Intersection signal timing dynamic cooperation method and system based on real-time traffic prediction

Abstract

The invention belongs to the technical field of intelligent traffic systems, and particularly relates to an intersection signal timing dynamic coordination system and method based on real-time traffic prediction.The method comprises the steps of collecting multi-source traffic data, generating a space-time prediction digital twin model, dynamically defining an intersection coordination cluster and executing cluster coordination optimization control. And signal timing and closed loop feedback are carried out. By adopting the technical scheme, the cooperative operation efficiency and the intelligent management level of the urban intersection group can be effectively improved, and the fundamental conversion from local and reactive active cooperative control to global and predictive active cooperative control is realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation systems technology, specifically a method and system for dynamic coordination of intersection signal timing based on real-time traffic prediction. Background Technology

[0002] In modern urban traffic management systems, intersections, as key nodes in the road network, directly determine the service level and traffic capacity of the entire network through their operational efficiency. Therefore, developing efficient and intelligent intersection signal control technology is of paramount practical significance and application value for alleviating traffic congestion, reducing travel delays, and improving traffic safety. Traffic signal control technology has evolved from fixed timing and sensor-based control to adaptive control. With the deep integration of vehicle-to-everything (V2X) communication technology and artificial intelligence algorithms, dynamic collaborative control based on real-time traffic condition perception and future traffic flow prediction has become a current research hotspot and cutting-edge development in this field.

[0003] Against this backdrop, those skilled in the art have conducted numerous beneficial explorations. For example, Chinese patent CN111710176B discloses a signal-vehicle trajectory cooperative control method for intersections in a vehicle-road cooperative environment. The core idea of ​​this method is to construct mutually independent intersection signal optimization models and vehicle trajectory optimization models. By solving these models separately, timing schemes are provided for traffic lights, and the travel trajectories of connected vehicles through the intersection are planned. At a specific stage of technological development, this scheme effectively solves the problem of vehicle throughput efficiency within a single intersection by decoupling and optimizing macroscopic signal control with microscopic vehicle behavior. To a certain extent, it reduces the number of starts and stops and delays of vehicles at approach lanes and within the intersection, constituting a significant improvement over traditional signal control methods. Another representative technology, such as the dynamic adjustment method for the signal cycle of staggered intersections disclosed in Chinese patent CN117334061B, focuses on responding to real-time traffic information. This method dynamically adjusts the signal cycle by collecting vehicle information from each lane of the intersection in real time and utilizing a signal cycle optimization model and an optimal signal timing scheme model. The contribution of this scheme is that it enables the signal control system to break away from rigid, fixed-cycle patterns and make adaptive adjustments based on the actual traffic flow detected at present, thereby improving the matching degree of the signal timing scheme to real-time traffic demand.

[0004] However, with the increasing complexity of urban transportation systems and the ever-increasing demands for refined management, the inherent characteristics of the aforementioned technical solutions at the principle level have gradually revealed their limitations in dealing with large-scale, highly dynamic modern traffic flows. These limitations do not stem from oversights in model construction, but are rooted in a deeper technical contradiction: the inherent conflict between local spatial optimization and temporal response lag. Specifically, existing technical paradigms generally simplify complex traffic network problems into a series of local, instantaneous optimization problems, while neglecting the indivisible spatiotemporal continuity of traffic flow itself.

[0005] The reason for this lies in the fact that urban traffic flow is essentially a dynamic system that is spatially interconnected and continuously evolving in time. Outflow traffic at one intersection inevitably constitutes inflow traffic at downstream intersections; and the current traffic state inevitably forms the basis for the evolution of traffic patterns in the near future. While the first technical solution (CN111710176B) achieves signal and vehicle coordination, its optimization spatial scale is strictly limited to a single intersection and its approach lanes, creating a kind of "spatial island" optimization. This optimization method may achieve local optima, but its decision-making process does not consider the potential chain reaction of this optimization behavior on adjacent intersections. For example, if an intersection rapidly releases a large volume of traffic to maximize its own traffic efficiency, it may cause traffic saturation at downstream intersections in a short period, forming new congestion points and thus reducing the overall operational efficiency of the road network. Correspondingly, while the second technical solution (CN117334061B) introduces a dynamic adjustment mechanism, its decision-making is mainly based on historical data and currently existing queue lengths and other "already occurring" state information. This reactive approach to traffic management, based on the current state, inherently suffers from a time lag. By the time the system detects congestion and begins adjusting signal timing, congestion has already formed, and the optimal intervention window has often been missed. Especially during rush hours or in the event of unforeseen circumstances, traffic flow changes rapidly, and this lag means that control strategies will always fall behind actual needs, making it difficult to achieve truly proactive and forward-looking traffic management.

[0006] Therefore, the deep dilemma of existing technologies lies in the fact that their control logic is based on the separation of the spatiotemporal correlation of traffic systems. On the one hand, local optimization in space ignores the transmission and diffusion effects of traffic flow in the road network; on the other hand, passive response in time lacks the ability to predict the development trend of traffic flow. These two limitations are superimposed on each other, which together lead to the core problems of insufficient adaptability, weak coordination and low overall efficiency of existing signal timing systems when facing complex traffic scenarios.

[0007] To this end, the present invention provides a method and system for dynamic coordination of intersection signal timing based on real-time traffic prediction. Summary of the Invention

[0008] The technical problem this invention aims to solve is to overcome the inherent conflict between local spatial optimization and temporal response lag in existing intersection signal control methods. Existing technological paradigms either limit optimization to a single intersection, neglecting the transmission effect of traffic flow throughout the road network, leading to "spatial island" optimization that may cause downstream congestion; or they rely on reactive control based on current traffic conditions, which, due to its inherent time lag, cannot proactively intervene in the dynamic evolution of traffic flow. This spatiotemporally fragmented control logic limits the adaptability, coordination, and overall efficiency of signal timing systems in complex traffic scenarios.

[0009] To achieve the aforementioned objectives, this invention provides a dynamic coordination method and system for intersection signal timing based on real-time traffic prediction. The core technical concept of this invention lies in constructing a digital twin system capable of accurately mapping and proactively predicting the spatiotemporal evolution of road network traffic flow. Based on the system's prediction of future traffic conditions, it dynamically identifies and divides interconnected intersection clusters, and then performs coordinated signal timing optimization on the entire cluster with the goal of maximizing network macro-efficiency. Finally, through a closed-loop feedback mechanism, it continuously corrects the prediction model and control strategy, thereby achieving a fundamental shift from "local, reactive" control to "global, predictive" proactive coordinated control.

[0010] The technical solution adopted by this invention to solve its technical problem is: the intersection signal timing dynamic coordination system based on real-time traffic prediction described in this invention, the structure of which includes: a multi-source heterogeneous data acquisition layer, a spatiotemporal prediction digital twin generation module, a dynamic coordination intersection cluster definition module, a cluster coordination optimization control module, and a signal timing execution and closed-loop feedback module.

[0011] Furthermore, the multi-source heterogeneous data acquisition layer is configured to comprehensively and in real-time acquire the basic data required for constructing a dynamic evolution model of traffic flow within the physical space of the traffic network. This data acquisition layer specifically includes an edge perception unit deployed at each controlled intersection, a vehicle-road cooperative communication unit, and a central cloud platform data interface. The edge perception unit consists of at least one millimeter-wave radar sensor and a high-resolution optical imaging sensor. The millimeter-wave radar sensor operates at a frequency of 77 GHz, has a horizontal field of view of no less than 120 degrees, and a maximum detection range of no less than 250 meters. It has the capability to synchronously and with high precision (distance accuracy better than 0.1 meters, speed accuracy better than 0.1 m / s) detect the distance, speed, and azimuth of a single vehicle target, thereby acquiring the precise kinematic parameters of all motor vehicles and non-motor vehicles at the intersection's approach lanes, exit lanes, and within the intersection. The high-resolution optical imaging sensor has an optical resolution of 4K (3840×2160 pixels), a frame rate of no less than 30 fps, and is equipped with built-in... A deep convolutional neural network-based visual recognition processor is used to identify vehicle types (e.g., passenger cars, trucks, buses), lane occupancy, queue length, and the presence and intentions of pedestrians and non-motorized vehicles in real time. The vehicle-road cooperative communication unit adopts the C-V2X communication technology standard and communicates directly with connected vehicles equipped with OBUs (On-Board Units) through RSUs (Roadside Units) deployed on the roadside. It receives and parses the Basic Safety Messages (BSMs) broadcast by connected vehicles. The data structure of the Basic Safety Messages includes a unique temporary identifier for the vehicle, three-dimensional spatial coordinates (longitude, latitude, and elevation), speed, heading angle, acceleration, and vehicle size. The central cloud platform data interface is used to access the historical traffic database, real-time weather information system, and public event release system of the urban traffic management center to obtain background information that has a macroscopic impact on traffic flow. All data collected through the above units are appended with precise timestamps, with timestamp accuracy reaching the millisecond level.

