Traffic signal control method and system based on physical prior constraint time sequence attention

Abstract

The present application relates to the technical field of intelligent transportation system, in order to solve the problem that the prior art cannot identify the risk in advance and the core risk lacks direct quantification and constraint, provide a traffic signal control method and system based on physical prior constraint time sequence diagram attention, the method comprises the following steps: calculating the physical storage capacity of the road section, then combining the current number of vehicles to obtain the remaining storage space of the downstream road section, i.e. overflow safety buffer; the overflow safety buffer is converted into overflow risk cost, and a constrained Markov decision process model is constructed with traffic efficiency as reward and overflow risk as cost; the pre-trained time sequence diagram attention network is used to solve the constrained Markov decision process model, and the optimal signal control action balancing traffic efficiency and overflow risk is obtained. It can identify and actively avoid overflow risk in advance, so as to actively avoid the cascading congestion caused by queue overflow.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation systems, and in particular to a physical prior constraint time series graph attention traffic signal control method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Under the condition of high-density vehicle operation, the city traffic system often triggers a chain reaction due to queue spillover at local intersections, eventually leading to large-scale gridlock, which seriously affects the traffic efficiency and system resilience of the road network. Traditional traffic signal control methods, such as fixed timing or local induction control, are difficult to intervene in the formation and propagation of congestion in a forward-looking manner due to their limited information sources and lagging reaction modes, especially in complex topological structures and densely populated short road networks. In urban networks, it is difficult to effectively suppress the risk of cascading congestion.

[0004] In the prior art, deep reinforcement learning (Deep Reinforcement Learning) and graph neural network (Graph Neural Network, GNN) methods are combined to control adaptive traffic signals, which can effectively model the spatial dependence between intersections and improve the ability of multi-intersection coordinated control. However, the prior art still has the following problems: (1) In the time dimension, the model usually only relies on fixed-length historical data or short-term memory units, making it difficult to accurately capture the formation, propagation, and dissipation of congestion waves and other dynamic processes, resulting in the inability of the strategy to identify risks in advance when the road network approaches a critical state. (2) The optimization goal of the strategy is mainly around efficiency indicators such as travel delay and queue length, and there is a lack of direct quantification and constraints for the core risk of "avoiding queue spillover" at the network level, making the learned strategy tend to passively relieve congestion after it occurs, rather than actively controlling pressure before spillover. SUMMARY

[0005] To solve the above technical problems, the present application provides a physical prior constraint time series graph attention traffic signal control method and system, which can prospectively identify and actively avoid spillover risks, achieve adaptive trade-offs between traffic efficiency and network safety, and thus actively avoid cascading congestion caused by queue spillover.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] The first aspect of the present application provides a physical prior constraint time series graph attention traffic signal control method.

[0008] In one or more embodiments, a physical prior constraint timing diagram attention traffic signal control method is provided, comprising:

[0009] According to the traffic flow model based on inventory conservation and real-time traffic data, the physical storage capacity of the road section is calculated, and then combined with the current number of vehicles to obtain the remaining storage space of the downstream road section, i.e. the overflow safety buffer;

[0010] According to the preset safety criterion, the overflow safety buffer is converted into an overflow risk cost, and a constrained Markov decision process model is constructed, which takes the passing efficiency as the reward and the overflow risk as the cost. The state of the constrained Markov decision process model is the state information of the intersection node obtained based on the road network topology and real-time traffic data; the action is to select the next signal phase;

[0011] The pre-trained timing diagram attention network is used to solve the constrained Markov decision process model to obtain the optimal signal control action that balances the passing efficiency and the overflow risk. In the training process of the timing diagram attention network, the Lagrange dual method optimization algorithm is used to adaptively adjust the penalty weight of the overflow risk cost.

[0012] As an implementation, the state information of the intersection node is represented by the state information of the intersection node, the lane occupancy rate, the queue length, the average speed, the signal phase and the overflow safety buffer.

