Urban traffic signal lamp adaptive control method based on artificial intelligence

Abstract

The invention discloses an urban traffic signal lamp adaptive control method based on artificial intelligence, and the method comprises the steps: firstly converting vehicle data into a fluid field matrix through employing a space grid mapping technology, calculating a virtual fluid dynamic pressure index combining kinetic energy density and acceleration divergence, so as to represent a congestion evolution direction; the system identifies a fluid infinitesimal state, generates a game weight, monitors back propagation characteristics of a dynamic pressure shock wave, sends a phase pre-wake-up request to an upstream intersection, edge computing nodes construct a game revenue function based on a node reputation value and downstream impedance, and searches an optimal timing strategy through distributed interaction. And the system triggers asynchronous signal switching according to the deviation between the optimal strategy and the current state, and executes phase difference nonlinear compensation. According to the method, through quantification of the traffic flow state and the cooperative game, potential congestion sources are effectively identified, non-cooperative behaviors of nodes are inhibited, and real-time blocking of burst shock waves and dynamic coordination of regional signals are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic control technology, specifically to an adaptive control method for urban traffic lights based on artificial intelligence. Background Technology

[0002] Urban traffic signal control systems are critical infrastructure for ensuring the efficiency and safety of urban road networks. By allocating the timing of traffic light phases at intersections, these systems coordinate the right-of-way for vehicles traveling in different directions, aiming to maximize intersection capacity and reduce average vehicle delays. With the continuous growth of urban motor vehicle ownership, traffic signal control technology has evolved from early single-point timing control to adaptive control capable of adjusting timing schemes based on real-time traffic conditions, becoming a core tool in modern urban traffic management.

[0003] In existing urban traffic signal adaptive control systems, sensing devices such as geomagnetic induction coils, microwave radar, or video detectors are typically deployed at intersection approach lanes to collect basic data such as vehicle passage times, traffic flow, and occupancy rates. Based on this detection data, the signal controller uses predetermined control algorithms to adjust cycle length, green light ratio, or phase difference in real time. For example, the system detects vehicle queuing at approach lanes; when a large number of vehicles are detected queuing in a particular direction, it automatically extends the green light time for that phase, or prematurely cuts off the green light when traffic flow is interrupted, thus achieving dynamic allocation of traffic resources according to changing demand.

[0004] However, existing adaptive control methods primarily use the physical queue length or the number of vehicles piled up before the stop line at intersections as the core basis for triggering signal switching and timing optimization. This control logic based on static congestion results exhibits significant lag. Since queue length only reflects the final state after congestion has formed, it cannot quantify the microscopic changes in vehicle speed and acceleration within the traffic flow. This makes it difficult for the control system to perceive the compression and aggregation trend or the dispersion and dissipation trend of traffic flow as it approaches the intersection. The system often only begins to implement diversion strategies after a large-scale standstill of vehicles has occurred, failing to identify potential congestion sources in the early stages of compression disturbances caused by speed differences in the traffic flow. Therefore, it is difficult to proactively intervene and block the evolution of traffic congestion. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based adaptive control method for urban traffic lights. This method solves the problem that existing urban traffic signal control systems primarily rely on fixed timing or inductive control based on current queue length, resulting in insufficient perception of micro-traffic flow patterns and poor regional coordination. Existing inductive control methods typically only collect the current static queue length at intersections, failing to effectively characterize changes in acceleration within the traffic flow and congestion evolution trends. Furthermore, the lack of effective coordination mechanisms between control units at different intersections can easily lead to downstream intersection overflow deadlocks caused by local intersections prioritizing maximum traffic efficiency. In addition, traditional periodic-based control logic exhibits lag in responding to sudden traffic flow surges, leading to congestion spread.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive control method for urban traffic lights based on artificial intelligence, which is applied to a system including a roadside sensing unit, an edge computing node, and a signal execution unit.

[0007] The method first performs physical field mapping of traffic flow multi-source features. The edge computing node receives vehicle data collected by the roadside sensing unit, uses spatial grid mapping technology to divide the intersection entrance area into spatial grid units, and transforms discrete individual vehicle data into a fluid field matrix that reflects the continuous spatial distribution characteristics of the intersection.

[0008] Based on the fluid field matrix, the system calculates the real-time virtual fluid dynamic pressure index of the road segment. Edge computing nodes perform kinetic energy analysis on the traffic flow velocity vector field in the fluid field matrix and divergence calculation on the acceleration vector field. Specifically, the edge computing nodes calculate the square of the normalized velocity vector magnitude of the spatial grid cells and multiply it by the normalized local fluid density to obtain the kinetic energy density component. Simultaneously, the finite difference method is used to calculate the longitudinal and lateral gradient values ​​of the normalized acceleration vector field in its spatial distribution. These two values ​​are then added together and multiplied by the normalized local fluid density to obtain the acceleration divergence component. When synthesizing the virtual fluid dynamic pressure index of the road segment, the system determines the sign of the acceleration divergence component: if it is negative, the traffic flow is determined to be in a compressed state, and the system calculates the difference between the kinetic energy term corresponding to the kinetic energy density component and the damping term corresponding to the acceleration divergence component, and accumulates this difference in the virtual fluid dynamic pressure index; if it is non-negative, only the kinetic energy term is accumulated.

[0009] The system discretizes traffic flow patterns and generates game-theoretic weights based on virtual fluid dynamic pressure indices. Edge computing nodes construct dynamic pressure gradient maps to identify micro-element state attributes. The system calculates the dynamic pressure gradient values ​​of spatial grid cells and, combined with the virtual fluid dynamic pressure index values, labels the spatial grid cells as free-flowing, stable car-following, compressive disturbance, or deadlock / blockage states. The system assigns sequentially increasing weight values ​​to each of these four states, generating a game-theoretic weight mapping matrix corresponding to the intersection spatial grid cell matrix.

[0010] Upon detecting abnormal flow patterns, the system performs prediction and early warning of cross-node dynamic pressure shock waves. Edge computing nodes monitor the rate of change of virtual fluid dynamic pressure indices. When backpropagation characteristics consistent with the shock wave model are detected, the system performs time-series correlation on the dynamic pressure anomaly center and calculates the longitudinal propagation velocity of the dynamic pressure shock wave. For objects identified as backward shock waves, the system uses the longitudinal propagation velocity to linearly extrapolate the trajectory of the dynamic pressure shock wave and calculates the time window in which the backward shock wave is expected to reach the stop line of the upstream adjacent intersection. The edge computing node at the upstream intersection receives the shock wave early warning data packet, establishes a topological mapping between lanes and phases across the intersection, determines the associated phases affected by the shock wave, and constructs a dynamic spatiotemporal influence window based on the shock wave intensity index value, generating a phase pre-wake-up request object.

[0011] The system further executes distributed game decision-making based on a reputation mechanism. Edge computing nodes construct a game payoff function based on real-time virtual fluid dynamic pressure indicators, received phase pre-wake-up requests, and their own node game reputation value. The node game reputation value is generated dynamically: the system obtains the expected dynamic pressure release value submitted by the edge computing nodes in the previous signal control cycle and calculates the actual dynamic pressure release value after the release of the current signal control cycle; it compares the actual dynamic pressure release value with the expected dynamic pressure release value to calculate a single prediction confidence score; if the actual dynamic pressure release value is less than the expected dynamic pressure release value, the single prediction confidence score decreases as the deviation ratio between the two increases; the system uses a moving average algorithm, combining the long-term reputation value of the previous moment with the current single prediction confidence score, to update the node game reputation value at the current moment.

[0012] When constructing the game payoff function, the system uses a game weight mapping matrix to perform a weighted summation of the virtual hydrodynamic pressure values ​​of the spatial grid cells covered by the controlled phase, and combines this with the green light duration of the candidate game strategies to calculate the local payoff. Simultaneously, the system calculates and predicts the traffic flow entering each downstream entrance lane, and combines this with the real-time virtual hydrodynamic pressure index of the adjacent downstream intersection entrance lanes to calculate the downstream impedance penalty term. The system uses the node game reputation value as a gain factor for the local payoff, the downstream impedance penalty term as a loss term, and superimposes a shock wave warning correction term determined based on the phase pre-wake-up request to synthesize the comprehensive game payoff function. For the shock wave warning correction term, if the phase pre-wake-up request level is a level indicating the suppression of green light duration, a negative penalty value is assigned; if the level is a blocking level indicating mandatory red light, a negative infinity value is assigned. Edge computing nodes interact with neighboring edge computing nodes through a distributed communication network to obtain a neighborhood strategy set, and correct the downstream impedance penalty term in the game payoff function based on the neighborhood strategy set. The system then traverses the candidate game strategies and selects the optimal timing strategy that maximizes the game payoff function value.

[0013] Based on the generated strategy, the system executes asynchronous event-driven signal control. Edge computing nodes monitor the current phase's already running green light duration and remaining green light duration in real time, calculating the difference between the target green light duration set by the optimal timing strategy and the total planned green light duration for the current phase, thus obtaining a time deviation value. If the absolute value of the time deviation exceeds the asynchronous trigger threshold, and the current phase meets the minimum safe green light time constraint, the system generates a green light extension command or a green light truncation command and sends it to the signal execution unit. Furthermore, after executing the asynchronous switch, the system also performs phase difference nonlinear compensation: obtaining the physical length of the road segment and the average spatial speed between the current controlled intersection and the downstream adjacent intersection, calculating the ideal free-flow transmission time; establishing a nonlinear positive correlation mapping relationship between the compensation duration and the downstream virtual fluid dynamic pressure index, calculating the phase difference nonlinear compensation term; superimposing the ideal free-flow transmission time and the phase difference nonlinear compensation term to obtain the target phase difference, and generating a phase difference smooth transition command to send to the edge computing nodes of the downstream adjacent intersection.

