City traffic system simulation method based on digital twinning

Abstract

The application relates to the field of computers and digital twinning technology, and discloses a city traffic system simulation method based on digital twinning. The method comprises the following steps: constructing a static road network model comprising roads, intersections, lanes and signal phases; fusing multi-source real-time traffic data; initializing and continuously updating a high-fidelity digital twin; deploying a multi-agent path planning module in the twin, dynamically re-planning paths and outputting traffic prediction; taking the prediction as a feedforward input to a signal coordination control module based on reinforcement learning to generate a globally optimized signal timing scheme; and returning response data after execution in the physical world to calibrate the model, forming a closed-loop optimization. The application realizes the unification of maximum road network traffic efficiency and individual travel experience optimization by deeply coupling signal control and path induction.

Description

Technical Field

[0001] This invention belongs to the field of computer and digital twin technology, specifically relating to a simulation method for urban transportation systems based on digital twins. Background Technology

[0002] With accelerated urbanization and a continuous increase in the number of motor vehicles, urban transportation systems face increasingly severe problems of congestion, inefficiency, and carbon emissions. Intelligent Transportation Systems (ITS), as a means of alleviating these problems, rely heavily on two major technical modules: signal control and navigation route guidance. Signal control systems optimize local traffic capacity by adjusting intersection phases and timings, while navigation platforms provide route suggestions to individual vehicles based on real-time or historical traffic data to avoid congestion. However, currently, these two types of systems generally operate independently in actual deployments, lacking information sharing and decision-making coordination mechanisms, resulting in limited overall road network control effectiveness.

[0003] Building high-fidelity traffic simulation environments based on the digital twin concept has become an important direction for achieving refined traffic management. Digital twins, by integrating the physical world with virtual models, support real-time mapping and dynamic simulation of traffic flow states. Signal control strategies can be optimized online using the simulation environment, while navigation guidance can also utilize virtual road networks to pre-simulate the consequences of path selection. However, existing digital twin simulation platforms mostly focus on modeling single functional modules, failing to effectively integrate the interactive feedback mechanism between signal control agents and vehicle group behavior, making it difficult to support global collaborative optimization.

[0004] In existing technologies, signal control algorithms typically perform local optimization based solely on traffic flow data from the intersection or nearby detectors, neglecting the large-scale traffic redistribution effect caused by navigation guidance. Mainstream navigation applications employ short-sighted shortest or fastest path strategies, failing to consider the impact of collective behavior on signal control stability, easily leading to oscillating phenomena of "guidance-congestion-re-guidance" in hotspot areas. Furthermore, vehicle route selection inherently constitutes a non-cooperative game process, and existing simulation methods lack the ability to model the joint convergence of collective agent strategies and signal control strategies, resulting in discrepancies between simulation results and real road network dynamics. Therefore, there is an urgent need for a digital twin simulation method for urban traffic that can deeply integrate signal control agents and vehicle collective agents, achieving bidirectional feedback and collaborative evolution, to overcome the current system's fragmented nature, delayed response, and global suboptimal performance. Summary of the Invention

[0005] This invention provides a digital twin-based simulation method for urban traffic systems, aiming to solve the technical problem that existing signal control and navigation path guidance systems are disconnected and lack a coordination mechanism, resulting in the inability to achieve dynamic optimization of global traffic flow. This method constructs a high-fidelity, full-element, real-time interactive digital twin of urban traffic, deeply coupling the traffic signal control system in the physical world with a multi-agent path guidance model in the virtual world. This forms a closed-loop feedback, bidirectional driving collaborative optimization architecture, thereby maximizing road network efficiency at the macro level and ensuring the rationality of individual travel experience at the micro level.