[0012] Furthermore, the spatiotemporal prediction digital twin generation module communicates with the multi-source heterogeneous data acquisition layer. Its core function is to fuse real-time multi-source data to construct a virtual digital model that is highly synchronized with the physical road network state and can perform high-fidelity projections into the future. This module integrates three sub-modules: a multi-source data fusion and state estimation sub-module, a hybrid granularity traffic flow propagation modeling sub-module, and a future state sequence generation sub-module. The multi-source data fusion and state estimation sub-module receives heterogeneous data streams from the edge sensing unit, the vehicle-road cooperative communication unit, and the central cloud platform data interface. This sub-module employs a... This fusion algorithm, based on Extended Kalman Filter (EKF), correlates and fuses target trajectories detected by millimeter-wave radar, vehicle attributes and classification information identified by optical sensors, and BSM data reported by connected vehicles. This generates a unified, continuous, and accurate state vector for each traffic participant in the road network (including connected vehicles, non-connected vehicles, non-motorized vehicles, and pedestrians). This state vector includes information such as its position, velocity, acceleration, type, and lane affiliation in the global coordinate system. This fusion process resolves the limitations of single-sensor perspectives and potential conflicts between data sources, generating a snapshot of the global traffic state at the current moment (T=0).

[0013] Furthermore, the hybrid granularity traffic flow propagation modeling submodule, using the aforementioned global traffic state snapshot as initial conditions, predicts the traffic flow evolution over a future period (e.g., the next 300 seconds). This submodule employs a hybrid modeling strategy to balance computational efficiency and prediction accuracy. For structured road segments such as arterial roads and expressways in the road network, a macroscopic model based on the Cell Transmission Model (CTM) is used. Specifically, the road segment is divided into several cells of length L, and the propagation of traffic waves is simulated by solving a set of difference equations. Its core state update equation is: Where ρ j (t) represents the density of cell j at time t, q j(t) represents the outflow of cell j at time t, which is a function of the upstream and downstream cell densities. For the intersection area and its adjacent complex weaving area, an agent-based micro-simulation model is adopted. Each traffic participant is modeled as an independent agent, whose behavior is driven by a set of refined driving behavior models (such as car-following model and lane-changing model) and strictly follows the constraints of traffic lights. The macro model and the micro model exchange data and couple states at the boundary of the area to ensure the continuity of the entire road network prediction. The key parameters of the model, such as the maximum flow, critical density, and congestion density in CTM, as well as the driver reaction time and expected headway in the micro model, are continuously optimized by an online parameter self-calibration mechanism using the residual between the measured data and the predicted data.

[0014] Furthermore, the future state sequence generation submodule formats and outputs the calculation results of the hybrid granularity traffic flow propagation modeling submodule as a three-dimensional data tensor P(i,j,k), where the first dimension i represents the future discrete time step (e.g., from t+1 seconds to t+300 seconds), the second dimension j represents the smallest spatial unit defined in the road network (e.g., lane segment or CTM cell), and the third dimension k represents the set of predicted traffic state parameters for that spatial unit at that future time. This set of parameters specifically includes: predicted average speed, predicted traffic density, predicted flow rate, predicted queue length, predicted saturation, etc. This data tensor constitutes the sole data basis for subsequent optimization decisions.

[0015] Furthermore, the dynamic cooperative intersection cluster delineation module is connected to the spatiotemporal prediction digital twin generation module. Its function is to dynamically and adaptively identify a set of intersections with strong correlations in traffic behavior based on the prediction of future traffic flow. This module first uses the prediction data output by the future state sequence generation submodule to calculate the intersection cluster definition for any two adjacent or nearby intersections (I...) in the road network. a ,I b The traffic association strength coefficient C(a,b) between Q and Q is calculated using the formula: C(a,b) = f(Q). out,a→b (t),T travel,a→b (t)), where Q out,a→b (t) is the predicted traffic flow function from intersection a to intersection b, where T is the flow rate. travel,a→b(t) is the predicted travel time function of the traffic flow between the two points. When it is predicted that the traffic flow from point a to point b will be large and the travel time will be short in a future control period, the value of C(a,b) will be high. Subsequently, the module models the entire controlled road network as an undirected weighted graph G = (V,E,W), where the node set V represents all intersections, the edge set E represents the connection relationship between adjacent intersections, and the weight set W is the traffic association strength coefficient C just calculated. Finally, the module executes a community detection algorithm on the graph, specifically the Louvain algorithm. This algorithm divides the entire graph into several subgraphs by iteratively optimizing the modularity index. Each subgraph constitutes a dynamic cooperative intersection cluster.

[0016] (Dynamic Collaborative Intersection Cluster, DCIC) This partitioning method ensures that the intersections within a cluster have the most frequent traffic interactions in the future, while the traffic impact between clusters is relatively weak. This provides a reasonable computational boundary for subsequent distributed collaborative optimization. The cluster partitioning result is dynamic. For example, during the morning rush hour, multiple intersections along the main road direction will be divided into a large linear cluster; while during the off-peak period, multiple small local clusters centered on commercial areas may be formed.

[0017] Furthermore, the cluster collaborative optimization control module is the core of the decision-making process of this invention. After the dynamic collaborative intersection cluster delineation module completes the cluster division, this module independently performs a collaborative signal timing optimization calculation for each defined DCIC. This module contains a multi-objective optimization function construction submodule, a hard constraint definition submodule, and a distributed collaborative solution algorithm submodule.

[0018] Furthermore, the multi-objective optimization function construction submodule aims to maximize the macroscopic operational efficiency of the entire intersection cluster within the predicted future time domain. Its constructed optimization objective function J is a weighted sum form:

[0019] minJ=w1·D cluster +w2·S cluster +w3·U cluster , where D cluster This is the total delay time of all vehicles within the cluster in the prediction time domain. This value is calculated by integrating the vehicle trajectories output by the spatiotemporal prediction digital twin model. cluster This is the total number of stops for all vehicles within the cluster in the prediction time domain. This value is calculated by detecting the number of events in the predicted trajectory where the speed drops to zero. cluster It is a network equilibrium penalty term, defined as the variance of the average saturation of all inlet channels within the cluster in the prediction time domain, Ucluster =Var(Sat i ), of which Sat i is the predicted saturation of the i-th approach lane. This penalty term is used to avoid some approach lanes being oversaturated or undersaturated for a long time due to local over-optimization, thereby improving the fairness and stability of the entire cluster's right-of-way allocation. The weight coefficients w1, w2, and w3 are determined by advanced traffic strategies. For example, during peak hours, the weight of w1 can be increased to prioritize traffic efficiency; at night, the weight of w2 can be increased to reduce unnecessary starts and stops, improve the driving experience, and reduce energy consumption.

[0020] Furthermore, the hard constraint definition submodule sets physical and safety boundary conditions that must be followed for the optimization process. These constraints include: minimum green light time for each phase (e.g., not less than 7 seconds to ensure pedestrian safety), maximum green light time (e.g., not more than 120 seconds to avoid excessive waiting), yellow light time (fixed at 3 seconds), full red clear time (calculated based on the intersection geometry), and a fixed signal phase sequence (e.g., for a standard cross intersection, a four-phase or eight-phase structure of "straight ahead - left turn ahead"). All optimization results must strictly meet these constraints.