[0014] This invention provides an adaptive control method for urban traffic lights based on artificial intelligence. It has the following beneficial effects:

[0015] 1. This invention introduces a virtual fluid dynamic pressure index and combines kinetic energy density and acceleration divergence within a spatial grid for dual physical field mapping. It uses acceleration divergence components to quantify the compression and aggregation or dispersion trends within the traffic flow. Compared to traditional detection methods that rely solely on static queue length, this index can identify potential congestion sources in a state of compression disturbance. This allows the control system to detect the deterioration trend of traffic flow before vehicles have completely stopped, thereby achieving accurate characterization and forward-looking prediction of the evolution direction of traffic congestion at intersections.

[0016] 2. This invention constructs a distributed game decision-making model based on a reputation mechanism, mapping the historical prediction accuracy of edge computing nodes to dynamically evolving reputation values ​​and using them as weight gain factors in the game payoff function. At the same time, a downstream impedance penalty term is introduced to form a flow impedance mechanism based on pressure gradient, forcing each intersection node to truthfully reflect traffic demand and consider the impact on downstream areas during the bidding process. This effectively suppresses non-cooperative behavior caused by local nodes falsely reporting demand, and realizes the convergence of regional traffic control strategies from single-point optimization to global collaborative equilibrium.

[0017] 3. This invention implements a prediction and early warning mechanism for cross-node dynamic pressure shock waves, uses the reverse propagation speed of congestion waves to linearly extrapolate the shock wave trajectory to construct a dynamic spatiotemporal influence window, and combines asynchronous event-driven signal control logic to trigger signal switching when the deviation between the optimal strategy and the current state exceeds a threshold. This reduces the injected flow or cuts off the green light at the upstream intersection in advance to block deadlock paths, eliminates the time lag of fixed-cycle control, and achieves a dynamic balance between real-time response to sudden traffic flow shock waves and the coordination stability of trunk green waves. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see the appendix Figure 1 This invention provides an adaptive control method for urban traffic lights based on artificial intelligence.

[0021] The adaptive control system for urban traffic lights provided in this embodiment of the invention mainly consists of a roadside sensing unit, edge computing nodes, a distributed communication network, and a signal execution unit. The roadside sensing unit is deployed at each approach lane of the intersection and is equipped with millimeter-wave radar and a high-pixel optical camera to collect vehicle driving data covering a preset area of ​​the intersection. The edge computing nodes are physically connected to the traffic signal controller at the intersection and have built-in high-performance processors for running fluid dynamics modeling algorithms and multi-agent game algorithms. The distributed communication network is built between edge computing nodes at adjacent intersections, using fiber optic Ethernet or 5G-V2X communication protocols to achieve peer-to-peer transmission of virtual fluid dynamic pressure indicators and control signals between adjacent nodes. The signal execution unit is connected to the driving circuit of the traffic lights, receives phase control commands from the edge computing nodes, and executes light color switching.

[0022] Based on the above system architecture, the overall workflow provided by the embodiments of the present invention is described as follows:

[0023] In step S101, the system performs physical field mapping of traffic flow multi-source characteristics.

[0024] The roadside sensing unit scans the intersection approach lane area at millisecond-level frequencies to acquire the instantaneous position coordinates, vector velocity, and acceleration of discrete vehicles. Edge computing nodes receive the above discrete data and use spatial grid mapping technology to divide the intersection approach lane area into spatial grid cells of preset size. The system statistically analyzes the vehicle density and average velocity vector within each spatial grid cell, transforming the discrete individual vehicle data into a fluid field matrix reflecting the continuous spatial distribution characteristics of the intersection.

[0025] In step S102, the system calculates the real-time virtual hydrodynamic pressure index of the road segment.

[0026] Based on the fluid field matrix constructed in step S101, the edge computing node calculates the virtual fluid dynamic pressure index that reflects the traffic pressure state of the road segment. This calculation process involves kinetic energy analysis of the traffic flow velocity vector field and divergence calculation of the acceleration vector field using the finite difference method. The virtual fluid dynamic pressure index not only characterizes the current static load of vehicle queuing, but also characterizes the kinetic energy decay trend and congestion evolution direction inside the traffic flow through the acceleration divergence term.

[0027] In step S103, the system performs traffic flow pattern discretization and game weight generation.

[0028] Edge computing nodes construct dynamic pressure gradient maps based on dynamic pressure indices, identify four micro-states: free flow, stable following, compression disturbance, and deadlock blocking, and generate the weight mapping matrix required for non-cooperative game based on these indices.

[0029] In step S104, the system performs prediction and early warning of cross-node dynamic pressure shock waves.

[0030] Through a distributed communication network, edge computing nodes continuously monitor the rate of change of virtual hydrodynamic pressure indices at their own intersection and adjacent intersections. When the virtual hydrodynamic pressure indices are detected to exhibit reverse propagation characteristics in their spatial distribution and the propagation speed conforms to the shock wave model, the system determines that a traffic flow shock wave exists. At this time, the associated edge computing nodes located on the shock wave propagation path calculate the time window for the shock wave to arrive at the stop line of the controlled intersection based on the shock wave propagation speed and generate a phase pre-wake-up request. This request includes the expected arrival time of the shock wave and the corresponding affected phase number.

[0031] In step S105, the system performs distributed game decision-making based on a reputation mechanism.

[0032] Each edge computing node at an intersection acts as an independent game player. It constructs a game payoff function based on the intersection's real-time virtual hydrodynamic pressure index, received shock wave warning levels, and its own historical reputation value. Neighboring nodes interact through multiple rounds of information exchange to search for the optimal timing strategy that minimizes overall regional hydrodynamic pressure fluctuations. This strategy corresponds to the green light allocation weights for each traffic phase at the current moment. During this process, if a high-priority pre-wake-up request is received, the game algorithm increases the payoff coefficient of the affected phase in the payoff function, giving it a higher probability of green light allocation during the Nash equilibrium solution process, thus prioritizing the passage time required for shock wave dissipation.

[0033] In step S106, the system performs asynchronous event-driven signal control.

[0034] Based on the game result obtained in step S104, the edge computing node generates a signal phase switching command. The triggering of this command does not depend on a fixed clock cycle, but is event-driven based on changes in virtual fluid dynamic pressure index. When the deviation between the calculated optimal phase duration and the current execution state exceeds a preset threshold, the system immediately issues a switching command to the signal execution unit. At the same time, the signal execution unit uses phase difference dynamic compensation logic to adjust the start time of the signal switching and performs linear smoothing on the green light interval time to maintain fluid stability between intersections.

[0035] For the radar and visual data fusion calibration involved in the roadside perception unit, as well as the specific handshake protocol in the distributed communication network, those skilled in the art can use existing multi-sensor fusion technology and network communication standards to implement it. The specific implementation details are well-known technologies in this field and will not be elaborated here.

[0036] To achieve the overall workflow described above, the edge computing node runs a complete set of algorithmic logic, which includes physical modeling of traffic flow characteristics, feature recognition of dynamic pressure shock waves, and state machine-based phase control. The specific implementation steps of these key processes will be described in detail below.

[0037] The edge computing node performs the steps of extracting and normalizing traffic flow feature parameters, specifically including steps S111 to S114:

[0038] In step S111, multidimensional vehicle trajectory data is acquired.

[0039] Edge computing nodes receive fused data frames generated by millimeter-wave radar and cameras via a data interface. These fused data frames contain the target vehicle's unique identifier (ID) for the detection period, its lateral and longitudinal coordinates in the intersection coordinate system, its longitudinal velocity along the lane direction, its lateral velocity perpendicular to the lane direction, and its longitudinal acceleration. The edge computing nodes employ a Kalman filter algorithm to smooth the raw observations, eliminating measurement noise interference with subsequent differential calculations.

[0040] In step S112, a traffic flow spatial grid mapping is established.

[0041] Edge computing nodes divide the effective driving area of ​​the intersection's approach lane into a fixed-size spatial grid cell matrix. Based on the vehicle's real-time coordinates, the edge computing nodes determine the grid position where the vehicle's geometric center is located. For cases where the vehicle's geometry exceeds the range of a single grid cell, the edge computing nodes execute mapping logic based on area overlap rate. The edge computing nodes calculate the area percentage of the vehicle within different grid cells, assigning the vehicle's presence identifier to the multiple spatial grid cells containing the vehicle according to this area percentage, and using this area percentage as a weighting coefficient for subsequent calculations.

[0042] In step S113, the microfluidic parameters within the spatial grid cell are calculated.

[0043] The edge computing node traverses the spatial grid cell matrix. For non-empty spatial grid cells, the edge computing node statistically analyzes the vehicle characteristics inside the spatial grid cell and calculates the local fluid density, average velocity vector, and average acceleration vector.

[0044] The specific calculation logic is as follows:

[0045] Local fluid density calculation: Edge computing nodes accumulate the weight coefficients mapped to all vehicles within the current spatial grid cell to obtain the weighted vehicle count. The edge computing nodes calculate the ratio of the weighted vehicle count to the physical length of the spatial grid cell to obtain the linear vehicle density. The edge computing nodes divide the linear vehicle density by the lane's design saturation capacity to obtain the dimensionless local fluid density, reflecting the congestion level of the current spatial grid cell.

[0046] Average velocity vector calculation: The edge computing node performs a weighted average calculation of the instantaneous velocity vectors of all vehicles within the current spatial grid cell. Specifically, the edge computing node multiplies the instantaneous velocity vector of each vehicle by its weight coefficient to obtain a weighted value for each vehicle's velocity; the edge computing node sums the weighted values ​​of the velocity of all vehicles within the current spatial grid cell to obtain a weighted sum of velocities; the edge computing node divides this weighted sum of velocities by the sum of the weight coefficients of all vehicles within the current spatial grid cell to obtain the average velocity vector of the current spatial grid cell.