[0006] This invention provides a digital twin-based urban transportation system simulation method, comprising: Construct a static topology model of the urban road network; Access real-time traffic flow data sources; Based on the static topology model and the real-time traffic flow data source, an dynamically evolving urban traffic digital twin is initialized and continuously updated. The urban traffic digital twin synchronizes the traffic status of the physical world once per second and uses an embedded microscopic traffic simulation engine to extrapolate and complete the traffic flow in unobserved areas. A multi-agent path planning module is deployed within the urban transportation digital twin. This module assigns an initial path to each virtual travel agent entering the road network and dynamically replans the path based on real-time travel time, queue length, and signal waiting time within the urban transportation digital twin. The next-time road network traffic distribution prediction result output by the multi-agent path planning module is transmitted to the signal control strategy generation module as a feedforward input signal. The signal control strategy generation module adopts a multi-intersection cooperative control algorithm based on reinforcement learning. The optimization objective is to minimize the total delay and total number of stops of vehicles in the entire road network. It combines the current signal phase status, the predicted value of the queue length of each approach lane and the prediction result of the feedforward traffic distribution to generate the green ratio and phase switching sequence of each intersection in the next control cycle. The control commands output by the signal control strategy generation module are sent to the signal controller in the physical world for execution. At the same time, the actual traffic response data after execution is sent back to the urban traffic digital twin for calibration of simulation model parameters and initiation of the next round of collaborative optimization cycle.

[0007] Preferably, the static topology model for constructing the urban road network specifically includes: Extract road centerline vector data from the city's basic geographic information system, and label road segments with attributes based on road grade, design speed, and number of lanes; Identify all at-grade intersections and determine the number of left-turn, straight-ahead, and right-turn lanes for each approach based on the channelization design drawings; Based on the traffic signal timing scheme document, establish a signal phase-lane mapping table for each intersection, and clarify the set of lanes served by each phase; All road segments, nodes, lanes, and phase information are integrated into a directed graph data structure. Nodes represent intersections, edges represent road segments, and the weight attributes of the edges include free-flow travel time, maximum capacity, and current saturation.

[0008] Preferably, accessing real-time traffic flow data sources specifically includes: The system uses the vehicle detection interface built into the traffic signal controller to acquire the number of arriving vehicles, departing vehicles, and queue overflow status of each approach lane section every 5 seconds. By deploying video analytics servers at intersections, the surveillance video stream is processed frame by frame, and the instantaneous speed, vehicle type classification, and queue length estimate of each lane are output. Receive de-identified trajectory point sequences from partner ride-hailing platforms and logistics fleet management systems, and convert them into average travel speeds for road segments after spatiotemporal matching; By integrating real-time traffic flow data sources and using a Kalman filter to fuse different observations of the same physical quantity, a traffic state estimation vector with a unified spatiotemporal reference is generated.

[0009] Preferably, the initialized and continuously updated dynamically evolving urban transportation digital twin specifically includes: A microscopic traffic simulation kernel based on cellular automata is adopted, dividing each road segment into several cellular units with a length of 7.5 meters, each cell accommodating a standard passenger car; Each vehicle agent entering the simulation is assigned a unique identifier, origin, destination, desired speed, and car-following sensitivity parameters. At the beginning of each simulation step, the observed vehicle positions and speeds are forcibly synchronized based on the received real-time traffic flow data, while the state of unobserved areas is deduced based on the car-following model and the lane-changing model. The simulation kernel runs with a basic time step of 100 milliseconds and outputs a traffic state snapshot synchronized with the physical world every 10 time steps for use by the upper-level decision-making module.

[0010] Preferably, the multi-agent path planning module assigns an initial path to each virtual travel agent entering the road network, and dynamically replans the path based on the real-time travel time, queue length, and signal waiting time within the urban traffic digital twin. Specifically, this includes: When a new travel request is injected into the city's traffic digital twin, the improved Dijkstra algorithm is invoked to calculate multiple alternative routes, with its origin and destination as endpoints. The path cost function of the improved Dijkstra algorithm comprehensively considers the free-flow travel time of the road segment, the current saturation-converted delay, and the expected red light waiting time at downstream intersections. During the vehicle's operation, at each intersection, the route selection for the remaining journey is reassessed. If a new route has an estimated total journey time that is more than 15% shorter than the current route, a route switch is triggered. All route replanning decisions are made within the city's traffic digital twin, and its output traffic allocation matrix for each road segment in the next 15 minutes is encapsulated as a feedforward signal.