[0021] Furthermore, the distributed collaborative solution algorithm submodule is responsible for solving the aforementioned constrained multi-objective optimization problem. Considering the complexity and high dimensionality of the problem (a cluster may contain multiple intersections, each with multiple phases, requiring optimization over multiple future cycles), this invention employs a distributed collaborative evolutionary algorithm based on a Model Predictive Control (MPC) framework. The specific process is as follows: First, a central coordinator (logically existing, but physically deployed in the cloud or at an edge node) sets weight coefficients according to a high-level strategy and decomposes the overall optimization objective J of the cluster into each intersection controller within the cluster. Each intersection controller, based on future traffic flow prediction data about itself and adjacent intersections obtained from a spatiotemporal prediction digital twin module (as boundary conditions of its model), utilizes an improved genetic algorithm. The Genetic Algorithm (GA) generates a set of candidate signal timing schemes locally (i.e., a green ratio sequence for the next N cycles). Then, each intersection controller broadcasts its optimal candidate scheme and its expected impact on the boundary traffic of adjacent intersections to other controllers in the cluster. After receiving the information from its neighbors, each controller updates its boundary conditions and runs its local genetic algorithm again for the next round of iterative optimization. This process is repeated several times until the timing schemes of all controllers in the cluster converge or reach the preset computation time limit. The converged timing scheme is the optimal collaborative timing scheme of the DCIC in the next control cycle. This distributed architecture decomposes the large-scale centralized optimization problem into multiple parallel, low-dimensional local optimization problems, and achieves global approximate optimality through iterative coordination, which significantly improves the solution efficiency.

[0022] Furthermore, the signal timing execution and closed-loop feedback module serves as a bridge connecting virtual optimization and the physical world. This module receives the optimal coordinated timing scheme output from the cluster coordinated optimization control module. This scheme is a time series accurate to the second, specifying the start and end times of each phase at each intersection within the cluster within a future execution cycle (e.g., 5-10 minutes). The module transforms this high-level logic scheme into low-level control commands conforming to a specific brand's signal controller communication protocol (such as NTCIP), and distributes them to the physical traffic signal controller at the intersection via industrial Ethernet or dedicated fiber optic cable. The controller then drives the switching of red, yellow, and green lights. Simultaneously, the core function of this module is to achieve closed-loop feedback of the control system. It integrates the actually executed signal timing scheme with data obtained from the multi-source heterogeneous data acquisition layer... The actual traffic flow data generated under this timing scheme (i.e., actual queue length, delay, vehicle speed, etc.) is fed back to the spatiotemporal prediction digital twin generation module. This fed-back data is used as an important input for the next round of prediction and optimization of the system: on the one hand, it is used to calculate the error between the output of the prediction model and the actual observation value, and based on this error, the parameters in the mixed-granularity traffic flow propagation modeling submodule are calibrated and finely adjusted online through methods such as backpropagation or gradient descent; on the other hand, this error is also used to evaluate the actual effect of the previous cycle's collaborative optimization control strategy, providing data support for dynamically adjusting the weight coefficients in the optimization objective function. Through this continuous "prediction-control-observation-calibration" closed loop, the system proposed in this invention can continuously learn and adapt to the dynamic changes in the traffic environment, and continuously improve its prediction accuracy and control performance.

[0023] The dynamic coordination method for intersection signal timing based on real-time traffic prediction provided by this invention has specific implementation steps that correspond one-to-one with the functions of each module in the above system, including the following steps:

[0024] S1: Real-time collection of multi-source heterogeneous traffic data. Through millimeter-wave radar, high-resolution cameras, vehicle-road cooperative RSUs and central cloud platform interfaces deployed in the road network, continuously collect vehicle kinematic parameters, vehicle attributes, queuing information, connected vehicle BSM data and macro background information covering the entire area with timestamps.

[0025] S2: Generate a spatiotemporal prediction digital twin model. Input the collected multi-source heterogeneous data into the multi-source data fusion and state estimation submodule. Use the extended Kalman filter algorithm to generate a global traffic state snapshot at the current moment. Use this snapshot as the initial condition to drive the hybrid granularity traffic flow propagation modeling submodule to perform high-fidelity simulation prediction of the road network traffic flow evolution within a future preset time domain (e.g., 300 seconds). Finally, generate a future state sequence data tensor containing the predicted traffic state parameters of each road segment at future moments.

[0026] S3: Dynamically define the intersection cooperative clusters. Based on the future state sequence data generated in step two, calculate the traffic association strength coefficient between any two intersections in the road network, abstract the road network into a weighted graph, and run the Louvain community detection algorithm on this graph. According to the inherent structure of traffic association strength, dynamically divide the road network into several independent, highly interconnected dynamic cooperative intersection clusters (DCIC).

[0027] S4: Execute cluster-level collaborative optimization control. For each defined DCIC, construct a weighted multi-objective optimization function with the total cluster delay, total number of stops, and network balance as core indicators. Under the condition of satisfying hard constraints such as minimum green light time and phase sequence, adopt a distributed collaborative evolutionary algorithm based on model predictive control framework to jointly solve a set of collaborative signal timing schemes that can achieve the best overall benefits in future control cycles for all intersections in the cluster.

[0028] S5: Execute signal timing and implement closed-loop feedback. The optimal coordinated timing scheme generated in step four is transformed into low-level control commands and sent to the physical signal controllers at each intersection for execution. At the same time, the actual timing scheme is compared with the real traffic flow data collected in the new round, the prediction error is calculated, and this error is fed back to the spatiotemporal prediction digital twin generation module for online calibration of model parameters, and also fed back to the cluster collaborative optimization control module for evaluation and adjustment of optimization strategies.

[0029] The beneficial effects of this invention are as follows:

[0030] 1. The intersection signal timing dynamic coordination method and system based on real-time traffic prediction described in this invention uses a digital twin system to proactively predict the spatiotemporal evolution of traffic flow, combined with dynamic cluster collaborative optimization, to effectively solve the "spatial island" problem and time lag caused by local optimization in existing technologies, avoid downstream congestion, and achieve proactive intervention in traffic dynamics.

[0031] 2. The method and system for dynamic coordination of intersection signal timing based on real-time traffic prediction described in this invention dynamically defines intersection clusters based on traffic association strength, ensuring close traffic interaction within the clusters, achieving coordinated timing with the goal of optimizing the macro-level benefits of the network, enhancing the coordination and integrity of signal control in complex road networks, and reducing the disorderly transmission of cross-regional traffic flow.

[0032] 3. The intersection signal timing dynamic coordination method and system based on real-time traffic prediction described in this invention constructs a multi-objective function that includes total delay, total number of stops, and network balance. In the optimization process, it takes into account both efficiency and fairness, which can reduce vehicle delay and unnecessary starts and stops (reduce energy consumption and emissions), and avoid local oversaturation or undersaturation, thereby improving the stability of right-of-way allocation.

[0033] 4. The intersection signal timing dynamic coordination method and system based on real-time traffic prediction described in this invention continuously calibrates the prediction model parameters and control strategies by comparing the predicted data with the actual traffic conditions in real time, enabling the system to dynamically adapt to random fluctuations in traffic flow (such as sudden accidents and weather changes) and maintain high-precision prediction and efficient control performance over the long term.

[0034] 5. The intersection signal timing dynamic coordination method and system based on real-time traffic prediction described in this invention integrates data from millimeter-wave radar, optical imaging, vehicle-road coordination, and cloud platform, and generates a global traffic state snapshot through high-precision fusion. This provides a comprehensive and accurate data foundation for prediction and optimization, effectively avoids the limitations and errors of a single data source, and improves the scientific nature of signal timing decisions. Attached Figure Description

[0035] The invention will now be further described with reference to the accompanying drawings.