[0047] Average acceleration vector calculation: Edge computing nodes use the same weighted averaging logic as the average velocity vector. Each edge computing node multiplies the instantaneous acceleration vector of each vehicle by its weight coefficient to obtain a weighted acceleration value for each vehicle. The edge computing node then sums the weighted acceleration values ​​of all vehicles within the current spatial grid cell to obtain a weighted acceleration sum. Finally, the edge computing node divides this weighted acceleration sum by the sum of the weight coefficients of all vehicles within the current spatial grid cell to obtain the average acceleration vector for the current spatial grid cell.

[0048] In step S114, a continuous traffic fluid field is constructed and normalized.

[0049] The edge computing nodes interpolate and smooth the discrete grid data generated in step S113 to construct a continuously distributed velocity and acceleration field. Specifically, the edge computing nodes use a Gaussian kernel function to perform convolution operations on the spatial grid cell matrix, and fill empty grids with values ​​from surrounding neighboring grids to eliminate data gaps and form a continuous physical field.

[0050] Subsequently, the edge computing nodes perform dimensionless normalization on the continuously distributed velocity and acceleration fields. The edge computing nodes divide the velocity values ​​in the continuous velocity field by the maximum design speed for the current road segment to obtain the normalized velocity; similarly, they divide the acceleration values ​​in the continuous acceleration field by a preset reference value for maximum vehicle braking acceleration to obtain the normalized acceleration. Through this normalization process, the edge computing nodes unify traffic flow data from different road levels into the same dimensional range, providing a unified data benchmark for subsequent comparisons of virtual hydrodynamic pressure indices across multiple intersections.

[0051] The edge computing node performs the virtual fluid dynamic pressure calculation steps, specifically including steps S121 to S124:

[0052] In step S121, normalized fluid field data is read.

[0053] The edge computing node reads the continuously distributed normalized local fluid density field, normalized velocity field, and normalized acceleration field generated in step S114 from memory. The edge computing node determines the target road segment range for which virtual fluid dynamic pressure indices need to be calculated, and extracts the fluid parameters of all spatial grid cells within the target road segment range.

[0054] In step S122, the kinetic energy density component of the traffic fluid is calculated.

[0055] Edge computing nodes traverse every spatial grid cell within the target road segment, calculating the kinetic energy density component that reflects the current kinetic energy of the traffic flow. The higher the value of the kinetic energy density component, the greater the throughput required to maintain the current state of traffic flow.

[0056] The specific calculation logic is as follows:

[0057] The edge computing node obtains the normalized velocity vector of the current spatial grid cell, and calculates the square of the modulus of the normalized velocity vector. Then, the edge computing node multiplies the square of the modulus with the normalized local fluid density of the current spatial grid cell, and multiplies the result by a preset kinetic energy coefficient (e.g., 0.5) to obtain the kinetic energy density component of the current spatial grid cell.

[0058] In step S123, the traffic fluid acceleration divergence component is calculated.

[0059] Edge computing nodes calculate the divergence of the normalized acceleration vector field in its spatial distribution to quantify the convergence or dispersion trend of vehicles within the traffic flow.

[0060] The specific calculation logic is as follows:

[0061] Edge computing nodes employ finite difference logic for computation. Each edge computing node obtains the longitudinal acceleration components of its downstream and upstream adjacent spatial grid cells along the driving direction of the current spatial grid cell. The node then calculates the difference between these components and divides the difference by twice the physical length of the spatial grid cell along the driving direction to obtain the longitudinal gradient value.

[0062] Similarly, the edge computing node calculates the difference between the lateral acceleration components of the current spatial grid cell and the two adjacent grid cells in the direction perpendicular to the lane, and divides the difference by twice the physical width of the spatial grid cell in the direction perpendicular to the lane to obtain the lateral gradient value.

[0063] Edge computing nodes add the vertical gradient value to the horizontal gradient value and multiply the sum by the normalized local fluid density of the current spatial grid cell to obtain the acceleration divergence component of the current spatial grid cell. For spatial grid cells located at the boundary, edge computing nodes use lateral difference values ​​instead of central difference values ​​for calculation.

[0064] In step S124, the virtual hydrodynamic pressure index of the road section is synthesized.

[0065] Edge computing nodes perform weighted integration of the kinetic energy density component and acceleration divergence component of all spatial grid cells within the target road segment to obtain a virtual hydrodynamic pressure index that reflects the overall traffic pressure of the target road segment.

[0066] The specific calculation logic is as follows:

[0067] The edge computing node initializes the virtual hydrodynamic pressure index to zero. The edge computing node then traverses every spatial grid cell within the target road segment. For each spatial grid cell, the edge computing node determines the sign of the acceleration divergence component.

[0068] If the acceleration divergence component is negative, it means that the traffic flow is in a state of compression and aggregation. The edge computing node multiplies the acceleration divergence component by the preset compression damping coefficient to obtain the damping term. The edge computing node multiplies the kinetic energy density component by the preset kinetic energy weighting coefficient to obtain the kinetic energy term. The edge computing node calculates the difference between the kinetic energy term and the damping term and adds the difference to the virtual fluid dynamic pressure index value.

[0069] If the acceleration divergence component is non-negative, it means that the traffic flow is in a discrete or uniform state. The edge computing node only adds the value of the kinetic energy density component multiplied by the preset kinetic energy weight coefficient to the virtual fluid dynamic pressure index value, ignoring the contribution of the acceleration divergence component.

[0070] The edge computing node reads the corresponding kinetic energy weighting coefficient and compression damping coefficient from a preset parameter table based on the road grade and lane function of the target road segment. Through the above cumulative calculation, the edge computing node finally outputs the virtual hydrodynamic pressure index of the target road segment.

[0071] The edge computing node performs the discretization and infinitesimal element definition steps of the traffic flow state. The edge computing node transforms the continuously changing virtual fluid dynamic pressure field into a discrete state matrix that can be used for subsequent logical judgments, specifically including steps S131 to S134:

[0072] In step S131, a dynamic pressure gradient distribution map is constructed.

[0073] Based on the numerical distribution of the virtual hydrodynamic pressure index of the target road segment output in step S124 on each spatial grid cell, the edge computing node calculates the dynamic pressure gradient along the driving direction of the lane.

[0074] The specific calculation logic is as follows:

[0075] Edge computing nodes traverse every spatial grid cell within the target road segment. For the current spatial grid cell, the edge computing node obtains its virtual hydrodynamic pressure index value. The edge computing node then obtains the virtual hydrodynamic pressure index values ​​of the adjacent spatial grid cells upstream of the current spatial grid cell in the vehicle's direction of travel. The edge computing node calculates the difference between the virtual hydrodynamic pressure index value of the upstream adjacent spatial grid cell and the virtual hydrodynamic pressure index value of the current spatial grid cell. The edge computing node divides this difference by the physical length of the spatial grid cell to obtain the dynamic pressure gradient value of the current spatial grid cell. The dynamic pressure gradient value reflects the degree of change in traffic pressure spatially. If the dynamic pressure gradient value is positive and exceeds a preset gradient threshold, it indicates that the current spatial grid cell is a potential point for the formation of a congestion tail.

[0076] In step S132, the state attributes of the micro-element are defined.

[0077] Edge computing nodes classify each spatial grid cell into a specific fluid state category based on its virtual hydrodynamic pressure index and hydrodynamic pressure gradient. The edge computing nodes have a pre-defined state classification table, which includes free-flow state, steady-following state, compressible disturbance state, and deadlock / blockage state.

[0078] The specific judgment logic is as follows:

[0079] If the virtual hydrodynamic pressure index of the current spatial grid cell is lower than the preset free flow threshold, the edge computing node marks the current spatial grid cell as a free flow state;

[0080] If the virtual hydrodynamic pressure index value of the current spatial grid cell is between the preset free flow threshold and the preset blockage threshold, and the dynamic pressure gradient value of the current spatial grid cell is lower than the preset gradient threshold, the edge computing node marks the current spatial grid cell as a stable following state.

[0081] If the dynamic pressure gradient value of the current spatial grid cell exceeds the preset gradient threshold, the edge computing node ignores the magnitude of the virtual fluid dynamic pressure index value and directly marks the current spatial grid cell as a compression disturbance state. The compression disturbance state indicates that there is a speed conflict within the traffic flow.

[0082] If the virtual hydrodynamic pressure index of the current spatial grid cell exceeds the preset blocking threshold, and the average velocity vector of the current spatial grid cell is close to zero, the edge computing node marks the current spatial grid cell as deadlocked and blocked.

[0083] In step S133, the set of key pressure micro-elements is extracted.

[0084] Edge computing nodes perform connected component analysis on spatial grid cells marked as compressed perturbation states and spatial grid cells marked as deadlocked or blocked states. Using a connected component labeling algorithm, edge computing nodes identify spatially adjacent spatial grid cells of the same type of anomalous state. Each edge computing node defines a set of consecutively adjacent spatial grid cells of the same type of anomalous state as a critical pressure micro-element set. The edge computing node calculates the geometric center coordinates and coverage length of the critical pressure micro-element set, and records the geometric center coordinates of the critical pressure micro-element set as the location of the dynamic pressure anomaly center.

[0085] In step S134, the game weight mapping matrix is ​​generated.

[0086] Based on the micro-element state attributes determined in step S132, the edge computing node assigns corresponding game decision weights to each spatial grid cell.

[0087] The specific allocation logic is as follows:

[0088] Edge computing nodes assign basic weight values ​​(e.g., 1.0) to spatial grid cells in a free-flow state.

[0089] Edge computing nodes assign intermediate weight values ​​(e.g., 1.5) to spatial grid cells in a stable catastrophe state.