[0013] As an implementation, according to the preset safety criterion, the overflow safety buffer is converted into an overflow risk cost:

[0014] ;

[0015] Wherein, is a safety buffer threshold; is a penalty curvature; is the overflow risk cost at time t; is the overflow safety buffer of the road section at time t; is a set of road sections; represents taking the maximum value. As an implementation, the calculation process of the overflow safety buffer is:

[0016]

[0017] ;

[0018] ;

[0019] Wherein, is the overflow safety buffer of the road section at time t; is the overflow safety buffer of the road section at time t; ​​​A constant overflow safety buffer; It is a section of road The effective length, It is the upper limit of the blocking density. It refers to the number of lanes; For road section exist Number of vehicles at any given time; It refers to storage capacity, i.e., the maximum number of vehicles that can be accommodated.

[0020] As one implementation method, Lagrange multipliers are introduced during the training of the temporal graph attention network. The constrained Markov decision process model described above is transformed into an unconstrained dual problem, with the corresponding Lagrangian function set as follows: :

[0021] ;

[0022] in, These are Lagrange multipliers, representing the dynamic price of risk; For policy network parameters; In strategy The following expectations; A discount factor for future returns; for The cost of spillover risks at any given moment; For traffic efficiency.

[0023] In one implementation, the time-series graph attention network comprises a multi-head graph attention network and a gated recurrent unit. The multi-head graph attention network is used to capture the attention weights between intersection nodes in the road network topology, and the gated recurrent unit is used to propagate states along the time axis and learn the temporal dynamics of congestion wave propagation. The multi-head graph attention network captures attention weights from intersection nodes... To the intersection node attention weights for:

[0024] ;

[0025] in, and It is node embedding. and It is a learnable parameter matrix. It is the coefficient of the time penalty term; It is a safety buffer threshold; for Transpose of; This represents an activation function with a leakage slope; It is the intersection node Relevant timestamps / update times; This represents a vector concatenation operation; It indicates that they are directly proportional.

[0026] As one implementation method, during the training of the temporal graph attention network, a proximal policy optimization algorithm is used to update the policy network parameters. Its alternative objective function also includes the advantage of return. and cost advantage :

[0027] ;

[0028] Using Lagrange multipliers The algorithm is updated using a subgradient ascent method, enabling it to dynamically adjust the penalty for risk.

[0029] ;

[0030] in It's the learning rate; when actual risk exceeds the budget, Increase the punishment; conversely, decrease it. Optimize the objective function for the PPO strategy; This indicates a batch of time steps sampled. Take the average; It represents the probability ratio; The shear threshold of PPO; The updated Lagrange multipliers will be used for the next round of training / optimization; For the first Lagrange multipliers in the next iteration; In the current strategy Below, the price Expectations; Cost-based upper limit / risk budget; It is a data processing function used to limit numerical values ​​to a specified range; values ​​that exceed the range will be truncated to boundary values. It takes the minimum value; It takes the maximum value.

[0031] A second aspect of the present invention provides a timing graph-based attention traffic signal control system based on physical prior constraints.

[0032] In one or more embodiments, a physical prior constraint-based timing graph attention traffic signal control system includes:

[0033] An overflow safety buffer calculation module is configured to calculate a physical storage capacity of a road section according to a traffic flow model based on inventory conservation and real-time traffic data, and then obtain a remaining storage space of a downstream road section, i.e., an overflow safety buffer, in combination with a current number of vehicles;

[0034] A decision process model construction module is configured to convert the overflow safety buffer into an overflow risk cost according to a preset safety criterion, and construct a constrained Markov decision process model with a passing efficiency as a reward and an overflow risk as a cost; wherein a state of the constrained Markov decision process model is state information of an intersection node obtained based on a road network topology structure and real-time traffic data; and an action is to select a next signal phase;

[0035] An optimal signal control action determination module is configured to solve the constrained Markov decision process model by using a pre-trained time series graph attention network, and obtain an optimal signal control action balancing the passing efficiency and the overflow risk; wherein in a training process, the time series graph attention network adaptively adjusts a penalty weight for the overflow risk cost by using a Lagrange dual method optimization algorithm.

[0036] A third aspect of the present application provides a computer-readable storage medium.

[0037] A computer-readable storage medium has a computer program stored thereon, and the program, when executed by a processor, implements the steps in the physical-prior-constrained time series graph attention traffic signal control method described above.