[0011] Preferably, the signal control strategy generation module adopts a multi-intersection cooperative control algorithm based on reinforcement learning, with the optimization objective of minimizing the total vehicle delay and total number of stops across the entire road network. Combining the current signal phase state, predicted queue lengths for each approach lane, and predicted feedforward traffic distribution, it generates the green light ratio and phase switching sequence for each intersection in the next control cycle, specifically including: The entire control area is modeled as a Markov decision process, whose state space consists of the queue length of each approach lane at each intersection, the remaining time of the current phase, and the feedforward flow prediction vector, and whose action space is the phase switching instruction set of each intersection. A deep Q-network is used as the policy network. Its input is a state space vector and its output is the expected cumulative reward value of each action. The reward function is defined as the negative weighted sum of the total delay of vehicles and the total number of stops in the entire road network. The policy network is trained online through continuous interaction with the city's traffic digital twin. An experience replay sample is collected after each control cycle to update the network parameters. At the beginning of each control cycle, the policy network outputs the optimal action based on the current state. This optimal action is parsed into a specific green ratio allocation scheme and a phase switching timetable.

[0012] Preferably, the actual traffic response data after execution is fed back to the urban traffic digital twin for calibrating simulation model parameters and initiating the next round of collaborative optimization loop, specifically including: After the signal control command is executed, the actual vehicle passage data of each intersection is collected, and the measured queue dissipation time, saturation flow rate and travel time are calculated. These measured indicators are compared with the simulation output of the urban traffic digital twin during the same period, and the error vector is calculated. Based on the error vector, the gradient descent method is used to adjust the car-following model parameters, lane-changing intention threshold, and driver reaction time constant in the micro-simulation kernel, so that the simulation output gradually approximates the real response in the physical world. After parameter calibration is completed, the system automatically enters the next collaborative optimization cycle.

[0013] Preferably, the microscopic traffic simulation engine uses an improved intelligent driver model as the car-following model, and its acceleration... The calculation formula is: ; This is the maximum acceleration; The vehicle speed at the current moment; For the desired speed; The acceleration index; This represents the current distance between the front of the vehicle; It is the relative velocity; This represents the desired front-end spacing.

[0014] Preferably, the path cost function Defined as: ; For path, For road segments on the path, For free-flow travel time on the road segment, To delay waiting in line, Signal waiting time and These are the weighting coefficients.

[0015] Preferably, the reward function The expression is: ; Total vehicle delays across the entire road network, expressed in seconds; This represents the total number of stops for vehicles across the entire road network. and These are the corresponding weighting coefficients.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by constructing a high-fidelity, full-element, real-time interactive digital twin of urban traffic, has for the first time achieved deep coupling and collaborative optimization of the two major subsystems of signal control and route guidance.

[0017] 2. Urban traffic digital twins can not only accurately reproduce the current traffic conditions, but also make high-precision predictions of future traffic flows, thereby providing reliable feedforward information for signal control and avoiding the lag of traditional inductive control that relies solely on historical data.

[0018] 3. Route guidance decisions are completed within the urban traffic digital twin. The replanning behavior fully considers the expected effects of the signal control strategy, eliminating the "oscillation" effect and secondary congestion caused by blind guidance. The signal control strategy generation module introduces a reinforcement learning algorithm, which can learn online and adapt to complex traffic dynamics. Its optimization objective is directly aimed at global performance indicators, rather than single-point or local optima.

[0019] 4. The entire system forms a complete closed loop of "perception-simulation-decision-execution-learning", which enables urban traffic management to shift from passive response to proactive guidance, improves the overall traffic efficiency of the road network, reduces the average travel time and number of stops for vehicles, and alleviates urban traffic congestion. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention. Figure 2 This is a schematic diagram of the core principle framework of deep coupling and collaborative optimization of signal control and path induction in this invention; Figure 3 This is a logical flowchart of the fusion of static road network modeling and multi-source real-time traffic flow data in this invention. Figure 4 This is a flowchart illustrating the logical process of constructing and updating the dynamically evolving urban traffic digital twin in this invention. Figure 5 This is a logical flowchart of the multi-agent path planning and signal cooperative control feedforward interaction in this invention. Figure 6 This is a schematic diagram of the closed-loop feedback calibration and online learning interaction relationship and data flow between the physical world and the digital twin of urban transportation in this invention. Detailed Implementation

[0021] refer to Figures 1 to 6 This invention provides a digital twin-based simulation method for urban traffic systems, aiming to solve the technical problem that existing signal control and navigation path guidance systems are disconnected and lack a coordination mechanism, resulting in the inability to achieve dynamic optimization of global traffic flow. This method constructs a high-fidelity, full-element, real-time interactive digital twin of urban traffic, deeply coupling the traffic signal control system in the physical world with a multi-agent path guidance model in the virtual world. This forms a closed-loop feedback, bidirectional driving collaborative optimization architecture, thereby maximizing road network efficiency at the macro level and ensuring the rationality of individual travel experience at the micro level.