[0090] Edge computing nodes assign high-level weight values ​​(e.g., 3.0) to spatial grid cells in a state of compressed perturbation.

[0091] Edge computing nodes assign the highest weight value (e.g., 5.0) to spatial grid cells that are in a deadlock state.

[0092] Edge computing nodes generate game weight mapping matrices that correspond one-to-one with the spatial grid cell matrix of the intersection. In subsequent non-cooperative game steps, edge computing nodes use the game weight mapping matrix to weight the traffic demand in different areas. By assigning higher weights to spatial grid cells in a compressed disturbance state and those in a deadlocked or blocked state, the system ensures that the game algorithm prioritizes responding to traffic micro-elements on the verge of deterioration.

[0093] The edge computing node performs the feature identification step of the dynamic pressure shock wave. Based on the change in the location of the dynamic pressure anomaly center in the time series, the edge computing node calculates the propagation speed and direction of the traffic flow congestion wave, specifically including steps S141 to S144:

[0094] In step S141, time-series correlation of the dynamic pressure anomaly center is performed.

[0095] Edge computing nodes allocate a circular buffer in memory to store information about the key pressure micro-elements set and the coordinates of the corresponding dynamic pressure anomaly centers for the most recent N sampling periods. The edge computing node retrieves the list of dynamic pressure anomaly centers at the current sampling time t, and then calls the circular buffer to retrieve the list of dynamic pressure anomaly centers at the previous sampling time t-1.

[0096] Edge computing nodes employ a minimum Euclidean distance-first matching logic for spatiotemporal association. Each edge computing node iterates through the list of dynamic pressure anomaly centers at the current sampling time. For each dynamic pressure anomaly center at the current sampling time, the edge computing node calculates the Euclidean distance between the current sampling time's dynamic pressure anomaly center and every record point in the previous sampling time's list of dynamic pressure anomaly centers.

[0097] If the calculated minimum Euclidean distance is less than a preset association threshold (e.g., 10% of the lane's physical length), the edge computing node determines that the dynamic pressure anomaly center at the current sampling time and the dynamic pressure anomaly center at the previous sampling time with the corresponding minimum Euclidean distance belong to the same continuous traffic flow compression state. The edge computing node establishes a unique trajectory index association between the dynamic pressure anomaly center at the current sampling time and the dynamic pressure anomaly center at the previous sampling time.

[0098] In step S142, the physical propagation velocity of the dynamic pressure shock wave is calculated.

[0099] For each pair of successfully established trajectory-uniquely associated dynamic pressure anomaly centers, the edge computing nodes calculate the instantaneous propagation velocity vector of the dynamic pressure shock wave.

[0100] The specific calculation logic is as follows:

[0101] The edge computing node is defined with the stop line at the intersection as the origin of the coordinate system, and the positive direction of the longitudinal coordinate axis is the opposite direction of vehicle travel (i.e., the direction pointing upstream of the road segment). The edge computing node calculates the difference between the longitudinal coordinate of the center of dynamic pressure anomaly at the current sampling time and the longitudinal coordinate of the center of dynamic pressure anomaly at the previous sampling time, thus obtaining the longitudinal displacement. The edge computing node divides the longitudinal displacement by the duration of the sampling period to obtain the longitudinal propagation velocity of the dynamic pressure shock wave.

[0102] Edge computing nodes determine the nature of the shock wave based on the sign of the longitudinal propagation velocity. A positive longitudinal propagation velocity indicates that the center of the dynamic pressure anomaly is moving away from the stop line at the intersection, and the edge computing node classifies the dynamic pressure shock wave as a backward shock wave. A backward shock wave indicates that the congestion queue is extending upstream. A negative longitudinal propagation velocity indicates that the center of the dynamic pressure anomaly is moving closer to the stop line at the intersection, and the edge computing node classifies the dynamic pressure shock wave as a forward shock wave. A forward shock wave indicates that the congestion queue is dissipating.

[0103] In step S143, the continuity and intensity of the shock wave are verified.

[0104] Edge computing nodes perform shock wave validity verification to eliminate instantaneous numerical fluctuations caused by sensor noise. The edge computing nodes count the duration of continuous existence of the dynamic pressure anomaly center associated with the unique trajectory index. The edge computing node confirms the dynamic pressure shock wave as valid only if the duration of continuous existence of the dynamic pressure anomaly center's trajectory exceeds a preset time threshold (e.g., 3 seconds).

[0105] Simultaneously, the edge computing nodes calculate the intensity parameters of the hydrodynamic shock wave. The edge computing nodes obtain the virtual hydrodynamic pressure parameters of the spatial grid cells at the wavefront position (i.e., the side closer to the stop line). The edge computing nodes also obtain the virtual hydrodynamic pressure parameters of the spatial grid cells at the waveback position (i.e., the side farther from the stop line). The edge computing nodes calculate the difference between the virtual hydrodynamic pressure parameters at the waveback position and the wavefront position, and record this difference as the shock wave intensity.

[0106] In step S144, the time when the shock wave arrives at the upstream intersection is predicted.

[0107] The edge computing node executes prediction logic only for objects identified as backward shock waves. The edge computing node uses the longitudinal propagation velocity of the backward shock wave calculated in step S142 to linearly extrapolate the trajectory of the dynamic pressure shock wave.

[0108] The specific prediction logic is as follows:

[0109] The edge computing node obtains the longitudinal coordinate position of the dynamic pressure anomaly center at the current sampling time. The edge computing node reads the physical length of the road segment between the current controlled intersection and the upstream adjacent intersection from the high-precision map data. The edge computing node calculates the remaining propagation distance by subtracting the value of the longitudinal coordinate position of the dynamic pressure anomaly center at the current sampling time from the physical length of the road segment. The edge computing node divides the remaining propagation distance by the longitudinal propagation velocity of the backward shock wave to obtain the time window in which the backward shock wave is expected to reach the stop line of the upstream adjacent intersection.

[0110] The edge computing node encapsulates the estimated time window for the backward shock wave to reach the stop line of the upstream adjacent intersection, the shock wave intensity, and the affected lane number into a shock wave warning data packet. The edge computing node then sends this shock wave warning data packet to the edge computing node at the upstream adjacent intersection via a distributed communication network. The edge computing node at the upstream adjacent intersection uses the shock wave warning data packet to trigger phase pre-wake-up logic.

[0111] The edge computing node at the upstream intersection performs the cross-intersection spatiotemporal window association mapping step. The edge computing node at the upstream intersection converts the received shock wave warning data from the downstream intersection into phase scheduling constraints for the local signal controller, specifically including steps S145 to S148:

[0112] In step S145, the shock wave warning data packet is parsed.

[0113] The edge computing node at the upstream intersection receives shock wave warning data packets from the edge computing nodes at the adjacent downstream intersection via a distributed communication network. The edge computing node at the upstream intersection performs protocol parsing on the shock wave warning data packets and extracts the key fields contained in the shock wave warning data packets: the expected start time of the backward shock wave reaching the stop line at the upstream intersection, the intensity index value of the backward shock wave, and the number of the affected approach lane at the downstream intersection.

[0114] In step S146, a lane-phase topology mapping is established across the intersection.

[0115] The edge computing node at the upstream intersection queries the pre-stored intersection topology database. The intersection topology database stores a physical connection mapping table of "upstream exit lane - downstream entrance lane" and a flow direction mapping table of "upstream release phase - upstream exit lane".

[0116] The specific mapping logic is as follows:

[0117] The edge computing node at the upstream intersection uses the affected approach lane number at the downstream intersection as the index key to look up the physically connected upstream intersection exit lane number in the "Upstream Exit Lane - Downstream Approach Lane" physical connection mapping table. Then, using the upstream intersection exit lane number as the index key, the edge computing node at the upstream intersection looks up the upstream intersection traffic phase number responsible for delivering traffic flow to the upstream intersection exit lane in the "Upstream Release Phase - Upstream Exit Lane" flow direction mapping table. The edge computing node at the upstream intersection marks the found upstream intersection traffic phase number as the associated phase affected by the shock wave.

[0118] For example, if the northbound straight lane of the downstream intersection is included in the shock wave warning data packet, the edge computing node of the upstream intersection identifies that the southbound straight lane of the upstream intersection is physically connected to the northbound straight lane of the downstream intersection, and the edge computing node of the upstream intersection locks the southbound straight phase of the upstream intersection as the associated phase affected by the shock wave.

[0119] In step S147, a dynamic spatiotemporal influence window is constructed.

[0120] The edge computing nodes at the upstream intersection define the restricted time period of the associated phase affected by the shock wave on the time axis based on the start time parameter and intensity index value parameter in the shock wave warning data packet.

[0121] The specific construction logic is as follows:

[0122] The edge computing node at the upstream intersection sets the expected start time of the rearward shock wave reaching the stop line at the upstream intersection as the opening time of the spatiotemporal influence window. The edge computing node at the upstream intersection calculates the duration of the spatiotemporal influence window using the intensity index value of the rearward shock wave. The edge computing node at the upstream intersection multiplies the intensity index value of the rearward shock wave by a preset time conversion factor to obtain the base duration. The time conversion factor represents the average number of seconds required for a shock wave of unit intensity to completely dissipate. The edge computing node at the upstream intersection adds the opening time to the base duration to obtain the ending time of the spatiotemporal influence window.

[0123] In step S148, a phase pre-wake-up request object is generated.

[0124] The edge computing node at the upstream intersection generates a phase pre-wake-up request object for input to the game decision module based on the associated phase affected by the shock wave determined in step S146 and the dynamic spatiotemporal influence window constructed in step S147.

[0125] The phase pre-wake-up request object contains the following information: associated phase number, the opening time of the spatiotemporal influence window, the ending time of the spatiotemporal influence window, and the pre-wake-up priority level.