[0038] A fourth aspect of the present application provides an electronic device.

[0039] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the steps in the physical-prior-constrained time series graph attention traffic signal control method described above when executing the program.

[0040] Compared with the prior art, the present application has the following beneficial effects:

[0041] The present application introduces an overflow safety buffer and converts it into an overflow risk cost, and then constructs a constrained Markov decision process model with a passing efficiency as a reward and an overflow risk as a cost, solves the constrained Markov decision process model by using a pre-trained time series graph attention network, determines an optimal signal control action balancing the passing efficiency and the overflow risk, and realizes active avoidance control of cascading congestion in a city road network by combining traffic engineering physical priors and a time series graph attention network, effectively solves the technical problems that a traditional control method is difficult to suppress queue overflow due to lack of foresight and an existing learning method tends to "passive clearing" due to a single optimization goal, and finally realizes the coordinated improvement of traffic efficiency and safety and stability of a city complex road network. BRIEF DESCRIPTION OF DRAWINGS

[0042] The accompanying drawings, which form a part of the specification, are included to provide further understanding of the application and are incorporated in and constitute a part of this specification. The illustrative embodiments of the application and their description are used to explain the application.

[0043] Figure 1 is a flow chart of a traffic signal control method based on physical prior constraint timing diagram attention of an embodiment of the application;

[0044] Figure 2 is a structural schematic diagram of a traffic signal control system based on physical prior constraint timing diagram attention of an embodiment of the application;

[0045] Figure 3 is a throughput comparison chart of a traffic signal control method based on physical prior constraint timing diagram attention and a traditional algorithm of an embodiment of the application;

[0046] Figure 4 is a delay rate comparison chart of a traffic signal control method based on physical prior constraint timing diagram attention and a traditional algorithm of an embodiment of the application;

[0047] Figure 5 is an average travel time comparison chart of a traffic signal control method based on physical prior constraint timing diagram attention and a traditional algorithm of an embodiment of the application;

[0048] Figure 6 is an average queue length comparison chart of a traffic signal control method based on physical prior constraint timing diagram attention and a traditional algorithm of an embodiment of the application;

[0049] Figure 7 is a schematic diagram of an electronic device of an embodiment of the application. DETAILED DESCRIPTION

[0050] The application will be further described below in conjunction with the drawings and embodiments.

[0051] It should be noted that the following detailed description is merely exemplary and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0052] It is to be understood that the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. It is to be understood that the terms "comprising," "including," and "having" can be used interchangeably.

[0053] Terminology Explanation:

[0054] Queue Spillover: This refers to the phenomenon where, at an intersection, the length of the waiting queue exceeds the physical storage capacity of the road segment, with the tail of the queue extending to the upstream intersection, thus blocking upstream traffic.

[0055] Gridlock: refers to a state in urban road networks where vehicles become immobile due to large-scale queue overflows and intertwining, paralyzing traffic in the entire area.

[0056] Graph Neural Network (GNN): A deep learning model specifically designed for processing graph-structured data. It learns node representations by passing and aggregating information between nodes, making it well-suited for modeling the interactions between intersections.

[0057] Constrained Markov Decision Process (CMDP): A mathematical framework for reinforcement learning that aims to maximize cumulative reward while satisfying one or more constraints on cumulative cost.

[0058] Figure 1 A schematic diagram of the timing graph attention-based traffic signal control method based on physical prior constraints, according to an embodiment of the present invention, is provided. Figure 1 The physical prior constraint time graph attention traffic signal control method of this embodiment may include the following steps S101 to S103.

[0059] The specific implementation process of steps S101 to S103 is as follows:

[0060] Step S101: Calculate the physical storage capacity of the road segment based on the traffic flow model based on inventory conservation and real-time traffic data, and then combine it with the current number of vehicles to obtain the remaining storage space of the downstream road segment, i.e., the overflow safety buffer.