[0036] Figure 1 This is a system structure block diagram of the present invention;

[0037] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0038] Figure 3 This is a schematic diagram illustrating the principle of closed-loop control of the physical road network and digital twin system in this invention;

[0039] In the diagram: 100, Multi-source heterogeneous data acquisition layer; 110, Edge sensing unit; 120, Vehicle-road cooperative communication unit; 130, Central cloud platform data interface; 200, Spatiotemporal prediction digital twin generation module; 210, Multi-source data fusion and state estimation sub-module; 220, Hybrid granularity traffic flow propagation modeling sub-module; 230, Future state sequence generation sub-module; 300, Dynamic cooperative intersection cluster definition module; 400, Cluster cooperative optimization control module; 410, Multi-objective optimization function construction sub-module; 420, Hard constraint definition sub-module; 430, Distributed cooperative solution algorithm sub-module; 500, Signal timing execution and closed-loop feedback module. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, will illustrate the dynamic coordination method and system for intersection signal timing based on real-time traffic prediction provided by this invention. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

[0041] like Figure 1 and Figure 3 As shown in the embodiment of the present invention, the intersection signal timing dynamic coordination system based on real-time traffic prediction is logically divided into five interrelated functional modules, which together constitute a complete technical link from data perception to control execution and then to closed-loop feedback.

[0042] First, the technical solution of this invention is rooted in a comprehensive data foundation, which is constructed by a multi-source heterogeneous data acquisition layer 100. This data acquisition layer 100 is deployed in a distributed manner within the physical traffic network. Its core mission is to capture, in real time and without blind spots, all-dimensional data required to construct a dynamic evolution model of traffic flow. Specifically, this layer includes an edge perception unit 110 deployed at each controlled intersection, a vehicle-road cooperative communication unit 120, and a central cloud platform data interface 130. In a specific implementation scenario, at least one set of edge perception units 110 is installed on a high pole approximately 100 to 150 meters from the stop line in the direction of the approach lane at each intersection. Each unit 110 integrates a 77GHz millimeter-wave radar sensor and a 4K resolution optical imaging sensor. The millimeter-wave radar sensor uses modulation... The frequency continuous wave (FMCW) system uses a 4T transmitting antenna and a 16R receiving antenna to form a virtual MIMO antenna array, thereby achieving high-precision angular resolution. Its operating bandwidth is 4GHz, ensuring a range resolution better than 0.1 meters. In terms of signal processing, the radar's raw data is first processed by a two-dimensional fast Fourier transform (FFT) to generate a range-velocity spectrum. Then, the target azimuth angle is accurately estimated using either the multiple signal classification (MUSIC) algorithm or the rotationally invariant subspace (ESPRIT) algorithm. With this configuration, the radar can cover a horizontal field of view of no less than 120 degrees and a maximum detection range of no less than 250 meters. It can simultaneously track up to 256 motor vehicles and non-motor vehicles within the field of view, and output the velocity vector and position coordinates of each target in real time with an accuracy of 0.1 meters per second.

[0043] Alongside it is a high-resolution optical imaging sensor with a 1 / 1.8-inch CMOS sensor and an effective pixel count of 3840×2160. It is equipped with a motorized zoom lens with a focal length range of 5mm to 50mm, allowing for flexible adjustment of the field of view based on the installation height to fully cover all lanes of the intersection approach and the queuing area behind the stop line. This sensor acquires video streams at a rate of at least 30 frames per second and sends the video stream directly to the built-in, lightweight design.

[0044] YOLOv5 (YouOnlyLookOnce, version 5) is a deep convolutional neural network-based visual recognition processor. The model it runs is trained on a finely labeled dataset of over five million images covering various lighting, weather, and traffic congestion conditions. It can identify vehicle types (such as small passenger cars, medium-sized passenger cars, large trucks, buses, and motorcycles), lane occupancy status, real-time queue lengths (accurate to the number of vehicles) and the number and waiting status of pedestrians in crosswalk areas with over 98% accuracy in real time.

[0045] The vehicle-road cooperative communication unit 120 consists of a standard roadside unit (RSU) deployed at the intersection. This RSU conforms to the C-V2X communication technology standard and supports the LTE-V2X PC5 direct communication mode. It continuously listens for and decodes the basic safety messages (BSM) broadcast by connected vehicles equipped with on-board units (OBUs) in the vicinity. The parsing of BSM messages strictly follows the standard "Dedicated Short Range Communication for Cooperative Intelligent Transport Systems Part 3: Message Sets" (GB / T31024.3), extracting key dynamic information such as the vehicle's temporary identification ID (PTID), high-precision GNSS positioning information (longitude, latitude, and elevation), speed, heading angle, lateral and longitudinal acceleration, and vehicle dimensions (length and width). At the same time, this unit 120 is also responsible for receiving and parsing the signal phase and timing messages (SPaT) and map messages (MAP) broadcast by the intersection signal controller, providing decision support for connected vehicles, and can also serve as a data source for internal status verification of this system.

[0046] The central cloud platform data interface 130 connects to the backend server of the city traffic management center via a secure RESTful API. This interface periodically (e.g., every 5 minutes) pulls historical traffic flow data at the city level, 24-hour weather forecast information provided by the meteorological department (such as factors that may affect driving behavior, such as rain and fog), and public event information released by the emergency management department (such as road construction, traffic accidents, and large-scale events). All data collected through the above three sub-units are forcibly appended with a unified timestamp synchronized by the NTP (Network Time Protocol) service with millisecond-level accuracy when entering the system, providing a guarantee for subsequent spatiotemporal data alignment and fusion.

[0047] After acquiring massive amounts of real-time data, the system enters the core prediction phase, executed by the spatiotemporal prediction digital twin generation module 200, such as... Figure 1As shown, module 200 is tightly coupled with data acquisition layer 100. Its function is to transform discrete, heterogeneous observation data into a continuous, dynamic, and future-predictable virtual traffic world. This module is internally divided into three collaborative sub-modules. The first is the multi-source data fusion and state estimation sub-module 210, which receives all timestamped data streams from radar, cameras, V2X, and the cloud platform. The core of this sub-module is a track-level data fusion algorithm based on extended Kalman filtering (EKF). For each perceived and identified traffic participant in the road network, the system creates a state vector x = [p x ,p y ,p z ,v x ,v y ,v z ,a x ,a y ,a z ,w,l,type] T This includes the vehicle's three-dimensional spatial position, velocity, acceleration, width, length, and vehicle type determined by visual recognition. When a new sensor measurement (e.g., a point detected by radar, or BSM data reported by a connected vehicle) arrives, the algorithm first associates it with existing tracks through a gating mechanism. After successful association, the state vector of the corresponding track is updated using the measurement. The prediction step of EKF uses a constant acceleration motion model to predict the prior estimate of the state, while the update step selects different measurement functions h(x) according to the source of the measurement (radar, vision, or BSM), maps the state vector to the measurement space, and calculates the Kalman gain to achieve the posterior optimal estimate of the state vector. This process not only smooths the measurement noise of individual sensors, but more importantly, it resolves the conflict between data sources and generates a unified, high-precision dynamic state description for all traffic participants in the road network, whether connected or non-connected vehicles, forming a global traffic state snapshot at the current moment (T=0).

[0048] Furthermore, the hybrid granularity traffic flow propagation modeling submodule 220 uses this global state snapshot as initial conditions to simulate and predict the traffic flow evolution within a preset time domain, such as 300 seconds. This submodule creatively adopts a hybrid modeling strategy that combines macroscopic and microscopic models. For structured road segments with relatively uniform traffic flow characteristics, such as urban arterial roads and expressways, the system adopts a macroscopic model based on the Cell Transmission Model (CTM). In specific implementation, the road segment is discretized into a series of cells with a length L of 7.5 meters (approximately the length of a standard vehicle). Within each time step Δt (e.g., 1 second), the propagation of traffic waves is simulated by solving a set of coupled difference equations. The core state update equation is:

[0049] Where ρ j (t) represents the density of cell j at time t, q in,j (t) and q out,j (t) represents the flow rates in and out of cell j, respectively. The calculation of flow rates relies on a basic traffic flow graph describing the flow-density-velocity relationship, whose parameters, such as the maximum flow rate q, are... max Critical density ρ crit Congestion density ρ jam The accuracy is obtained through statistical calibration of historical measured data. For complex weaving areas such as intersections and their entrances / exits, the system switches to an agent-based micro-simulation model. In this model, each traffic participant (vehicle, pedestrian) is modeled as an agent with independent decision-making capabilities. The behavior of vehicle agents is driven by sophisticated driving behavior models. For example, their car-following behavior uses an Intelligent Driver Model (IDM), and their lane-changing decisions use a lane-changing model that minimizes overall braking excitation (MOBIL). The behavior of these agents is strictly influenced by their location. Constrained by the traffic light state (which evolves according to a baseline timing scheme during prediction), the macroscopic CTM model and the microscopic agent model are seamlessly coupled at the defined region boundary: the flow from the macroscopic region to the microscopic region determines the generation rate and initial state of vehicles within the microscopic simulation region; conversely, the statistical characteristics of vehicles flowing out of the microscopic region serve as the flow input to the downstream boundary of the macroscopic CTM region. This hybrid modeling method ensures high-fidelity simulation of intersection areas while leveraging the efficiency of the macroscopic model to handle large-scale road segments, achieving the best balance between computational efficiency and prediction accuracy.