[0126] The edge computing nodes at the upstream intersection determine the pre-wake priority level based on the intensity index value of the backward shock wave:

[0127] If the intensity index of the backward shock wave is lower than the preset blocking threshold, the edge computing node of the upstream intersection sets the pre-wake priority level to "attention level". "Attention level" indicates that the subsequent game algorithm adds a green light duration penalty term for the associated phase in the game payoff function within the time range of the spatiotemporal influence window to suppress the green light extension time of the associated phase.

[0128] If the intensity index of the backward shock wave is higher than or equal to the preset blocking threshold, the edge computing node of the upstream intersection sets the pre-wake priority level to "blocking level". "Blocking level" indicates that the subsequent game algorithm will force the game weight of the associated phase to be reduced to zero or force the associated phase to remain in red light state within the time range of the spatiotemporal influence window, so as to avoid traffic flow injection causing deadlock overflow at the downstream intersection.

[0129] As a specific implementation of the reputation-based distributed game decision-making in step S105, the edge computing node first executes the sub-step of dynamic evolution of the node's game reputation value:

[0130] Step S151: The edge computing node quantifies the prediction confidence of the current intersection in regional collaborative control by comparing the expected effect of historical game strategies with the physical feedback after actual execution. This specifically includes steps S1511 to S1514:

[0131] In step S1511, historical game commitment data is obtained.

[0132] Edge computing nodes query the locally stored historical game data database. They extract the expected dynamic pressure release value submitted by the current edge computing node during the distributed game in the previous signal control cycle. The expected dynamic pressure release value is defined as the numerical value of the virtual hydrodynamic pressure index that the edge computing node calculates during the previous round of bidding, and is expected to eliminate for the current road segment after obtaining the right-of-way in a specific phase.

[0133] In step S1512, the actual dynamic pressure release feedback is calculated.

[0134] After the current signal control cycle's right-of-way phase ends, the edge computing node performs a data backtracking comparison of the road segment status. The edge computing node reads the real-time virtual hydrodynamic pressure index of the target road segment at the end of the right-of-way period. The edge computing node also reads the historical virtual hydrodynamic pressure index of the target road segment at the start of the right-of-way period. The edge computing node calculates the difference between the historical virtual hydrodynamic pressure index value at the start of the right-of-way period and the real-time virtual hydrodynamic pressure index value at the end of the right-of-way period, obtaining the actual hydrodynamic pressure release value. The actual hydrodynamic pressure release value reflects the actual physical effect of traffic flow dissipation after obtaining the green light right-of-way.

[0135] In step S1513, the confidence score for a single prediction is calculated.

[0136] Edge computing nodes compare the expected dynamic pressure release value with the actual dynamic pressure release value, and calculate a single prediction confidence score reflecting the accuracy of the game strategy prediction. This step helps edge computing nodes identify negative prediction bias behavior where the actual effect is far lower than the expected promise.

[0137] The specific calculation logic is as follows:

[0138] Edge computing nodes determine the relationship between the actual dynamic pressure release value and the expected dynamic pressure release value.

[0139] If the actual dynamic pressure release value is greater than or equal to the expected dynamic pressure release value, the edge computing node determines that the game strategy has been executed effectively. The edge computing node sets the confidence score for each prediction to a preset maximum value (e.g., 1.0).

[0140] If the actual dynamic pressure release value is less than the expected dynamic pressure release value, the edge computing node determines that there is a deviation in the game strategy. The edge computing node calculates the single prediction confidence score using the following confidence score calculation formula. :

[0141]

[0142] In the confidence score calculation formula This represents the confidence score for a single prediction. This represents the expected release value of dynamic pressure submitted in the previous round of the game; Represents the actual dynamic pressure release value; This represents the preset deviation sensitivity coefficient, which is used to adjust the system's tolerance to prediction errors. This is a preset, extremely small positive number (e.g., 0.001) to prevent calculation errors caused by a zero denominator. The larger the prediction bias, the lower the confidence score for a single prediction.

[0143] In step S1514, the node's long-term reputation value is updated.

[0144] Edge computing nodes use a moving average algorithm, combining the long-term reputation value from the previous time step with the single-prediction confidence score calculated in step S1513, to update the current long-term reputation value of the edge computing node. The long-term reputation value is used as a weighting factor in the subsequent game payoff function. During the system initialization phase, the edge computing nodes set the initial value of their long-term reputation value to a preset maximum value (e.g., 1.0).

[0145] The specific update logic is as follows:

[0146] Edge computing nodes calculate their long-term reputation value at the current moment using the following reputation value evolution formula. :

[0147]

[0148] In the formula for the evolution of reputation value, This represents the long-term reputation value as updated at the current moment. Represents the long-term reputation score at the previous moment; This represents the confidence score of a single prediction calculated at the current moment; This represents the historical forgetting factor, which ranges from 0 to 1 (e.g., 0.8).

[0149] Edge computing nodes employ a historical forgetting factor to ensure that their long-term reputation values ​​remain relatively stable while also enabling them to respond quickly to recent negative prediction biases. If an edge computing node repeatedly experiences actual dynamic pressure release values ​​significantly lower than expected dynamic pressure release values, its long-term reputation value will decline significantly, reducing its competitiveness for resources in future game decisions.

[0150] Step S152: The edge computing node transforms the physical state of traffic flow, the mutual influence between intersections, and the reputation evaluation results into a mathematical optimization objective function, specifically including steps S1521 to S1524:

[0151] In step S1521, the game strategy space is defined.

[0152] Edge computing nodes determine the set of currently available signal control strategies. At preset time steps (e.g., 1 second), the edge computing nodes generate a series of discrete green light duration options between the minimum safe green light duration and the maximum allowed green light duration. Each discrete green light duration option constitutes a candidate game strategy.

[0153] In step S1522, the weighted dynamic pressure release benefit term is calculated.

[0154] Edge computing nodes calculate the degree of local traffic congestion relief that can be achieved by implementing a candidate game strategy. The edge computing nodes incorporate the game weight mapping matrix generated in step S134 to perform differentiated calculations on spatial grid cells in different states.

[0155] The specific calculation logic is as follows:

[0156] Edge computing nodes construct a positive correlation function based on the weight values ​​of spatial grid cells, dynamic pressure values, and green light duration. In this embodiment, the edge computing nodes calculate local revenue using the following linearly weighted weighted dynamic pressure release revenue formula. :

[0157]

[0158] In the weighted dynamic pressure release benefit formula Represents local revenue; This represents the total number of spatial grid cells currently covered by the controlled phase; For the index of the spatial grid cell; This represents the weight value of the corresponding spatial grid cell in the game weight mapping matrix generated in step S134; The virtual hydrodynamic pressure value represents a single spatial grid cell, which is equal to the difference between the kinetic energy term and the damping term obtained when calculating the spatial grid cell in step S124. The duration of the green light corresponding to the candidate game strategy.

[0159] It should be noted that the above formula is only a preferred method for calculating local revenue. In other embodiments, edge computing nodes may also use nonlinear functions (e.g., taking the logarithm of the green light duration to reflect diminishing marginal returns) to characterize local revenue. As long as the monotonicity characteristics of "high-weighted grids, high dynamic pressure values, and long green light durations corresponding to high revenue values" are met, they all fall within the scope of protection of this invention.

[0160] In step S1523, the downstream impedance penalty term is calculated.

[0161] Edge computing nodes quantify the impact of the current traffic flow at an intersection on downstream adjacent intersections. These nodes acquire real-time virtual hydrodynamic pressure indices of the approach lanes at downstream adjacent intersections via a distributed communication network.

[0162] The specific calculation logic is as follows:

[0163] The edge computing node first calculates the predicted traffic flow into the downstream lane. The edge computing node obtains the saturation flow rate parameter of the currently allowed lane, and multiplies the saturation flow rate parameter by the green light duration of the candidate game strategy to obtain the predicted traffic flow into the downstream lane. Traffic flow of downstream approach lanes .

[0164] Subsequently, the edge computing node constructs a penalty function based on the predicted flow rate and downstream dynamic pressure indicators. In this embodiment, the edge computing node calculates the external cost using the following downstream impedance penalty formula. :

[0165]

[0166] In the downstream impedance penalty formula, This represents the number of downstream approach lanes associated with the current release phase; Represents the downstream adjacent intersection number The current virtual hydrodynamic pressure index value of the approach lane. The physical meaning of the downstream impedance penalty formula is: if the current virtual hydrodynamic pressure index value of the downstream intersection is high, the edge computing node will incur a high penalty cost when sending flow to the high-pressure area, thus forming a flow impedance mechanism based on pressure gradient.

[0167] In step S1524, the comprehensive game payoff function is synthesized.

[0168] The edge computing node combines the long-term reputation value updated in step S1514, the local revenue calculated in step S1522, the external cost calculated in step S1523, and the phase pre-wake-up request generated in step S148 to construct the final comprehensive game payoff function.

[0169] Edge computing nodes use long-term reputation value as a gain factor for local revenue and external costs as a loss term. In this embodiment, the edge computing node uses a comprehensive game payoff formula of the following linear combination form to calculate the total payoff for a specific candidate game strategy. :

[0170]

[0171] In the comprehensive game payoff formula The total payoff value representing the candidate game strategy; This represents the long-term reputation value of the edge computing node updated in step S1514. This represents the weighted average dynamic pressure release benefit item; Represents the downstream impedance penalty term; This represents a preset collaborative damping coefficient, used to adjust the balance between local and global interests; This represents the shock wave warning correction item.

[0172] Constructed using this formula, the system establishes the following game rules: when the long-term reputation value of the edge computing node... Even at lower levels, local revenue Very high, its total revenue It will also be significantly reduced. Mathematically, this is equivalent to lowering the node's priority in resource contention, thereby forcing nodes with low reputation to reduce their green light requirements until they restore their reputation through honest behavior.