[0022] The method includes the following steps: S1, Construct a static topological model of the urban road network; S2 connects to real-time traffic flow data sources; S3. Based on the static topology model and the real-time traffic flow data source, initialize and continuously update the dynamically evolving urban traffic digital twin. S4. Deploy a multi-agent path planning module within the urban traffic digital twin; S5, the next moment's road network traffic distribution prediction result output by the multi-agent path planning module is transmitted to the signal control strategy generation module as a feedforward input signal; S6, the signal control strategy generation module adopts a multi-intersection collaborative control algorithm based on reinforcement learning to generate the green light ratio and phase switching timing of each intersection in the next control cycle; S7, the control command output by the signal control strategy generation module is sent to the signal controller in the physical world for execution, and the actual traffic response data after execution is sent back to the urban traffic digital twin for calibration of simulation model parameters and starting the next round of collaborative optimization cycle.

[0023] In step S1, a static topology model of the urban road network is constructed. This static topology model includes all road segments, intersection nodes, lane attributes, signal phase configurations, and geospatial coordinate information. Specifically, road centerline vector data is extracted from the urban basic geographic information system, and road segments are labeled with attributes based on road grade, design speed, and number of lanes. Road grades are divided into four categories: expressways, arterial roads, secondary arterial roads, and local roads, with corresponding design speeds of 80 km / h, 60 km / h, 40 km / h, and 30 km / h, respectively. The number of lanes is determined based on field survey data, ranging from 1 to 6. All at-grade intersections are identified, and the number of left-turn, straight-ahead, and right-turn lanes for each approach is determined based on channelization design drawings. For each intersection, the total number of approach lanes is counted, typically 3 to 5, with each approach lane containing 1 to 3 lanes.

[0024] Based on the traffic signal timing scheme document, a signal phase-lane mapping table is established for each intersection, clarifying the set of lanes served by each phase. For example, in a four-phase intersection, Phase 1 serves the eastbound and westbound through lanes, Phase 2 serves the southbound and northbound left-turn lanes, Phase 3 serves the southbound and northbound through lanes, and Phase 4 serves the eastbound and westbound left-turn lanes. All road segments, nodes, lanes, and phase information are integrated into a directed graph data structure. Nodes represent intersections, and edges represent road segments. The weight attributes of the edges include free-flow travel time, maximum capacity, and current saturation. The free-flow travel time is the ratio of road segment length to design speed, with a passenger car equivalent of 1800 vehicles per hour per lane, and the initial current saturation value is set to 0.5.

[0025] In step S2, a real-time traffic flow data source is accessed. This data source includes geomagnetic coil detectors, video structured analysis units, floating car trajectory data, and travel speed data reported by an internet map platform. Specifically, the number of arriving vehicles, departing vehicles, and queue overflow status at each approach lane section are acquired every 5 seconds via the vehicle detection interface built into the traffic signal controller. The number of arriving vehicles is counted by the geomagnetic coil detectors, and the number of departing vehicles is confirmed by the downstream geomagnetic coils or video analysis units. The queue overflow status is defined as the queue length exceeding 90% of the approach lane's storage capacity. A video analysis server deployed at the intersection processes the monitoring video stream frame by frame, outputting the instantaneous speed of each lane, vehicle type classification, and estimated queue length.