[0061] The Cell Transmission Model (CTM) is used to model traffic flow on road segments. The traffic flow model based on inventory conservation is as follows: Road segment exist Number of vehicles at any given time Depend on Number of vehicles and inflow at any given time and outflow The decision is made based on the downstream road segment's capacity to receive the outflow from the upstream segment, which in turn determines the inflow to the downstream segment. Strict constraints:

[0062] ;

[0063] ;

[0064] wherein the calculation process of the overflow safety buffer is as follows:

[0065] ;

[0066] ;

[0067] wherein, is the effective length of the road section at the time of overflow safety buffer; is the effective length of the road section , is the upper limit of the congestion density, is the number of lanes; is the number of vehicles on the road section at the time of , is the storage capacity, i.e., the maximum number of vehicles that can be accommodated. The overflow safety buffer

[0068] , i.e., the remaining storage space (receiving capacity) of the downstream road section. The overflow safety buffer intuitively quantifies “how far is it from overflowing”, and when , the risk increases dramatically.

[0069] The embodiment of the present application introduces the “overflow safety buffer” index derived from the traffic physics theory, so that the model can identify and respond to potential queue overflow risks earlier, and realize the transition from “passive response” to “active avoidance”.

[0070] Step S102: According to the preset safety criterion, the overflow safety buffer is converted into an overflow risk cost, and a constrained Markov decision process model is constructed, which takes the passing efficiency as the reward and the overflow risk as the cost; wherein the state of the constrained Markov decision process model is the state information of the intersection node obtained based on the road network topology and real-time traffic data; the action is to select the next signal phase.

[0071] To guide the control decision, a single-cycle safety criterion is proposed: for a given release phase, if the number of net inflow vehicles in the most unfavorable case exceeds the current safety buffer of the downstream road section within its duration, the phase is considered to be at risk. The above safety criterion is converted into a continuous, differentiable convex cost function :

[0072] ;

[0073] wherein,​ It is a safety buffer threshold (e.g., 15% of storage capacity). It is a penalty curvature; for The cost of spillover risks at any given moment; It is a section of road exist A constant overflow safety buffer; For road segments; This indicates that the function takes the maximum value. The function handles overflow "close" to the maximum value (i.e., ...). Apply a smooth and progressively increasing penalty to the state.

[0074] Formal definition of CMDP (Constrained Markov Decision Process Model). The adaptive signal control problem is formulated as a CMDP tuple. ,in:

[0075] state Status information of intersection nodes;

[0076] action Select the next signal phase;

[0077] Transfer Determined by the traffic simulation environment;

[0078] Return This refers to traffic efficiency, which is usually a negative cumulative delay or pressure value at intersections.

[0079] cost That is, the cost of spillover risk. ;

[0080] Discount This refers to the discount factor for future returns.

[0081] The urban transportation network is abstracted as a directed graph. , where nodes Represents an intersection, with directed edges. A road segment representing a direction of travel. Real-time traffic status information for each road segment is obtained through sensors deployed at intersections (such as geomagnetic coils and video detectors).

[0082] The acquired status information It not only includes traditional traffic variables, but also integrates physical priors that serve the core of this invention. The state information of the intersection node is represented as follows: ;

[0083] in, for Status information of intersection nodes at any time; for lane occupancy rate of the intersection node at time t, is queue length of the intersection node at time t, is average speed of the intersection node at time t, is signal phase of the intersection node at time t, is overflow safety buffer of the intersection node at time t.

[0084] The optimization goal of CMDP is to maximize the long-term cumulative return while satisfying the long-term average risk cost below the preset budget :

[0085] ;

[0086] where, is the expectation under policy ; is traffic efficiency at time t; is the discount factor of future return; is the set time period; is the maximum value under policy . This formalization turns the "active risk aversion" from a vague concept to a solvable optimization problem with a clear mathematical definition. In the unified constrained reinforcement learning framework, the traffic efficiency (return) and network safety (cost constraint) are optimized collaboratively to achieve an adaptive and interpretable trade-off between the two, rather than simply pursuing a single indicator.

[0087] Step S103: using the pre-trained time series graph attention network to solve the constrained Markov decision process model to obtain the optimal signal control action balancing traffic efficiency and overflow risk; wherein, in the training process of the time series graph attention network, the Lagrange dual method optimization algorithm is used to adaptively adjust the penalty weight of the overflow risk cost.