[0173] Correction item for shock wave warning The edge computing nodes are assigned numerical values ​​according to the pre-wake-up priority level determined in step S148:

[0174] If there is no phase pre-wake-up request, the edge computing node will The value is assigned to zero.

[0175] If a phase pre-wake-up request with a priority level of "Attention" exists, the edge computing node will... The value is assigned a preset negative penalty to reduce the competitiveness of the current candidate game strategy.

[0176] If a phase pre-wake-up request with a priority level of "blocking" exists, the edge computing node will... Assigning a value of negative infinity forces the current candidate game strategy to exit the solution space.

[0177] Step S153: The edge computing nodes, through multiple rounds of strategy interaction and iterative calculation with neighboring edge computing nodes, search for the optimal signal control scheme that enables the overall regional benefit to reach a Nash equilibrium state. This specifically includes steps S1531 to S1535:

[0178] In step S1531, the game iteration parameters are initialized.

[0179] The edge computing node establishes a game iteration counter to record the negotiation rounds and then resets the game iteration counter to zero. The edge computing node determines the initial game strategy for the currently controlled phase. In this embodiment, the edge computing node sets the initial game strategy to the currently executing signal timing scheme. The edge computing node sets a maximum allowed number of iterations threshold (e.g., 10 iterations) and a convergence threshold for judging the stability of the result (e.g., the change in green light duration is less than 1 second) to prevent the algorithm from timeout or failing to converge within a finite time.

[0180] In step S1532, neighborhood policy interaction is performed.

[0181] Edge computing nodes broadcast their local game strategy for the current iteration to all geographically adjacent edge computing nodes via a distributed communication network. Simultaneously, each edge computing node receives the current game strategies from all its neighboring edge computing nodes. The edge computing node defines the set of strategies received from all neighboring edge computing nodes as the neighborhood strategy set. The neighborhood strategy set contains the planned green light durations for all neighboring intersections that influence traffic flow input or output at the current intersection in the current iteration.

[0182] In step S1533, the optimal response strategy is calculated.

[0183] Based on the received neighborhood policy set, the edge computing node searches within its local policy space for a policy that maximizes the comprehensive game payoff function constructed in step S1524.

[0184] The specific calculation logic is as follows:

[0185] The edge computing node first adjusts the downstream impedance penalty term in the comprehensive game payoff function based on the neighborhood strategy set. The edge computing node predicts the changing trend of the virtual hydrodynamic pressure index of the downstream approach lanes when a specific neighborhood strategy (i.e., a specific green light duration) is implemented at a downstream adjacent intersection. If the neighborhood strategy indicates an increased green light duration, it means the downstream intersection's traffic flow capacity is enhanced, and the edge computing node accordingly lowers the value of the downstream impedance penalty term in the comprehensive game payoff function; conversely, if the neighborhood strategy indicates a decreased green light duration, the edge computing node increases the value of the downstream impedance penalty term.

[0186] Subsequently, the edge computing nodes traverse each discrete candidate game strategy defined in step S1521. For each candidate game strategy, the edge computing nodes calculate the total payoff value corresponding to that candidate game strategy, incorporating the modified downstream impedance penalty term. The edge computing nodes compare the total payoff values ​​of all candidate game strategies and select the candidate game strategy with the largest total payoff value as the update strategy for the next round. This process is logically equivalent to finding the optimal response solution for the current neighborhood environment.

[0187] In step S1534, the Nash equilibrium convergence determination is performed.

[0188] The edge computing node calculates the magnitude of the difference between the update strategy generated in this round and the current strategy in the previous round. In this embodiment, the magnitude of the difference is represented by the absolute difference in the duration of the green light.

[0189] If the difference is less than the preset convergence threshold, or the game iteration counter reaches the preset maximum iteration threshold, the edge computing node determines that the game process has reached a stable state. At this time, the edge computing node terminates the iteration and confirms the update strategy generated in this round as the final optimal timing strategy, and passes it to step S1535 for parsing.

[0190] If the difference is greater than or equal to the convergence threshold but has not reached the maximum number of iterations threshold, the edge computing node will increment the game iteration counter by one unit and return to execution step S1532 to enter the next round of information interaction and strategy update.

[0191] In step S1535, the final control command is output.

[0192] The edge computing node parses the optimal timing strategy confirmed in step S1534 and extracts the green light duration parameter included in the optimal timing strategy. The edge computing node converts the green light duration parameter into a specific phase switching time point. During the conversion process, the edge computing node combines the remaining time of the current signal cycle to calculate the absolute timestamp of the target phase opening or ending, generates a signal phase switching command and caches it, waiting for the trigger logic call in step S106.

[0193] As the basic state management step S106, the edge computing node performs the maintenance and transition steps of the signal phase pre-wake-up state machine. The edge computing node instantiates an independent finite state machine object for each traffic signal phase to manage the control state transition of the traffic signal phase in response to dynamic pressure shock waves, specifically including steps S161 to S164:

[0194] In step S161, the state machine instance and state definitions are initialized.

[0195] Edge computing nodes establish state machine instances for each controllable traffic signal phase at the intersection. Each state machine instance contains four mutually exclusive operating states: normal scan state, pre-lock state, active suppression state, and damped recovery state.

[0196] The normal scanning state is defined as the state in which the traffic signal phase is not affected by dynamic pressure shock waves and operates according to conventional game control logic.

[0197] The pre-lock state is defined as the state in which the system enters the control parameter locking phase when the traffic signal phase has received a shock wave warning but the shock wave has not yet physically reached the stop line of the upstream intersection.

[0198] Active suppression state is defined as the state in which the system is currently within the effective range of the dynamic spatiotemporal influence window, and the edge computing nodes apply intervention measures to the traffic signal phase;

[0199] The damping recovery state is defined as the state in which the dynamic spatiotemporal influence window has ended, the edge computing nodes gradually withdraw intervention measures, and the traffic signal phase smoothly transitions back to the normal scanning state.

[0200] In step S162, clock synchronization and status trigger monitoring are performed.

[0201] Edge computing nodes read the system's current timestamp at millisecond intervals. The edge computing nodes then iterate through all traffic signal phase state machine instances that are in a normal scanning state.

[0202] For each traffic signal phase state machine instance, the edge computing node checks whether there is an associated phase pre-wake-up request object. If a phase pre-wake-up request object exists, the edge computing node compares the current system timestamp with the spatiotemporal influence window opening time in the phase pre-wake-up request object.

[0203] If the system's current timestamp falls within a preset buffer time interval before the opening time of the spatiotemporal influence window (e.g., 5 seconds in advance), the edge computing node switches the traffic signal phase state machine instance from the normal scanning state to the pre-locking state. In the pre-locking state, the edge computing node locks the minimum green light time parameter of the current traffic signal phase and prohibits the current traffic signal phase from responding to other temporary insertion requests, ensuring that the current traffic signal phase remains stable before the shock wave arrives.

[0204] In step S163, the control strategy under active suppression state is executed.

[0205] When the current timestamp of the system is greater than or equal to the opening time of the spatiotemporal influence window and less than the closing time of the spatiotemporal influence window, the edge computing node switches the traffic signal phase state machine instance to the active suppression state.

[0206] In active suppression mode, edge computing nodes perform differentiated control based on the pre-wake priority level in the phase pre-wake request object:

[0207] If the pre-wake-up priority level is "Attention Level," the edge computing node calls the game weight correction interface. In the subsequent calculation of the game payoff function, the edge computing node adds a negative penalty term to the payoff value of the current traffic signal phase. The value of the negative penalty term is proportional to the shock wave intensity index value. By adding the negative penalty term, the edge computing node reduces the probability of the current traffic signal phase obtaining the green light right-of-way in the game competition, thereby reducing traffic flow entering the downstream congestion area.

[0208] If the pre-wake-up priority level is "blocking level," the edge computing node activates hard constraint interception logic. The edge computing node forcibly modifies the maximum permissible green light time for the current traffic signal phase to the minimum safe green light time (e.g., 5 to 10 seconds), or the edge computing node directly blocks the bidding qualification for the current traffic signal phase in the game theory algorithm, forcing the current traffic signal phase to remain in a red state. The aforementioned forced red light operation aims to physically cut off traffic flow to the congestion source, preventing deadlock overflow.

[0209] In step S164, the execution state exits and damping is restored.

[0210] When the current timestamp of the system is greater than the end time of the spatiotemporal influence window, the edge computing node switches the traffic signal phase state machine instance from the active suppression state to the damped recovery state.

[0211] In the damped recovery state, the edge computing node does not immediately zero out the negative penalty term or remove the maximum allowable green light time limit. Instead, the edge computing node starts a recovery cycle counter. At the end of each signal control cycle, the edge computing node calculates the current recovery progress percentage.

[0212] The specific recovery logic is as follows:

[0213] The edge computing node calculates the ratio of the current recovery cycle counter value to the preset total recovery cycle. Using a linear interpolation algorithm, the edge computing node gradually decreases the negative penalty term or gradually increases the maximum allowed green light time based on this ratio. For example, if the preset total recovery cycle is 3 cycles, at the end of the first recovery cycle, the edge computing node will reduce the negative penalty term value to 66% of its initial value.

[0214] Edge computing nodes monitor the actual queue length of traffic signal phases in real time. If the actual queue length is lower than the preset normal threshold and the negative penalty term has decayed to zero, the edge computing node switches the traffic signal phase state machine instance back to the normal scanning state, completing a full pre-wake-up control loop.

[0215] Based on maintaining the aforementioned state machine, the edge computing node executes step S165, which is the asynchronous triggered signal switching logic sub-step. The edge computing node dynamically determines whether to immediately interrupt or extend the current phase based on the deviation between the game decision result and the current execution state. This specifically includes steps S1651 to S1654:

[0216] In step S1651, the current phase execution status is monitored in real time.