[0026] The video analysis employs a convolutional neural network-based target detection and tracking algorithm with a frame rate of at least 25 Hz. Vehicle types are categorized into passenger cars, buses, and trucks, with instantaneous speed accuracy within ±5 km / h. Anonymized trajectory point sequences are received from partner ride-hailing platforms and logistics fleet management systems, and converted into average travel speeds after spatiotemporal matching. The timestamp accuracy of the trajectory point sequences is 1 second, and the spatial coordinate accuracy is 5 meters. Spatiotemporal matching uses a dynamic time warping algorithm to map trajectory points to the nearest road segment. Integrating the above multi-source heterogeneous data, a Kalman filter is used to fuse different observations of the same physical quantity, generating a traffic state estimation vector with a unified spatiotemporal reference. The Kalman filter's state vector includes three dimensions: average road segment speed, approach lane queue length, and intersection saturation. The process noise covariance matrix and observation noise covariance matrix are statistically set based on historical data to ensure the confidence level of the fusion result is higher than 95%.

[0027] In step S3, based on the static topology model and the real-time traffic flow data source, a dynamically evolving urban traffic digital twin is initialized and continuously updated. This digital twin synchronizes with the physical world's traffic state once per second and uses an embedded microscopic traffic simulation engine to extrapolate and complete traffic flow in unobserved areas. A microscopic traffic simulation kernel based on cellular automata is used, dividing each road segment into several 7.5-meter-long cellular units, each capable of accommodating a standard passenger car. The cell length is determined by adding a 3-meter safety margin to the standard passenger car length of 4.5 meters. Each vehicle agent entering the simulation is assigned a unique identifier, origin, destination, desired speed, and car-following sensitivity parameter. The unique identifier is a 16-digit hexadecimal string, the origin and destination are intersection node numbers, the desired speed is randomly generated within the range of 30 to 80 kilometers per hour, and the car-following sensitivity parameter follows a normal distribution with a mean of 1.5 and a standard deviation of 0.3.

[0028] At the start of each simulation step, based on the received real-time traffic flow data, the observed vehicle positions and speeds are forcibly synchronized. For unobserved areas, state deduction is performed based on the car-following model and the lane-changing model. The car-following model employs an improved intelligent driver model with acceleration... The calculation formula is: ; This is the maximum acceleration; The vehicle speed at the current moment; For the desired speed; The acceleration index; This represents the current distance between the front of the vehicle; It is the relative velocity; The formula for calculating the desired front-end spacing is: ; The minimum safe distance is set at 2.5 meters. For safety, a time interval of 1.6 seconds is used; For comfortable deceleration, a value of 2.0 meters per second squared is used. The lane-changing model employs a decision-making mechanism based on utility theory, where vehicles trigger lane-changing behavior under conditions that satisfy safety clearance, speed gain, and lane preference. The simulation kernel runs with a basic time step of 100 milliseconds, outputting a traffic state snapshot synchronized with the physical world every 10 time steps for use by the upper-level decision-making module. The state snapshot contains the position, speed, lane number, and direction of travel of all vehicles, in structured text format, and is generated once per second.

[0029] In step S4, a multi-agent path planning module is deployed within the urban traffic digital twin. The multi-agent path planning module assigns an initial path to each virtual travel agent entering the road network and dynamically replans the path based on the real-time travel time, queue length, and signal waiting time within the urban traffic digital twin.

[0030] When a new travel request is injected into the city's traffic digital twin, using its origin and destination as endpoints, an improved Dijkstra algorithm is invoked to calculate multiple alternative routes. The improved Dijkstra algorithm's path cost function comprehensively considers the free-flow travel time of road segments, the current saturation-based delay, and the expected red-light waiting time at downstream intersections. Path cost function Defined as: ; For path, For road segments on the path, For free-flow travel time on the road segment, To delay waiting in line, Signal waiting time and These are weighting coefficients, with values ​​of 0.7 and 0.3 respectively. Queue delay time. Based on the current saturation of the road segment The calculation yields the following formula: Signal waiting time The prediction is based on the remaining time and cycle length of the current phase at the downstream intersection. During the vehicle's proxy driving process, the route selection for its remaining journey is reassessed every time it passes an intersection node. If a new route has an estimated total journey time that is more than 15% shorter than the current route, a route switch is triggered.

[0031] Route switching decisions are made independently by the route planning module, without external intervention. All route replanning decisions are completed within the city's traffic digital twin, and its output, the traffic allocation matrix for each road segment in the next 15 minutes, is encapsulated as a feedforward signal. The traffic allocation matrix is ​​a two-dimensional array, with row indices representing time periods (one row per minute), column indices representing road segment numbers, and element values ​​representing the number of vehicles expected to pass through that road segment within that time period.