[0088] In the process of training the time series graph attention network, the Lagrange multiplier is introduced to convert the above constrained Markov decision process model into an unconstrained dual problem, and the corresponding Lagrange function is set as :

[0089] ;

[0090] where, is the dynamic price of risk, which is automatically learned by the algorithm; is the policy network parameter; is the maximum value under policy The expectation below, that is, "according to the strategy" The expected value of the trajectory (state-action-reward-cost sequence) generated by interaction with the environment is simply calculated. In other words, it is to calculate the average effect after considering randomness. A discount factor for future returns; for The cost of spillover risks at any given moment; For traffic efficiency.

[0091] The temporal graph attention network consists of a multi-head graph attention network (GAT) and a gated recurrent unit (GRU). The GAT captures the attention weights between intersection nodes in the road network topology, while the GRU propagates states along the time axis and learns the temporal dynamics of congestion wave propagation. The GAT captures the attention weights between intersection nodes... To the intersection node attention weights for:

[0092] ;

[0093] in, and It is node embedding. and It is a learnable parameter matrix. It is the coefficient of the time penalty term (hyperparameter or learnable parameter), and the larger it is, the more "stale information is penalized"; It is a safety buffer threshold; for Transpose of; This represents an activation function with a leakage slope; It is the intersection node Relevant timestamps / update times; This represents a vector concatenation operation; It indicates that they are directly proportional.

[0094] GAT's attention weight In the calculation, a negative bias term based on the downstream safety buffer is explicitly introduced; the above formula ensures that when neighboring nodes... downstream buffer When it is nearing exhaustion, it is passed to the node. The information weight will increase significantly, forcing the model to "pay attention" to potential sources of risk.

[0095] During the training of the temporal graph attention network, the proximal policy optimization algorithm is used to update the policy network parameters. Its alternative objective function also includes the advantage of return. and cost advantage :

[0096] ;

[0097] Utilize Lagrange multiplier Update by sub-gradient ascent method, which can dynamically adjust the punishment strength of risk:

[0098] ;

[0099] Wherein is the learning rate; when the actual risk exceeds the budget, increase, strengthen the punishment; otherwise, reduce; is the PPO strategy optimization objective function; Indicates that a batch of time steps sampled Take the average; is the probability ratio; is the PPO clipping threshold; is the updated Lagrange multiplier, used for the next round of training / optimization; is the Lagrange multiplier at the th iteration; is the expected cost under the current strategy ; is the upper limit of the cost constraint / risk budget; is a data processing function for limiting numerical values within a specified range, and values exceeding the range will be truncated to the boundary value; is the minimum value; is the maximum value.

[0100] The embodiment of the present application integrates the risk indicator into the attention mechanism and action filtering layer of the graph neural network, thereby strengthening the sensitivity of the model to the critical state of the network from two aspects of feature representation and decision space, and improving the robustness of the strategy.

[0101] At the strategy execution level, the embodiment of the present application adds a final action filtering layer. For any candidate phase proposed by the strategy network, the system will be tested according to the safety criterion. If the phase is determined to have high overflow risk under the current state, it will be removed from the candidate action set, ensuring that the final executed decision always meets the most basic physical safety boundary.

[0102] Through the implementation of the above steps, the embodiment of the present application is based on the deep integration of traffic flow physical priori and time series graph attention network, and through the collaborative optimization of passing efficiency and overflow risk in the constraint reinforcement learning framework, effectively suppresses the propagation of queue overflow and cascading congestion in urban road network, realizes the forward-looking and stable traffic signal control, and provides strong support for improving the resilience and intelligent level of modern urban transportation system.

[0103] The embodiment of the present application combines the physical theory of traffic engineering and deep reinforcement learning technology, uses the interpretable physical quantity of "overflow safety buffer" to guide and constrain the decision-making process of the artificial intelligence model, thereby improving the traffic efficiency and operational stability of the traffic network. The method not only reduces the average driving delay, but also significantly reduces the probability of gridlock. In addition, the embodiment of the present application also has good scalability and adaptability, and can continuously optimize the model through continuous training on new traffic data to adapt to the long-term changes of urban development and traffic patterns.