[0050] Finally, the future state sequence generation submodule 230 is responsible for organizing and outputting the massive amounts of data generated by the complex simulation process into structured prediction results. Its output is a standardized three-dimensional data tensor P(i,j,k). In this tensor, the first dimension index i represents the future discrete time step from the current moment, with a step size of 1 second and a range from 1 to 300; the second dimension index j represents the unique identifier of the smallest pre-divided spatial unit in the road network. This unit can be a CTM cell or a specific segment of a lane (e.g., 20 meters in length); the third dimension index k... This data tensor P represents a detailed set of traffic state parameters for the spatial unit at the corresponding future time. Specifically, this set of parameters includes: predicted average speed (unit: km / h), predicted traffic density (unit: veh / km / lane), predicted flow rate (unit: veh / h), predicted queue length (unit: meters or number of vehicles), and predicted saturation (defined as the ratio of predicted flow rate to lane capacity). This data tensor P, in a highly condensed and structured form, depicts a complete picture of the traffic situation of the entire road network in the next five minutes, forming a solid and unique data foundation for all subsequent optimization decisions.

[0051] Based on accurate future predictions, the system then proceeds to the first step of the decision-making phase: the dynamic collaborative intersection cluster delineation module 300. This module is directly connected to the spatiotemporal prediction digital twin generation module 200. Its core task is to move away from the traditional, static sub-region division method based on geographical location, and instead dynamically and adaptively identify the set of intersections with strong correlations in traffic behavior based on the predicted traffic flow interaction intensity. This module first uses the predicted data tensor P output by the future state sequence generation sub-module 230 to quantify the traffic association strength between any pair of adjacent or near intersections (denoted as I_a and I_b) in the road network. This strength is represented by a comprehensive association coefficient C(a,b), the calculation formula of which has been carefully designed, for example:

[0052]

[0053] Among them, Q pred,a→b (t) represents the average predicted flow of the connecting road segment from intersection a to intersection b within the future predicted time domain t. link,a→b This refers to the designed traffic capacity of this road section; T pred,a→b (t is the average predicted travel time for this road segment, while T) ff,a→b It is its free-flow travel time, weight w q and w tThe formula can be adjusted according to the control strategy. The physical meaning of this formula is that when it is predicted that the traffic flow from a to b will be close to saturation in the future, and the travel time will increase significantly due to queuing, the value of C(a,b,t) will approach the maximum. This indicates that the two intersections are functionally closely coupled, and adjusting the signal control of one side will drastically affect the other side.

[0054] After calculating the association strength coefficients of all adjacent intersection pairs within the future control period, the module abstracts the entire controlled road network into an undirected weighted graph G = (V, E, W). In this graph, each node in the node set V represents an intersection, the edge set E represents the physical connecting roads between intersections, and the weight W of each edge is the calculated traffic association strength coefficient C. Finally, the module executes an efficient community detection algorithm on this weighted graph, specifically the Louvain algorithm. The Louvain algorithm is a heuristic algorithm based on modularity optimization, which uses a two-stage iterative process to... Discovering the community structure in the graph: In the first phase, each node is treated as an independent community. All nodes are traversed, and attempts are made to move each node to the community of its neighboring nodes, calculating the modularity gain from the move. Finally, the node is placed in the community that maximizes its gain. In the second phase, all communities discovered in the first phase are aggregated into a new supernode, constructing a new network. The process of the first phase is then repeated iteratively until moving any node in the network can no longer improve the overall modularity. After the algorithm converges, each community (or subgraph) is defined as a dynamic collaborative intersection cluster.

[0055] The advantage of Dynamic Collaborative Intersection Cluster (DCIC) is that it ensures that the traffic interactions between intersections within a cluster are the most frequent and close in the future, while the traffic impact between clusters is relatively weak. This provides a reasonable and efficient computational boundary for subsequent distributed optimization control. The cluster partitioning results evolve dynamically with changes in traffic demand. For example, during the morning rush hour on weekdays, multiple consecutive intersections along the main urban roads may be automatically identified and divided into a large linear cluster due to their strong tidal flow characteristics. During the off-peak hours at night, the road network may be decomposed into multiple small, isolated local clusters centered around commercial or residential areas.

[0056] Once the cluster is defined, the system assigns control tasks to the cluster collaborative optimization control module 400, which is the decision-making center of the system of this invention. For each DCIC defined by module 300, this module will independently and in parallel perform a collaborative signal timing optimization calculation to generate the optimal signal control scheme for the next control cycle (e.g., 5 to 10 minutes). This module contains three closely cooperating sub-modules.

[0057] First is the multi-objective optimization function construction submodule 410, whose objective is to maximize the macroscopic operational efficiency of the entire intersection cluster within the predicted future time domain. Its optimization objective function J is a configurable weighted sum form:

[0058] minJ=w1·D cluster +w2·S cluster +w3·U cluster Among them, the first target D cluster This is the total delay time of all vehicles within the cluster in the prediction time domain. This value is calculated by integrating the predicted trajectory of each vehicle agent output by the spatiotemporal prediction digital twin model, i.e., D. i =T actual,i -T ideal,i T actual,i T is the actual predicted travel time of vehicle i. ideal,i It is its ideal travel time without any traffic light delays, D cluster That is, the sum of D_i of all vehicles i, and the second objective S. cluster It is the total number of stops for all vehicles in the cluster within the prediction time domain. This value is obtained by detecting the predicted speed trajectories of all vehicle agents and counting the total number of events where their speed drops from a non-zero value to zero. The third objective U cluster It is a network equilibrium penalty term designed to improve the fairness of right-of-way allocation. Specifically, it is defined as the variance of the average saturation of all approach lanes within the cluster in the prediction time domain, i.e., U. cluster =Var(Sat i )=E[(Sat i -E[Sat i ]) 2 ], of which Sat iIt is the predicted saturation of the i-th approach lane calculated using the predicted data tensor P. This penalty term is used to avoid some approach lanes being oversaturated and congested for a long time due to excessive pursuit of efficiency, while others are idle, thereby improving the service level stability of the entire cluster and the sense of fairness for drivers. The weight coefficients w1, w2, w3 are not fixed, but can be dynamically set by advanced traffic management strategies. For example, when dealing with sudden congestion, the weight of w1 can be significantly increased to prioritize traffic diversion and ensure traffic efficiency; while under the strategy of prioritizing environmental protection, the weight of w2 can be increased to reduce unnecessary vehicle starts and stops, thereby reducing fuel consumption and carbon emissions.

[0059] Secondly, the hard constraint definition submodule 420 sets insurmountable physical and safety boundaries for the optimization process. These constraints ensure the feasibility and legality of the optimization results in the real world. Specific constraints include: the minimum green light time for each signal phase, for example, to ensure that pedestrians have sufficient time to safely cross the road, the minimum green light time for the corresponding phase of the pedestrian crossing must not be less than 7 seconds; the maximum green light time for each phase, for example, to avoid long waiting times in one direction causing driver anxiety or detours, the maximum green light time for any motor vehicle phase is limited to within 120 seconds; a fixed yellow light time, usually set to 3 seconds according to local traffic regulations; and a full red clearing time precisely calculated based on the intersection geometry (such as intersection width and vehicle clearing speed). In addition, there are fixed signal phase sequence constraints, for example, for a standard crossroads, a preset, conflict-free four-phase or eight-phase loop structure must be followed. All candidate timing schemes generated by the optimization algorithm must first pass the verification of this submodule, and any scheme that violates the constraints will be directly discarded.