[0032] In step S5, the predicted road network traffic distribution at the next time step, output by the multi-agent path planning module, is transmitted as a feedforward input signal to the signal control strategy generation module. The feedforward signal is transmitted through a message queue middleware with a transmission delay of no more than 200 milliseconds. After receiving the feedforward signal, the signal control strategy generation module fuses it with the current traffic state data to form a complete control decision input.

[0033] In step S6, the signal control strategy generation module employs a reinforcement learning-based multi-intersection cooperative control algorithm. The optimization objective is to minimize the total vehicle delay and total number of stops across the entire road network. Combining the current signal phase state, predicted queue lengths for each approach lane, and feedforward traffic distribution prediction results, it generates the green light ratio and phase switching sequence for each intersection in the next control cycle. The entire control area is modeled as a Markov decision process. Its state space consists of the queue lengths for each approach lane at each intersection, the remaining time of the current phase, and the feedforward traffic prediction vector. The action space is the phase switching instruction set for each intersection. The state space dimension is the number of intersections multiplied by the number of approach lanes plus the number of intersections, and the action space size is the product of the number of phases at each intersection.

[0034] A deep Q-network is used as the policy network. Its input is a state space vector, and its output is the expected cumulative reward value for each action. The reward function is defined as the negative weighted sum of the total vehicle delay and the total number of stops across the entire road network. The expression is: ; Total vehicle delays across the entire road network, expressed in seconds; This represents the total number of stops for vehicles across the entire road network. and The corresponding weighting coefficients are set to 0.6 and 0.4, respectively. The policy network is trained online through continuous interaction with the city's traffic digital twin. After each control cycle, an experience replay sample is collected to update the network parameters. The experience replay pool has a capacity of 10,000 samples, employing a priority experience replay mechanism that prioritizes samples with high TD errors. At the beginning of each control cycle, the policy network outputs the optimal action based on the current state. This optimal action is parsed into a specific green light ratio allocation scheme and a phase switching timetable. The control cycle length is 90 seconds, and the green light ratio allocation scheme ensures that the sum of the green light times for each phase does not exceed the cycle length minus the sum of the yellow and all-red light times.

[0035] In step S7, the control commands output by the signal control strategy generation module are sent to the physical world signal controller for execution. Simultaneously, the actual traffic response data after execution is transmitted back to the urban traffic digital twin for calibrating simulation model parameters and initiating the next round of collaborative optimization. Specifically, after the signal control commands are executed, actual vehicle passage data at each intersection is collected, and the measured queue dissipation time, saturation flow rate, and travel time are calculated. The queue dissipation time is defined as the time from the green light turning on until all queued vehicles have passed the stop line. The saturation flow rate is calculated based on the number of vehicles passing within a unit of green light time, and the travel time is determined by the time difference between upstream and downstream detectors.

[0036] These measured indicators are compared with the simulation output of the urban traffic digital twin during the same time period to calculate the error vector. The error vector includes three components: queue length error, travel time error, and saturation flow rate error, with each component measured using the root mean square error. Based on the error vector, the gradient descent method is used to adjust the car-following model parameters, lane-changing intention threshold, and driver reaction time constant in the micro-simulation kernel, so that the simulation output gradually approximates the real response in the physical world. The parameter update step size is set to 0.01, and each calibration iteration does not exceed 10 times. After the parameter calibration is completed, the system automatically enters the next collaborative optimization cycle, forming a complete closed loop of "perception-simulation-decision-execution-learning".

[0037] The digital twin-based urban traffic system simulation system includes a static road network modeling unit, a multi-source data fusion unit, a digital twin engine unit, a multi-agent path planning unit, a signal coordination control unit, and a closed-loop feedback calibration unit. The static road network modeling unit constructs a static topological model of the urban road network, internally storing an interface for the urban basic geographic information system, a channelization design drawing parsing module, and a signal phase configuration database. The multi-source data fusion unit accesses real-time traffic flow data sources and includes a geomagnetic coil data receiver, a video analysis server interface, a floating car trajectory receiving module, and a Kalman filter fusion engine. The digital twin engine unit initializes and continuously updates the dynamically evolving urban traffic digital twin, employing a microscopic traffic simulation kernel based on cellular automata, and equipped with a vehicle agent manager, a state synchronization module, and a snapshot generator. The multi-agent path planning unit allocates and replans paths for virtual travel agents within the urban traffic digital twin, and includes an improved Dijkstra algorithm executor, a path cost calculator, and a traffic allocation matrix generator.