[0217] Edge computing nodes access the real-time operational status data interface of the traffic signal controller at a preset high-frequency sampling rate (e.g., 10Hz). The edge computing node reads the currently active signal phase number, the current green light duration, and the preset remaining green light duration for the current phase within the current cycle. Simultaneously, the edge computing node obtains the optimal timing strategy output in step S1535, which includes the target green light duration calculated based on the latest traffic flow pattern.

[0218] In step S1652, the timing strategy deviation is calculated.

[0219] Edge computing nodes quantify the difference between the currently executing timing scheme and the latest optimal timing strategy.

[0220] The specific calculation logic is as follows:

[0221] The time deviation value of edge computing nodes is calculated using the following deviation calculation formula. :

[0222]

[0223] In the formula for calculating the deviation... Represents the time deviation value; The target green light duration is set in the optimal timing strategy output by step S1535. This represents the duration of the green light that has been in operation for the current phase. This represents the remaining green light duration recorded under the current controller logic for the current phase. The sum of the values ​​represents the total green light duration that would have been planned for the current phase if no intervention had been taken.

[0224] like A positive value indicates that the optimal timing strategy requires extending the green light time for the current phase; if... A negative value indicates that the optimal timing strategy requires the green light time of the current phase to end early.

[0225] In step S1653, hysteresis comparison and safety constraint verification are performed.

[0226] Edge computing nodes incorporate a hysteresis threshold determination mechanism to prevent signal oscillations caused by frequent traffic light switching within a short period. Simultaneously, edge computing nodes perform traffic safety regulation constraint verification.

[0227] The specific judgment logic is as follows:

[0228] The edge computing node first checks the minimum safety constraints. It determines whether the green light duration of the current phase is less than the preset minimum safe green light time, and whether the current duration meets the minimum clearance time required for pedestrians to cross the street.

[0229] If the current phase's green light duration does not meet the minimum safe green light time or the minimum clearance time required for pedestrians to cross the street, the edge computing node will not trigger a switching operation and will maintain its current operating state, regardless of the time deviation value.

[0230] If the above minimum security constraints are met, the edge computing node further determines whether the absolute value of the time deviation exceeds the preset asynchronous trigger threshold (e.g., 3 seconds).

[0231] If the absolute value of the time deviation is less than the asynchronous trigger threshold, the edge computing node determines that the current timing deviation is within the allowable range, and the system maintains the original countdown logic without triggering the asynchronous event.

[0232] If the absolute value of the time deviation is greater than or equal to the asynchronous trigger threshold, the edge computing node determines that the current timing scheme is significantly different from the actual needs, and the system triggers an asynchronous switching signal.

[0233] In step S1654, an asynchronous switching instruction is generated.

[0234] Based on the determination result of step S1653, the edge computing node generates a forced switching instruction that overrides the original control logic.

[0235] If the time deviation value If the value is positive and its absolute value exceeds the asynchronous trigger threshold, the edge computing node generates a green light extension command. The edge computing node sends a countdown reset signal to the signal execution unit, updating the remaining green light duration for the current phase. The value.

[0236] If the time deviation value If the value is negative and its absolute value exceeds the asynchronous trigger threshold, the edge computing node generates a green light cutoff command. The edge computing node sends a forced jump signal to the signal execution unit, instructing the signal controller to immediately terminate the green light of the current phase and enter the yellow light transition state.

[0237] In step S166, after performing asynchronous signal switching, the edge computing node recalculates the coordinated phase difference with adjacent intersections to eliminate the risk of green wave coordination failure caused by asynchronous switching. This specifically includes steps S1661 to S1664:

[0238] In step S1661, road segment fluid transport parameters are extracted. The edge computing node reads the physical length of the road segment between the currently controlled intersection and the downstream adjacent intersection from the intersection topology database. Simultaneously, the edge computing node calls the normalized velocity field data generated in step S114. The edge computing node calculates the normalized average speed of the current traffic flow within the target road segment coverage area. The edge computing node multiplies the normalized average speed by the current road segment's maximum design speed, restoring the dimensionless value to a spatial average speed with physical units (meters per second). Furthermore, the edge computing node obtains the real-time virtual hydrodynamic pressure index values ​​of the coordinated approach lanes corresponding to the current traffic flow direction at the downstream adjacent intersection through a distributed communication network.

[0239] In step S1662, the ideal free-flow transport time is calculated. The edge computing nodes calculate the reference time required for a vehicle to pass through the target road segment without obstruction, based on the physical length of the road segment and the average spatial travel speed.

[0240] The specific calculation logic is as follows:

[0241] Edge computing nodes divide the physical length of a road segment by the average spatial speed to obtain the ideal free-flow transmission time. The ideal free-flow transmission time represents the shortest physical time required for a stream of vehicles released upstream to reach the downstream stop line under conditions of extremely low traffic density and no downstream queuing interference.

[0242] In step S1663, a nonlinear stagnation compensation model is constructed. Considering that increased downstream traffic flow density leads to a nonlinear decrease in vehicle speed and that downstream queue dissipation waves experience reverse propagation delays, the edge computing node introduces a nonlinear compensation term based on a virtual fluid dynamic pressure index.

[0243] The specific calculation logic is as follows:

[0244] Edge computing nodes establish a nonlinear positive correlation mapping between compensation duration and downstream virtual fluid dynamic pressure parameters. In this embodiment, the edge computing nodes use the following nonlinear phase difference compensation formula, which includes a nonlinear exponential term, to calculate the target phase difference. :

[0245]

[0246] In the phase difference nonlinear compensation formula This represents the target phase difference, which is defined as the time difference between the green light start time at the upstream intersection and the green light start time at the downstream intersection with coordinated phase. Represents the physical length of the road segment; The spatial average speed calculated in step S1661; The real-time virtual hydrodynamic pressure index represents the coordinated approach lanes at downstream adjacent intersections; This represents the preset resistance sensitivity coefficient (e.g., 0.2). Represents a nonlinear exponent (e.g., 1.5); This represents the start-up time constant (e.g., 2 seconds) caused by the vehicle's acceleration from a standstill.

[0247] It should be noted that the above formula is only one specific implementation of the nonlinear hysteresis compensation model. In other embodiments, the edge computing nodes can also use an exponential function model or a piecewise linear model instead of a power function model. This is as long as the model satisfies the condition that "as the downstream virtual fluid dynamic pressure index..." As the pressure increases, the value of the compensation term shows a monotonically increasing trend, especially in the high dynamic pressure range (e.g., The nonlinear characteristic of "the growth rate of ) is higher than that of the low dynamic pressure range" is covered within the protection scope of this step.

[0248] By introducing an exponential term When the virtual hydrodynamic pressure index value at the downstream intersection increases, the nonlinear compensation term increases significantly, leading to a substantial increase in the calculated target phase difference value. The increased phase difference delays the opening time of the downstream green light, thus allowing sufficient time for the downstream intersection to clear queuing vehicles in advance.

[0249] In step S1664, a phase difference correction instruction is applied. The edge computing node calculates the difference between the target phase difference and the execution phase difference set in the current signal controller.

[0250] If the absolute value of the difference is less than the preset fine-tuning dead zone threshold (e.g., 2 seconds), the edge computing node keeps the current execution phase difference unchanged in order to maintain the stability of the control system.

[0251] If the absolute value of the difference is greater than or equal to the fine-tuning dead zone threshold, the edge computing node generates a phase difference smooth transition command. The edge computing node encapsulates the target phase difference into a communication protocol packet and sends it to the edge computing nodes at downstream adjacent intersections. Upon receiving the command, the edge computing nodes at downstream adjacent intersections initiate step adjustment logic. In subsequent consecutive signal cycles, the edge computing nodes at downstream adjacent intersections adjust their local executed phase difference towards the target phase difference value by a fixed time step (e.g., no more than 3 seconds per cycle) until the executed phase difference matches the target phase difference.

[0252] In step S167, the edge computing node suppresses the traffic flow speed dispersion caused by asynchronous switching by dynamically adjusting the transition time parameter during signal phase switching, specifically including steps S1671 to S1674:

[0253] In step S1671, the non-steady-state characteristics of traffic flow are monitored.

[0254] Edge computing nodes analyze the dispersion of vehicle driving status within the intersection area in real time. In the time window before the signal phase is about to switch (e.g., 5 seconds before the switch), the edge computing node uses the normalized acceleration field data generated in step S114 to calculate the statistical variance of the acceleration of all vehicles in the intersection area. If the number of detected vehicle samples is insufficient, the edge computing node sets the traffic flow non-steady-state index to zero by default, and determines that the traffic flow is in a stable state.

[0255] The specific calculation logic is as follows:

[0256] Edge computing nodes statistically analyze the instantaneous acceleration of vehicles within the spatial grid cells covered by the current traffic release phase. If the number of detected vehicle samples exceeds a preset minimum sample threshold, the edge computing nodes calculate the statistical variance of these instantaneous accelerations and define this variance as the traffic flow non-steady-state index. The traffic flow non-steady-state index quantifies the dispersion of speed changes in traffic flow: a higher index value indicates that there are both rapid acceleration and rapid deceleration behaviors in the traffic flow, and the traffic flow is in a non-steady state; a lower index value indicates that the speed change trend of the traffic flow is consistent, and the traffic flow is in a stable car-following state.

[0257] In step S1672, the stability damping strength is determined.

[0258] Edge computing nodes introduce damping evaluation logic based on the rate of change of virtual fluid dynamic pressure to determine the control damping strength during signal switching.