[0104] The embodiment adopts an algorithm based on a cellular transmission model to quantify the risk of the road network, and provides prospective physical priori for the strategy by calculating the "overflow safety buffer"; a time series graph attention network (PC-STGAT) model customized for the traffic network is adopted, which captures the spatial coupling between intersections through a graph attention mechanism and learns the time dynamics of congestion propagation through a gated recurrent unit. A proximal policy optimization (PPO-Lagrangian) algorithm based on Lagrangian duality is used to jointly optimize the reward signal representing traffic efficiency and the cost signal representing congestion risk, and by dynamically adjusting the Lagrange multiplier, the two goals of "efficiency improvement" and "risk avoidance" are adaptively balanced. Through the deep coupling of physical priori, spatio-temporal representation learning and constraint optimization, the performance, robustness and safety of traffic signal control in high saturation and high uncertainty scenarios are significantly improved.

[0105] The embodiment aims to realize the "active avoidance" type control of cascading congestion in urban road networks by combining traffic engineering physical priori and time series graph attention networks, and at the same time, using deep reinforcement learning algorithms to learn the traffic dynamic evolution law under high density and high uncertainty in a simulation environment, to improve the sensitivity and adaptability of the control strategy to the critical state of the network. The urban traffic network under high load presents significant space-time coupling characteristics, and its congestion evolution is a highly nonlinear dynamic process. Deep learning models, especially spatio-temporal graph neural networks, can flexibly model such complex topological dependencies and dynamic evolution relationships. The method effectively solves the technical problems of traditional control methods lacking of foresight to suppress queue overflow, and existing learning methods tending to "passive clearing" due to single optimization goal, and finally realizes the coordinated improvement of traffic efficiency and safety stability of complex urban road networks.

[0106] In simulation tests, the present application not only improves the average traffic efficiency, but also significantly reduces the probability of gridlock, shortens the congestion recovery time, and shows superior network-level stability and resilience.

[0107] As Figure 2As shown, the traffic signal control system based on the physical prior constraint time sequence diagram attention provided by the embodiment of the application can be implemented in a software manner, and the traffic signal control system based on the physical prior constraint time sequence diagram attention includes the following software modules: an overflow safety buffer calculation module 201, a decision process model construction module 202, and an optimal signal control action determination module 203.

[0108] The functions of each software module in the traffic signal control system based on the physical prior constraint time sequence diagram attention are introduced as follows:

[0109] The overflow safety buffer calculation module 201 is used to calculate the physical storage capacity of a road section according to a traffic flow model based on inventory conservation and real-time traffic data, and then obtain the remaining storage space of a downstream road section, i.e., the overflow safety buffer, in combination with the current number of vehicles.

[0110] The decision process model construction module 202 is used to convert the overflow safety buffer into an overflow risk cost according to a preset safety criterion, and construct a constraint Markov decision process model with the passing efficiency as a reward and the overflow risk as a cost. The state of the constraint Markov decision process model is the state information of an intersection node obtained based on the road network topology and real-time traffic data, and the action is to select the next signal phase.

[0111] The optimal signal control action determination module 203 is used to solve the constraint Markov decision process model by using a pre-trained time sequence diagram attention network, so as to obtain an optimal signal control action balancing the passing efficiency and the overflow risk. In the training process of the time sequence diagram attention network, a Lagrange dual method optimization algorithm is used to adaptively adjust the penalty weight of the overflow risk cost.

[0112] It should be noted that each module in the traffic signal control system based on the physical prior constraint time sequence diagram attention of the embodiment of the application corresponds to each step in the traffic signal control method based on the physical prior constraint time sequence diagram attention of the above embodiment, and the specific implementation process is the same, which will not be described here.

[0113] Through experimental comparison and analysis, Figures 3-6 The comparative results of the traffic signal control method based on the physical prior constraint time sequence diagram attention of the embodiment of the application and the traditional algorithms CoLight, MAPG, IPPO, and MaxPressure in terms of the throughput, delay rate, average travel time, and average queue length are given respectively.