[0038] The signal coordination control unit generates signal control strategies and incorporates a deep Q-network strategy model, a Markov decision process state encoder, and a control command parser. The closed-loop feedback calibration unit executes control commands and transmits data back; it includes a signal controller communication interface, a measured performance calculation module, and a simulation parameter calibrator. All units are connected via a high-speed internal bus, and data exchange uses a unified message format to ensure the real-time performance and consistency of the overall system operation.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A simulation method for urban transportation systems based on digital twins, characterized in that, include: Construct a static topology model of the urban road network; Access real-time traffic flow data sources; Based on the static topology model and the real-time traffic flow data source, an dynamically evolving urban traffic digital twin is initialized and continuously updated. The urban traffic digital twin synchronizes the traffic status of the physical world once per second and uses an embedded microscopic traffic simulation engine to extrapolate and complete the traffic flow in unobserved areas. A multi-agent path planning module is deployed within the urban transportation digital twin. This module assigns an initial path to each virtual travel agent entering the road network and dynamically replans the path based on real-time travel time, queue length, and signal waiting time within the urban transportation digital twin. The next-time road network traffic distribution prediction result output by the multi-agent path planning module is transmitted to the signal control strategy generation module as a feedforward input signal. The signal control strategy generation module adopts a multi-intersection cooperative control algorithm based on reinforcement learning. The optimization objective is to minimize the total delay and total number of stops of vehicles in the entire road network. It combines the current signal phase status, the predicted value of the queue length of each approach lane and the prediction result of the feedforward traffic distribution to generate the green ratio and phase switching sequence of each intersection in the next control cycle. The control commands output by the signal control strategy generation module are sent to the signal controller in the physical world for execution. At the same time, the actual traffic response data after execution is sent back to the urban traffic digital twin for calibration of simulation model parameters and initiation of the next round of collaborative optimization cycle.

2. The urban transportation system simulation method based on digital twins according to claim 1, characterized in that, Constructing a static topology model of an urban road network specifically includes: Extract road centerline vector data from the city's basic geographic information system, and label road segments with attributes based on road grade, design speed, and number of lanes; Identify all at-grade intersections and determine the number of left-turn, straight-ahead, and right-turn lanes for each approach based on the channelization design drawings; Based on the traffic signal timing scheme document, establish a signal phase-lane mapping table for each intersection, and clarify the set of lanes served by each phase; All road segments, nodes, lanes, and phase information are integrated into a directed graph data structure. Nodes represent intersections, edges represent road segments, and the weight attributes of the edges include free-flow travel time, maximum capacity, and current saturation.

3. The urban transportation system simulation method based on digital twins according to claim 2, characterized in that, Access to real-time traffic flow data sources specifically includes: The system uses the vehicle detection interface built into the traffic signal controller to acquire the number of arriving vehicles, departing vehicles, and queue overflow status of each approach lane section every 5 seconds. By deploying video analysis servers at intersections, the surveillance video stream is processed frame by frame, and the instantaneous speed, vehicle type classification, and queue length estimate of each lane are output. Receive de-identified trajectory point sequences from partner ride-hailing platforms and logistics fleet management systems, and convert them into average travel speeds for road segments after spatiotemporal matching; By integrating real-time traffic flow data sources and using a Kalman filter to fuse different observations of the same physical quantity, a traffic state estimation vector with a unified spatiotemporal reference is generated.

4. The urban transportation system simulation method based on digital twins according to claim 3, characterized in that, Initializing and continuously updating the dynamically evolving urban transportation digital twin specifically includes: A microscopic traffic simulation kernel based on cellular automata is adopted, dividing each road segment into several cellular units with a length of 7.5 meters, each cell accommodating a standard passenger car; Each vehicle agent entering the simulation is assigned a unique identifier, origin, destination, desired speed, and car-following sensitivity parameters. At the beginning of each simulation step, the observed vehicle positions and speeds are forcibly synchronized based on the received real-time traffic flow data, while the state of unobserved areas is deduced based on the car-following model and the lane-changing model. The simulation kernel runs with a basic time step of 100 milliseconds and outputs a traffic state snapshot synchronized with the physical world every 10 time steps for use by the upper-level decision-making module.