[0259] The specific calculation logic is as follows:

[0260] The edge computing node first calculates the time rate of change of the overall virtual hydrodynamic pressure index of the currently controlled intersection. The edge computing node uses discrete differential logic to calculate the difference between the virtual hydrodynamic pressure index at the current sampling time and the virtual hydrodynamic pressure index at the previous sampling time, and divides the difference by the sampling time interval to obtain the rate of change of dynamic pressure, which characterizes the intensity of instantaneous fluctuations in congestion pressure.

[0261] Subsequently, the edge computing node constructs a damping strength evaluation model based on the dynamic pressure change rate and the traffic flow unsteady-state index obtained in step S1671. The edge computing node assumes a positive correlation between damping strength and both the dynamic pressure change rate and the traffic flow unsteady-state index. That is, when the virtual fluid dynamic pressure index at the intersection fluctuates drastically in a short period of time, or when the traffic flow unsteady-state index shows extremely discrete vehicle driving states, the edge computing node calculates a higher stability damping coefficient; conversely, if the dynamic pressure changes gradually and the traffic flow is stable, the edge computing node calculates a lower stability damping coefficient.

[0262] In step S1673, the green light interval time is dynamically adjusted.

[0263] The edge computing node dynamically compensates for the all-red time during the upcoming signal switching process based on the stability damping coefficient determined in step S1672. The all-red time refers to the clearing time during which all approach lanes are red after the yellow light of the previous phase ends and before the green light of the next phase turns on.

[0264] The specific adjustment logic is as follows:

[0265] The edge computing nodes use the baseline full-red time, which conforms to traffic safety regulations, as a benchmark. Based on the stability damping coefficient, the edge computing nodes calculate additional dynamic compensation time. They then add the baseline full-red time to the dynamic compensation time to obtain the target full-red time required for this switchover. Simultaneously, the edge computing nodes set a maximum allowable increase in time to ensure the target full-red time does not extend indefinitely.

[0266] Based on the above logic, when traffic flow is in a highly unsteady state or when dynamic pressure changes drastically, the system automatically extends the full red time. The increased full red time physically provides a longer clearance time at the intersection, ensuring that vehicles in the previous phase can completely leave the conflict area and preventing the released traffic flow of the next phase from clashing with the stranded vehicles.

[0267] In step S1674, damping attenuation and state reset are performed.

[0268] After the signal switch is completed and the green light of the next phase is turned on, the edge computing node continuously monitors the changes in the traffic flow non-steady-state index.

[0269] If the traffic flow non-steady-state index is detected to fall below the preset stability threshold, the edge computing node determines that the flow stability has been restored. At this time, in subsequent signal cycles, the edge computing node executes linear attenuation logic to gradually reduce the dynamic compensation duration of the full-red time until the full-red time is restored to the basic full-red time, and then releases the damping control state to restore the maximum traffic efficiency of the intersection.

[0270] 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. An adaptive control method for urban traffic lights based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Perform physical field mapping of traffic flow multi-source features: The collected vehicle driving data is acquired, and the intersection entrance lane area is divided into spatial grid units using spatial grid mapping technology. Discrete individual vehicle data is then transformed into a fluid field matrix that reflects the continuous spatial distribution characteristics of the intersection. Step S2, calculate the real-time virtual hydrodynamic pressure index of the road segment: Kinetic energy analysis is performed on the traffic flow velocity vector field in the fluid field matrix, and divergence calculation is performed on the acceleration vector field to calculate the virtual fluid dynamic pressure index. Step S3: Perform traffic flow pattern discretization and game weight generation: Based on the virtual fluid dynamic pressure index, a dynamic pressure gradient map is constructed, the state attributes of the micro-element are identified, and the game weight mapping matrix required for non-cooperative game is generated. Step S4: Perform prediction and early warning of cross-node dynamic pressure shock waves: The rate of change of the virtual fluid dynamic pressure index is monitored. When the back propagation characteristics that conform to the shock wave model are detected, a phase pre-wake-up request containing the expected arrival time of the shock wave and the affected phase number is generated. Step S5: Execute distributed game decision-making based on reputation mechanism: The game payoff function is constructed based on the real-time virtual fluid dynamic pressure index, the received phase pre-wake-up request, and the current node game reputation value. The optimal timing strategy is then searched through strategy interaction. Step S6, execute asynchronous event-driven signal control: Based on the deviation between the optimal timing strategy and the current execution state, a signal phase switching command is generated.

2. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 1, characterized in that, In step S2, the specific steps for calculating the virtual hydrodynamic pressure index include: Calculate the kinetic energy density components of traffic fluid: Calculate the square of the normalized velocity vector magnitude of the spatial grid cell and multiply it with the normalized local fluid density to obtain the kinetic energy density components; Calculate the traffic fluid acceleration divergence components: The longitudinal and lateral gradient values ​​of the normalized acceleration vector field in the spatial distribution are calculated using the finite difference method. The two values ​​are added together and multiplied by the normalized local fluid density to obtain the acceleration divergence components. Synthetic road section virtual hydrodynamic pressure index: Determine the sign of the acceleration divergence component. If it is negative, calculate the difference between the kinetic energy term corresponding to the kinetic energy density component and the damping term corresponding to the acceleration divergence component, and add it to the virtual hydrodynamic pressure index; if it is non-negative, only the kinetic energy term is added.

3. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 1, characterized in that, In step S3, the specific steps for identifying the state attributes of the micro-element include: Calculate the dynamic pressure gradient value of the spatial grid cell and, in conjunction with the virtual fluid dynamic pressure index value, mark the spatial grid cell as a free flow state, a steady following state, a compressive disturbance state, or a deadlock blockage state. The generation of the game weight mapping matrix includes: assigning sequentially increasing weight values ​​to the spatial grid cells in the free flow state, stable following state, compressed disturbance state, and deadlock blocking state, thereby generating a game weight mapping matrix corresponding to the intersection spatial grid cell matrix.

4. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 1, characterized in that, In step S4, the specific steps for performing prediction and early warning of cross-node dynamic pressure shock waves include: Time series correlation was performed on the center of dynamic pressure anomaly to calculate the longitudinal propagation velocity of the dynamic pressure shock wave; For objects identified as backward shock waves, the trajectory of the dynamic pressure shock wave is linearly extrapolated using the longitudinal propagation velocity, and the time window in which the backward shock wave is expected to reach the stop line of the adjacent upstream intersection is calculated. The shock wave warning data is acquired, a lane-phase topology mapping is established across the intersection, the associated phases affected by the shock wave are determined, and a dynamic spatiotemporal influence window is constructed based on the shock wave intensity index value to generate the phase pre-wake-up request.

5. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 1, characterized in that, In step S5, the node game reputation value evolves dynamically through the following steps: Obtain the expected dynamic pressure release value submitted in the previous signal control cycle, and calculate the actual dynamic pressure release value after the current signal control cycle ends. The actual dynamic pressure release value is compared with the expected dynamic pressure release value to calculate a single prediction confidence score; if the actual dynamic pressure release value is less than the expected dynamic pressure release value, the single prediction confidence score decreases as the deviation ratio between the two increases. By using the moving average algorithm, and combining the long-term reputation value of the previous time step with the current single prediction confidence score, the node game reputation value at the current time step is updated.

6. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 1, characterized in that, In step S5, the specific steps for constructing the game payoff function include: Using the game weight mapping matrix, the virtual hydrodynamic pressure values ​​of the spatial grid cells covered by the controlled phase are weighted and summed, and the local payoff is calculated in combination with the green light duration of the candidate game strategy. Calculate and predict the traffic flow entering each downstream inlet lane, and combine the real-time virtual hydrodynamic pressure index of the inlet lanes of adjacent downstream intersections to calculate the downstream impedance penalty term. The node game reputation value is used as a gain factor for local revenue, the downstream impedance penalty term is used as a loss term, and a shock wave warning correction term determined based on the phase pre-wake-up request is superimposed to synthesize a comprehensive game revenue function.

7. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 6, characterized in that, The assignment logic for the shock wave early warning correction term is as follows: Based on the pre-wake priority level in the phase pre-wake request, if the level is the attention level indicating the duration of green light suppression, then the shock wave warning correction item is assigned a negative penalty value; if the level is the blocking level indicating the forced red light, then the shock wave warning correction item is assigned a negative infinity.

8. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 1, characterized in that, In step S5, the specific steps for searching for the optimal timing strategy include: Obtain a set of neighborhood strategies through data interaction; The downstream impedance penalty term in the game payoff function is modified according to the neighborhood strategy set, and the candidate game strategies are traversed to select the strategy that maximizes the game payoff function value as the update strategy. Determine whether the difference between the updated strategy and the previous strategy is less than the convergence threshold. If the threshold is met, the updated strategy is confirmed as the optimal timing strategy.

9. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 1, characterized in that, In step S6, the specific steps for generating the signal phase switching command include: Real-time monitoring of the current phase's existing green light duration and remaining green light duration; The time deviation value is obtained by calculating the difference between the target green light duration set by the optimal timing strategy and the total green light duration originally planned for the current phase. If the absolute value of the time deviation exceeds the asynchronous trigger threshold, and the current phase meets the minimum safe green light time constraint, then a green light extension command or a green light cutoff command is generated.

10. The adaptive control method for urban traffic lights based on artificial intelligence according to claim 1, characterized in that, Following step S6, the process further includes performing phase difference nonlinear compensation: Obtain the physical length and average speed of the road segment between the current controlled intersection and the downstream adjacent intersection, and calculate the ideal free-flow transmission time; Establish a nonlinear positive correlation mapping relationship between compensation duration and downstream virtual fluid dynamic pressure index, and calculate the phase difference nonlinear compensation term; The ideal free-flow transmission time is superimposed with the phase difference nonlinear compensation term to obtain the target phase difference, and a phase difference smooth transition command is generated.

Images