization decision module: used to generate the initial scheduling strategy for each vehicle based on the dynamic constraint set through multi-agent reinforcement learning algorithm; Strategy Verification Module: This module uses digital twin technology to build a virtual model that is isomorphic to the real traffic system, simulates the execution process of the initial scheduling strategy, and generates strategy evaluation results. Output interaction module: used to receive the final dispatch strategy, generate visual instructions, and send them to the new energy vehicle battery management system, autonomous driving controller and dispatcher monitoring interface. Data acquisition module: used to collect multi-source heterogeneous data in real time and transmit it to the event analysis module and hypergraph modeling module.

2. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 1, characterized in that: The data acquisition module is used to collect multi-source heterogeneous data in real time, including IoT sensor data, government announcement texts, social media information, historical order data and vehicle status data. The IoT sensor data collected in the data acquisition module includes vehicle-mounted GPS trajectory data, license plate and speed data recognized by traffic cameras, charging status and remaining power data of charging piles, and the number and location data of parking spaces. The government announcement texts include traffic control notices, weather warning information and public event announcements.

3. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 2, characterized in that: The event parsing module uses named entity recognition technology and relationship extraction technology to extract "event-impact range-time" triples consisting of the event name, the geographical coordinate set of the affected area, and the effective time interval from unstructured text. The named entity recognition technology is based on a combination model of a bidirectional long short-term memory network and a conditional random field, and the relationship extraction technology is based on a neural network model with an attention mechanism.

4. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 3, characterized in that: In the event knowledge graph constructed by the knowledge graph construction module, node types include event nodes, road nodes, vehicle type nodes and route nodes, and edge types include "influence", "restriction" and "recommendation", among which the "influence" edge connects the event node and the road node, the "restriction" edge connects the event node and the vehicle type node, and the "recommendation" edge connects the event node and the route node.

5. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 4 is characterized by: In the spatiotemporal hypergraph model defined by the hypergraph modeling module, the node set V is strictly divided as follows: V={V vehicle ,V Ifst ,V Ctt ]; Among them, V vehicle Represents a set of vehicle nodes, each of which contains vehicle ID, current location, remaining power, load capacity, and vehicle type attributes. Ifst Represents a set of infrastructure nodes, including charging pile nodes, parking lot nodes, and bus station nodes. Ctt Represents a set of constraint nodes, including prohibited road section nodes, time window nodes, and vehicle load limit nodes. Each hyperedge e in the hyperedge set E i Connect at least V vehicle With V Ifst and V Ctt Any node in is used to represent the interaction constraints between vehicles and infrastructure and the adaptation relationship between vehicles and constraints.

6. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 5, characterized in that: The hypergraph modeling module includes a hypergraph embedding algorithm to transform the hypergraph structure into a low-dimensional vector space. The embedding function is expressed as follows: Among them, h v is the d-dimensional embedding vector of node v, including node attributes and association relationship features, h u is the embedding vector of other nodes u associated with node v in the hyperedge e, E v is the set of all hyperedges containing node v, |e| is the number of nodes contained in hyperedge e, W is a d×d-dimensional trainable weight matrix used for feature transformation, and b is a d-dimensional bias vector.

7. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 6 is characterized in that: The constraint processing engine adopts a two-layer optimization architecture consisting of a low-level real-time constraint processing layer and a high-level global optimization layer: The underlying real-time constraint processing layer uses an improved Dijkstra algorithm to handle constraints such as prohibited road sections, vehicle battery critical value, order cancellation events, and vehicle fault status, generating a feasible solution space that includes feasible driving routes, effective task allocation, and reasonable docking nodes. The high-level global optimization layer is based on a deep neural network combined with a Bayesian optimization algorithm. It searches for a multi-objective optimal solution in the feasible solution space with the goals of shortest scheduling time, lowest operating cost, minimum carbon emissions, and extreme vehicle load balance. The deep neural network input is vehicle status, constraints, and historical scheduling data, and the output is the optimization parameters of each objective weight.

8. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 7 is characterized in that: The optimization decision module adopts a strategy generation algorithm that combines Monte Carlo tree search and reinforcement learning: Based on the event knowledge graph, real-time constraints such as traffic control constraints, weather impact constraints, and passenger flow constraints for large-scale events are dynamically obtained. A multi-agent reinforcement learning algorithm is used to generate a local optimal driving strategy for each vehicle that takes into account the status of surrounding vehicles, real-time road conditions, and task priorities. Based on the Monte Carlo tree search algorithm, the local strategies of all vehicles are combined and verified to evaluate the global feasibility and goal achievement of the strategy combination. The reward function of reinforcement learning includes multi-dimensional parameters such as time efficiency, cost control, and user satisfaction.

9. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 8, characterized in that: The digital twin model constructed by the strategy verification module includes: Road network virtual mapping sub-model: used to synchronize traffic flow speed, signal light timing plan, road congestion level and temporary construction area road parameters in real time; Vehicle dynamic behavior simulation sub-model: used to simulate vehicle acceleration, deceleration, lane change, parking and charging driving behaviors and energy consumption processes; User demand prediction sub-model: This model is used to predict order volume, order type, and user expected waiting time in different time periods and regions based on spatiotemporal data, historical order data, and real-time event data. The digital twin model uses quantitative indicators such as vehicle delay time, empty driving rate, user waiting time, and energy consumption to verify the execution effect of the scheduling strategy and generate an evaluation report.

10. The big data-based dynamic vehicle dispatch intelligent decision-making system according to claim 9, characterized in that: The visualization instructions generated by the output interaction module include: A real-time route map of the vehicle with highlighted prohibited roads, charging station locations, and restricted time reminders; Vehicle fault warning, order surge warning, low battery warning and task timeout warning abnormal event prompt information; the vehicle terminal interface supports communication adaptation with the new energy vehicle battery management system, automatic driving controller's CAN protocol, Ethernet / IP protocol and MQTT protocol.

Images