5. The urban transportation system simulation method based on digital twins according to claim 4, characterized in that, The multi-agent path planning module assigns an initial path to each virtual travel agent entering the road network, and dynamically replans the path based on the real-time travel time, queue length, and signal waiting time within the urban traffic digital twin. Specifically, this includes: When a new travel request is injected into the city's traffic digital twin, the improved Dijkstra algorithm is invoked to calculate multiple alternative routes, with its origin and destination as endpoints. The path cost function of the improved Dijkstra algorithm comprehensively considers the free-flow travel time of the road segment, the current saturation-converted delay, and the expected red light waiting time at downstream intersections. During the vehicle's operation, at each intersection, the route selection for the remaining journey is reassessed. If a new route has an estimated total journey time that is more than 15% shorter than the current route, a route switch is triggered. All route replanning decisions are made within the digital twin, and its output, the traffic allocation matrix for each route segment in the next 15 minutes, is encapsulated as a feedforward signal.

6. The urban transportation system simulation method based on digital twins according to claim 5, characterized in that, The signal control strategy generation module employs a multi-intersection cooperative control algorithm based on reinforcement learning, with the optimization objective of minimizing the total vehicle delay and total number of stops across the entire road network. Combining the current signal phase state, predicted queue lengths for each approach lane, and predicted feedforward traffic distribution, it generates the green light ratio and phase switching sequence for each intersection in the next control cycle. Specifically, this includes: The entire control area is modeled as a Markov decision process, whose state space consists of the queue length of each approach lane at each intersection, the remaining time of the current phase, and the feedforward flow prediction vector, and whose action space is the phase switching instruction set of each intersection. A deep Q-network is used as the policy network. Its input is a state space vector and its output is the expected cumulative reward value of each action. The reward function is defined as the negative weighted sum of the total delay of vehicles and the total number of stops in the entire road network. The policy network is trained online through continuous interaction with the city's traffic digital twin. An experience replay sample is collected after each control cycle to update the network parameters. At the beginning of each control cycle, the policy network outputs the optimal action based on the current state. This optimal action is parsed into a specific green ratio allocation scheme and a phase switching timetable.

7. The urban transportation system simulation method based on digital twins according to claim 6, characterized in that, The actual traffic response data after execution is fed back to the urban traffic digital twin for calibration of simulation model parameters and initiation of the next round of collaborative optimization loop, specifically including: After the signal control command is executed, the actual vehicle passage data of each intersection is collected, and the measured queue dissipation time, saturation flow rate and travel time are calculated. These measured indicators are compared with the simulation output of the urban traffic digital twin during the same period, and the error vector is calculated. Based on the error vector, the gradient descent method is used to adjust the car-following model parameters, lane-changing intention threshold, and driver reaction time constant in the micro-simulation kernel, so that the simulation output gradually approximates the real response in the physical world. After parameter calibration is completed, the system automatically enters the next collaborative optimization cycle.

8. The urban transportation system simulation method based on digital twins according to claim 1, characterized in that, The microscopic traffic simulation engine uses an improved intelligent driver model as the car-following model, and its acceleration... The calculation formula is: ; This is the maximum acceleration; The vehicle speed at the current moment; For the desired speed; The acceleration index; This represents the current distance between the front of the vehicle; It is the relative velocity; This represents the desired front-end spacing.

9. The urban transportation system simulation method based on digital twins according to claim 5, characterized in that, The path cost function Defined as: ； For path, For road segments on the path, For free-flow travel time on the road segment, To delay waiting in line, Signal waiting time and These are the weighting coefficients.

10. The urban transportation system simulation method based on digital twins according to claim 6, characterized in that, The reward function The expression is: ； Total vehicle delays across the entire road network, expressed in seconds; This represents the total number of stops for vehicles across the entire road network. and These are the corresponding weighting coefficients.

Images