[0060] Finally, the distributed collaborative solution algorithm submodule 430 is responsible for solving the constructed multi-objective optimization problem under the premise of satisfying the above hard constraints. Considering the nonlinear and high-dimensional characteristics of this problem (a cluster may contain dozens of intersections, each intersection has multiple phases, and the decision variable is the green light duration of each phase in multiple future control cycles, constituting a huge search space), this invention adopts a distributed collaborative evolutionary algorithm based on the model predictive control (MPC) framework. Its specific workflow is as follows: First, a logical central coordinator (which can be physically deployed in the cloud or acted as the node with the strongest edge computing capability in the cluster) determines the current high-level... The traffic strategy sets weight coefficients w1, w2, and w3 in the objective function J, and decomposes the overall optimization objective J of the cluster into each intersection controller (i.e., the solver). Each intersection solver, based on future traffic flow prediction data about itself and adjacent intersections obtained from the spatiotemporal prediction digital twin module 200 (this data constitutes the dynamic boundary conditions of its local optimization model), independently runs an improved genetic algorithm (GA) locally. In this GA, a "chromosome" is encoded as a sequence representing the green light ratio or green light duration for the intersection over N future signal cycles (e.g., N=5). The population selects and crosses over... The algorithm evolves through operations such as forking and mutation. Its fitness function represents the contribution of the timing scheme to the local objective function (total delay, number of stops, etc.). After the first round of local optimization, each intersection's solution agent broadcasts its currently found optimal candidate scheme, along with the expected impact of this scheme on the boundary traffic (exit traffic) of its downstream neighboring intersections, to all other agents in the cluster via the internal network. Upon receiving information from all neighbors, each agent updates the boundary conditions of its model (i.e., its upstream input traffic and downstream acceptance capacity) and, based on this, runs its local genetic algorithm again for the next round of iterative optimization. This "local optimization-information exchange" process... The "boundary update" process is repeated several times (e.g., 5 to 10 times) until the timing schemes of all controllers in the cluster converge to a stable Nash equilibrium state, or reach the preset computation time limit (e.g., 20 seconds). The converged timing scheme is then regarded as the optimal collaborative timing scheme for the DCIC in the next physical execution cycle. This distributed architecture cleverly decomposes a difficult large-scale centralized optimization problem into a set of parallel, low-dimensional local optimization problems. It approaches the global optimal solution through iterative coordination, thereby greatly improving the solution efficiency while ensuring the optimization effect and meeting the needs of real-time control.

[0061] Ultimately, the signal timing execution and closed-loop feedback module 500 acts as a bridge connecting the virtual optimization world and the physical traffic world. This module receives the optimal coordinated timing scheme output from the cluster coordinated optimization control module 400. This scheme is a command time sequence accurate to the second, which specifies in detail the precise opening and closing time of each signal phase at each intersection within the cluster in a future execution cycle (e.g., 5 minutes). The module first transforms this high-level logic scheme into low-level control commands that conform to the communication protocols of the specific brand of traffic signal controllers (such as Siemens, Hisense, etc.) at the intersection. For example, for controllers that support the NTCIP (National Transportation Communications for ITS Protocol) protocol, the module generates a standard NTCIP object setting request and sends the new phase duration, cycle duration, and offset parameters to the physical traffic signal controller at the intersection via industrial Ethernet or a dedicated fiber optic network. After receiving the command, the traffic signal controller precisely drives the on / off switching of its connected red, yellow, and green signal light groups, thereby implementing the optimization results.

[0062] However, the core function of this module lies in realizing the closed-loop feedback of the entire control system. While issuing control commands, it continuously sends two types of key information back to the upstream modules of the system. The first type of information is the record of the actual signal timing scheme, along with the traffic flow data actually generated under this timing scheme obtained from the multi-source heterogeneous data acquisition layer 100 (i.e., queue length, vehicle delay, and traffic speed actually measured by intersection sensors), which is sent back to the spatiotemporal prediction digital twin generation module 200. This data is used as the basis for the next round of prediction and optimization of the system. On the one hand, it is used to calculate the error between the output of the prediction model (such as the predicted queue length) and the actual observation value. This error signal is used to improve the key parameters (such as the CTM model) in the mixed-granularity traffic flow propagation modeling submodule 220 through online learning algorithms such as backpropagation or gradient descent. The system performs online calibration and fine-tuning of various parameters, including free-flow velocity, critical density, driver reaction time, and expected safe distance in the microscopic model. The second type of information is the error itself, which is also used to evaluate the actual effect of the previous cycle's collaborative optimization control strategy. If the actual delay reduction is far less than expected, it may indicate that the weight coefficients in the optimization objective function are set improperly. This evaluation result is fed back to the cluster collaborative optimization control module 400, providing key data support for manual intervention or automatic algorithm dynamic adjustment of weight coefficients w1, w2, and w3. Through this cyclical and uninterrupted closed-loop workflow of "prediction-control-observation-calibration," the system proposed in this invention can learn continuously from its interaction with the physical world, like an organic living organism, and continuously adapt to the unsteady dynamic changes in the traffic environment, thereby steadily improving its prediction accuracy and control performance.

[0063] like Figure 2 As shown, the intersection signal timing dynamic coordination method based on real-time traffic prediction provided by this invention closely follows the functional logic of each module of the system in its specific implementation steps. Its core process includes the following consecutive steps:

[0064] S1: Real-time collection of multi-source heterogeneous traffic data is carried out through edge perception units 110 (including millimeter-wave radar and high-resolution cameras), vehicle-road cooperative communication units 120 (RSU), and central cloud platform data interfaces 130 connected to the city traffic brain, which are deployed in the target road network area. The data is collected continuously and synchronously, covering the entire area and with millisecond-level precision timestamps, including vehicle kinematic parameters, vehicle attributes and classification information, lane queue length information, BSM data broadcast by connected vehicles, and macro background information (weather, events, etc.) affecting traffic flow.

[0065] S2: Generate a spatiotemporal prediction digital twin model. The massive, heterogeneous data stream collected in step one is input into the spatiotemporal prediction digital twin generation module 200 in real time. Inside this module, the multi-source data fusion and state estimation submodule 210 first uses the extended Kalman filter algorithm to fuse data from different sensors, generating a unified and accurate state vector for each traffic participant in the road network, thereby constructing a global traffic state snapshot at the current time T=0. Then, using this snapshot as the initial condition, the hybrid granularity traffic flow propagation modeling submodule 220 is driven to perform high-fidelity simulation prediction of the road network traffic flow evolution within a preset time domain (e.g., 300 seconds). Finally, the future state sequence generation submodule 230 outputs a future state sequence data tensor containing the predicted traffic state parameters of each road segment and each lane in every second of the future.

[0066] S3: Dynamically define the intersection coordination clusters. Input the future state sequence data generated in step two into the dynamic coordination intersection cluster definition module 300. Based on the predicted future traffic flow and travel time, this module calculates the traffic association strength coefficient between any two intersections in the road network. Then, it abstracts the entire road network into an undirected graph with intersections as nodes and association strength as weights. The Louvain community detection algorithm is run on this graph. Based on the inherent association structure of traffic flow in the prediction time domain, the entire road network is dynamically and adaptively divided into several independent dynamic coordination intersection clusters (DCICs) with highly associated internal members.

[0067] S4: Execute cluster-level collaborative optimization control. For each DCIC defined in step three, the cluster collaborative optimization control module 400 independently initiates an optimization calculation. First, a weighted multi-objective optimization function is constructed with minimizing the total delay of the cluster, the total number of stops, and maximizing network balance (i.e., minimizing the saturation variance) as the core indicators. Then, under the condition of strictly satisfying a series of hard safety and physical constraints such as minimum green light time, maximum green light time, and phase sequence, a distributed collaborative evolutionary algorithm based on the model predictive control (MPC) framework is adopted. Through iterative coordination between the solvers of each intersection, a collaborative signal timing scheme that can achieve the best overall operating efficiency for all intersections in the cluster is jointly solved.

[0068] S5: Execute signal timing and implement closed-loop feedback. The optimal coordinated timing scheme generated in step four is converted into low-level control commands conforming to the on-site signal controller protocol through the signal timing execution and closed-loop feedback module 500, and sent to the physical signals at each intersection for precise execution. At the same time, this module continuously compares the new round of real traffic flow data obtained from the data acquisition layer 100 with the predicted data of the previous cycle, calculates the error between the prediction model and the real world, and feeds this error information back to the spatiotemporal prediction digital twin generation module 200 for online calibration and optimization of the model's internal parameters. At the same time, the evaluation results of the control effect are also fed back to the cluster collaborative optimization control module 400 to guide and adjust subsequent optimization strategies.

[0069] By repeatedly executing the above five steps, the method of this invention can ensure that every decision in signal control is based on an accurate prediction of future traffic conditions and a deep understanding of the global correlation of the road network. This achieves a fundamental shift from traditional passive and isolated control to active, predictive, and collaborative control, significantly improving the overall operational efficiency, safety, and intelligent management level of urban intersection clusters.

[0070] Example 1

[0071] To verify the actual effect of the technical solution of this invention, a micro-traffic simulation environment based on SUMO (Simulation of Urban Mobility) was built. The simulation scenario was set as a typical urban center area, containing a 3x3 intersection matrix (9 intersections in total), with an intersection spacing of 400 meters. The traffic demand was set as a typical weekday evening peak period (17:30-18:30). The main roads (the rows and columns running through the center of the matrix) have obvious tidal traffic flow characteristics, while the traffic flow of other secondary roads is relatively balanced. The simulation environment simulated all the hardware described in this invention, including virtual millimeter-wave radar, cameras, and RSUs configured for each intersection. The penetration rate of connected vehicles (OBU) was set to 30%.

[0072] In this embodiment, the system and method of the present invention are fully applied. After the system starts up:

[0073] Data Acquisition: The SUMO simulator outputs information such as the position, speed, and type of each vehicle in real time, simulating the function of the multi-source heterogeneous data acquisition layer 100.

[0074] Digital Twin Prediction: The spatiotemporal prediction digital twin generation module 200 receives simulation data, and the EKF fusion algorithm successfully establishes high-precision tracks for approximately 2,500 vehicles (including connected and non-connected vehicles) in the scenario. The hybrid granularity model predicts the traffic situation for the next 300 seconds, with an accuracy rate (comparing the predicted traffic flow of key road sections with the simulated actual traffic flow) of over 92%.

[0075] Cluster Delineation: The Dynamic Cooperative Intersection Cluster Delineation Module 300 calculates the correlation strength between all intersection pairs based on the prediction results. Due to the evening rush hour, the traffic flow out along the east-west main road is huge, resulting in the correlation strength of the three intersections in this direction (denoted as I4, I5, I6) and the three intersections on the north-south main road perpendicular to it (I2, I5, I8) being significantly higher than other connections. After the Louvain algorithm runs, these nine intersections are dynamically divided into two clusters: one is a large "+" shaped main road cluster consisting of {I2, I4, I5, I6, I8}, and the other is four independent small clusters (or clusters considered as single nodes) consisting of four corner intersections {I1, I3, I7, I9} with weaker mutual correlation.

[0076] Collaborative optimization: For the "+" shaped large cluster, the cluster collaborative optimization control module 400 initiates distributed collaborative optimization. The optimization objective function weights are set as w1 = 0.6 (delay priority), w2 = 0.2 (considering the number of stops), and w3 = 0.2 (maintaining network balance). After about 15 seconds of 10 rounds of iterative calculation, a collaborative timing scheme for the next 5 minutes is generated. This scheme exhibits obvious green wave characteristics, that is, the green light opening time along the east-west main road has a coordinated offset to adapt to the progress of the main traffic flow.

[0077] Execution and Feedback: The generated timing scheme is loaded into SUMO's traffic light controller for execution. The system runs continuously, performing a complete cycle of "prediction-clustering-optimization-execution" every 5 minutes.

[0078] Comparative Example 1

[0079] For comparison, under identical simulation environment, traffic demand, and hardware configuration, a currently advanced local adaptive control method is used as a comparative example. This method operates independently at each intersection, and its logic is similar to:

[0080] SCATS (Sydney Coordinated Adaptive Traffic System) employs a local control strategy where each intersection dynamically adjusts the green light duration for the current cycle based solely on real-time traffic flow and occupancy detected by upstream detectors (simulated cameras or loop detectors) at its approach lanes, using a pre-defined rule base (e.g., extending green lights in the direction of traffic flow, skipping phases without traffic flow, etc.). This method offers real-time responsiveness, but its decision-making perspective is limited to a single intersection, lacking awareness of downstream intersection conditions and the ability to predict future traffic flow. Furthermore, there is no direct communication or collaborative optimization between intersections.

[0081] Performance Comparison

[0082] During the one-hour simulation period, the key performance indicators of Example 1 and Comparative Example 1 were statistically analyzed and compared. The results are shown in the table below:

[0083] Performance indicators Comparative Example 1 (Local Adaptive Control) Example 1 (Solution of the Invention) Performance improvement rate Average network latency (seconds / vehicle) 112.5 76.8 31.7％ Average number of parking times per vehicle on the network 2.8 1.5 46.4％ Average speed on main roads (km / h) 22.4 35.1 56.7％ Import channel saturation variance 0.082 0.045 45.1％ Total network traffic (vehicles / hour) 18,540 21,320 15.0％

[0084] The data in the table above clearly shows that the intersection signal timing dynamic coordination method and system based on real-time traffic prediction proposed in this invention exhibits significant superiority over advanced local adaptive control methods in all key performance indicators. The substantial reduction in average network delay and average number of stops (by 31.7% and 46.4%, respectively) directly proves that this invention successfully reduces vehicle waiting and stopping at intersections through prediction and coordinated control, improving traffic flow smoothness. The significant increase in average speed on main roads (by 56.7%) reflects the macroscopic green wave effect formed by dynamic clustering and coordinated optimization. The reduction in the variance of the saturation of the approach lanes (by 45.1%) indicates that the network equilibrium penalty term in the multi-objective optimization function of this invention has effectively played its role, avoiding the worsening of local congestion and improving the fairness and stability of the entire road network resource utilization. Ultimately, these micro and meso-level improvements have jointly contributed to the improvement of the network's macro-level traffic efficiency, with total traffic volume increasing by 15.0%. These quantitative data powerfully demonstrate that this invention, through a fundamental shift from "local, reactive" control to "global, predictive" proactive collaborative control, can effectively solve existing technical problems and achieve a leapfrog improvement in the level of urban traffic signal control.

[0085] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

[0086] The scope of protection of an invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic coordination system for intersection signal timing based on real-time traffic prediction, characterized in that, The system includes: The multi-source heterogeneous data acquisition layer (100) is configured to acquire the basic data required for constructing a dynamic evolution model of traffic flow in real time in the physical space of the traffic network. The spatiotemporal prediction digital twin generation module (200) communicates with the multi-source heterogeneous data acquisition layer (100) and is configured to integrate the basic data to construct a virtual digital model that is synchronized with the physical road network status and can be extrapolated to the future, and generate prediction data for the traffic status in the future preset time domain based on it. The dynamic cooperative intersection cluster definition module (300) is connected to the spatiotemporal prediction digital twin generation module (200) and is configured to dynamically identify and delineate dynamic cooperative intersection clusters with strong correlation in traffic behavior based on the prediction data. The cluster collaborative optimization control module (400) is configured to independently perform a collaborative signal timing optimization calculation with the goal of maximizing the macro-efficiency of the network for each dynamic collaborative intersection cluster defined by the dynamic collaborative intersection cluster definition module (300), thereby generating an optimal collaborative timing scheme. The signal timing execution and closed-loop feedback module (500) is configured to convert the optimal coordinated timing scheme into a low-level control command and send it to the physical traffic signal controller at the intersection for execution. At the same time, it sends the actually executed signal timing scheme and the real traffic flow data obtained from the multi-source heterogeneous data acquisition layer (100) back to the spatiotemporal prediction digital twin generation module (200) to achieve continuous calibration of the virtual digital model.

2. The intersection signal timing dynamic coordination system based on real-time traffic prediction according to claim 1, wherein the multi-source heterogeneous data acquisition layer (100) specifically includes: An edge sensing unit (110) is deployed at each controlled intersection. The edge sensing unit (110) consists of at least one millimeter-wave radar sensor and a high-resolution optical imaging sensor to acquire kinematic parameters of motor vehicles and non-motor vehicles, vehicle type, lane occupancy and queue length information within the intersection area. The vehicle-road cooperative communication unit (120) adopts the C-V2X communication technology standard and communicates with connected vehicles through roadside units to receive and parse basic safety messages broadcast by connected vehicles, including vehicle position, speed, heading angle and acceleration information. The central cloud platform data interface (130) is used to access the city traffic management center to obtain historical traffic data, real-time weather information and public event information.

3. The intersection signal timing dynamic coordination system based on real-time traffic prediction according to claim 1, characterized in that, The spatiotemporal prediction digital twin generation module (200) integrates the following: The multi-source data fusion and state estimation submodule (210) is configured to receive and fuse data streams from the multi-source heterogeneous data acquisition layer (100), and use a fusion algorithm based on extended Kalman filtering to generate a unified, continuous and accurate state vector for each traffic participant in the road network, thereby forming a snapshot of the global traffic state at the current moment. The hybrid granularity traffic flow propagation modeling submodule (220) is set to predict the evolution of traffic flow in the future preset time domain based on the above global traffic state snapshot as the initial condition; The future state sequence generation submodule (230) is configured to output the calculation results of the mixed granularity traffic flow propagation modeling submodule (220) in a formatted manner to form a three-dimensional data tensor containing the predicted traffic state parameters of each spatial unit in the road network at each discrete time step in the future.

4. The intersection signal timing dynamic coordination system based on real-time traffic prediction according to claim 1, characterized in that, The hybrid modeling strategy adopted by the hybrid granularity traffic flow propagation modeling submodule (220) is as follows: For structured road segments such as arterial roads and expressways in the road network, a macroscopic model based on the cell transmission model is adopted. This model simulates the propagation of traffic waves by dividing the road segment into several cells and solving a set of difference equations describing the flow exchange between cells. For the intersection area and its adjacent complex weaving area, an agent-based micro-simulation model is adopted, in which each traffic participant is modeled as an independent agent whose behavior is driven by car-following and lane-changing models, and whose behavior strictly follows the constraints of traffic lights. The macro model and the micro model exchange data and couple states at the defined regional boundaries to ensure the continuity of the entire road network prediction.

5. The intersection signal timing dynamic coordination system based on real-time traffic prediction according to claim 1, characterized in that, The specific method by which the dynamic collaborative intersection cluster delineation module (300) performs dynamic identification and division is as follows: First, using the predicted data output by the spatiotemporal prediction digital twin generation module (200), the traffic association strength coefficient between any two adjacent or close intersections in the road network is calculated. The traffic association strength coefficient is determined based on the predicted traffic flow function from one intersection to another and the predicted travel time function between them. Then, the entire controlled road network is modeled as an undirected weighted graph, where the nodes of the graph represent intersections, the edges represent the connection relationships between intersections, and the weight of the edges is the traffic association strength coefficient. Finally, a community detection algorithm based on modularity optimization is executed on the undirected weighted graph to divide the graph into several subgraphs, each of which constitutes a dynamic collaborative intersection cluster.

6. The intersection signal timing dynamic coordination system based on real-time traffic prediction according to claim 1, characterized in that, The cluster collaborative optimization control module (400) contains: The multi-objective optimization function construction submodule (410) is configured to construct an optimization function with the objective of maximizing the macro-operational benefits of the entire intersection cluster in the prediction time domain; The hard constraint definition submodule (420) is set to define the minimum green light time, maximum green light time, yellow light time, all-red clear time and fixed signal phase sequence and other safety and physical boundary conditions that must be followed during the optimization process; The distributed collaborative solution algorithm submodule (430) is configured to solve the multi-objective optimization problem under the premise of satisfying the hard constraints, so as to obtain the optimal collaborative timing scheme.

7. The intersection signal timing dynamic coordination system based on real-time traffic prediction according to claim 6, characterized in that, The optimization objective function constructed by the multi-objective optimization function construction submodule (410) is a weighted sum form, and its optimization objectives specifically include: Minimize the total delay time of all vehicles in the cluster within the prediction time domain; Minimize the total number of stops for all vehicles in the cluster during the prediction time domain; To improve the fairness and stability of right-of-way allocation across the entire cluster, a network equilibrium penalty term defined as the variance of the average saturation of all inbound lanes within the cluster in the prediction time domain is minimized.

8. The intersection signal timing dynamic coordination system based on real-time traffic prediction according to claim 6, characterized in that, The distributed collaborative solution algorithm submodule (430) employs a distributed collaborative evolutionary algorithm based on a model predictive control framework, the specific process of which includes: First, a central coordinator breaks down the overall optimization goal of the cluster to each intersection controller within the cluster; Subsequently, each intersection controller uses a genetic algorithm to generate a set of candidate signal timing schemes based on its local and adjacent intersection future traffic flow prediction data. Then, each intersection controller broadcasts its optimal candidate solution and its expected impact on the boundary traffic of adjacent intersections; After receiving information from its neighbors, each intersection controller updates its boundary conditions and repeats local optimization and information exchange until the timing schemes of all controllers in the cluster converge.

9. The intersection signal timing dynamic coordination system based on real-time traffic prediction according to claim 1, characterized in that, The closed-loop feedback implemented by the signal timing execution and closed-loop feedback module (500) specifically includes two feedback paths: First feedback path: The actual signal timing scheme and the actual traffic flow data are sent back to the spatiotemporal prediction digital twin generation module (200) to calculate the error between the prediction model output and the actual observation value, and to perform online calibration of the model parameters in the mixed granularity traffic flow propagation modeling submodule (220) based on this error. The second feedback path is to use the error to evaluate the actual effect of the previous cycle's collaborative optimization control strategy and feed the evaluation results back to the cluster collaborative optimization control module (400) to provide data support for dynamically adjusting the weight coefficients in the multi-objective optimization function.

10. A dynamic coordination method for intersection signal timing based on real-time traffic prediction, characterized in that, This method is applicable to the intersection signal timing dynamic coordination system based on real-time traffic prediction as described in any of claims 1-9; the method includes the following steps: S1: Real-time collection of multi-source heterogeneous traffic data. Through millimeter-wave radar, high-resolution cameras, vehicle-road cooperative roadside units and central cloud platform interfaces deployed in the road network, continuous collection of vehicle kinematic parameters, vehicle attributes, queuing information, connected vehicle data and macro background information containing timestamps. S2: Generate a spatiotemporal prediction digital twin model, fuse the collected multi-source heterogeneous data, generate a snapshot of the global traffic state at the current moment, and use this as the initial condition to drive a mixed-granularity traffic flow propagation model to simulate and predict the evolution of road network traffic flow in the future preset time domain, and finally generate a data tensor containing future predicted traffic state parameters. S3: Dynamically define intersection collaborative clusters. Based on the future state prediction data generated in step two, calculate the traffic association strength between intersections in the road network, abstract the road network into a weighted graph, and run the community detection algorithm on this graph to dynamically divide the road network into several highly interconnected dynamic collaborative intersection clusters. S4: Execute cluster-level collaborative optimization control. For each defined dynamic collaborative intersection cluster, construct a weighted multi-objective optimization function with the total cluster delay, total number of stops, and network balance as the core indicators. Under the condition of satisfying hard constraints such as minimum green light time, use a distributed collaborative evolution algorithm to solve a set of optimal collaborative signal timing scheme. S5: Execute signal timing and implement closed-loop feedback. The optimal coordinated timing scheme generated in step four is sent to the physical signal controllers at each intersection for execution. At the same time, the actual timing scheme is compared with the real traffic flow data collected in the new round, the prediction error is calculated, and this error is fed back to calibrate the traffic flow propagation model online and to evaluate and adjust the optimization control strategy.

Images