[0114] In the field of traffic control, CoLight is a multi-agent traffic signal control model based on reinforcement learning and graph attention networks, which models traffic signal control as a Markov game, with each intersection controlled by an agent, selecting actions (such as phase switching) by observing local traffic states (such as lane vehicle count and phase), with the goal of minimizing average queue length.

[0115] MAPG, Multi-Agent Policy Gradient, is an extension of reinforcement learning in multi-agent systems, which directly optimizes random policy parameters through gradient ascent to maximize expected returns. The core idea is to evaluate the joint policy using global reward signals and calculate gradient updates for each agent to update its own policy.

[0116] IPPO, Independent PPO, is a fully decentralized algorithm that uses the single-agent algorithm PPO to train each agent.

[0117] MaxPressure algorithm is an intelligent traffic signal control method based on pressure difference. The core idea is to use the data accumulation degree of each node in the dynamic perception network as the core driving force for traffic scheduling, and to preferentially serve the upstream and downstream link pairs with the largest pressure, so as to achieve rapid relief of global congestion.

[0118] According to Figures 3-6 In a high-throughput environment of 3900 vehicles / hour, the physical prior constraint timing graph attention traffic signal control method of the embodiment of the present application achieved excellent results, with higher throughput than other mainstream comparative algorithms, and significantly lower average travel time, delay rate and queue length than traditional algorithms such as CoLight, MAPG, IPPO and MaxPressure.

[0119] The structure of the electronic device of the embodiment of the present application will be described in detail below. Figure 7 The composition structure diagram of the electronic device provided by the embodiment of the present application can be understood as follows. Figure 7 Only exemplary structures of the electronic device are shown, not all structures, and the shown part or all structures can be implemented as needed.

[0120] The electronic device provided by the embodiment of the present application comprises at least one processor 701, a memory 702, a user interface 703 and at least one network interface 704. The various components in the physical prior constraint timing diagram attention traffic signal control system are coupled together through a bus system 705. It can be understood that the bus system 705 is used to realize the connection communication between the components. The bus system 705 includes not only a data bus, but also a power bus, a control bus and a status signal bus. However, for the purpose of clear illustration, all the buses are marked as the bus system 705 in the Figure 7

[0121] The user interface 703 can include a display, a keyboard, a mouse, a trackball, a click wheel, a key, a button, a touchpad or a touch screen, etc.

[0122] It can be understood that the memory 702 can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. The memory 702 in the embodiment of the present application can store data to support the operation of the terminal. Examples of the data include any computer programs for operating on the terminal, such as an operating system and an application program. The operating system contains various system programs, such as a framework layer, a core library layer and a driver layer, for realizing various basic services and processing hardware-based tasks. The application program can include various application programs.

[0123] In some embodiments, the physical prior constraint timing diagram attention traffic signal control system provided by the embodiment of the present application can be realized in a combination of software and hardware. As an example, the physical prior constraint timing diagram attention traffic signal control system provided by the embodiment of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the physical prior constraint timing diagram attention traffic signal control method provided by the embodiment of the present application. For example, the processor in the form of a hardware decoding processor can use one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs) or other electronic elements.

[0124] ​As an example, the processor 701 can be an integrated circuit chip having a processing capability of signals, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor.

[0125] As an example, the processor 701 can be an integrated circuit chip having a processing capability of signals, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor.

[0126] The memory 702 in the embodiment of the present application is used to store various types of data to support the operation of the physical prior constraint-based timing diagram attention traffic signal control system, or to store program codes for executing the method shown in the embodiment of the present application. Figure 1 Examples of these data include any executable instructions for operating on the physical prior constraint-based timing diagram attention traffic signal control system, such as executable instructions, and the program implementing the physical prior constraint-based timing diagram attention traffic signal control method of the embodiment of the present application can be included in the executable instructions.

[0127] In particular, according to the embodiments of the present application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiments of the present application include a computer program product comprising a computer program carrying on a computer readable medium, and the computer program contains program codes for executing the method shown in the embodiments of the present application. Figure 1 In such embodiments, the computer program can be downloaded and installed from a network by a communication part, and / or installed from a detachable medium. When the computer program is executed by the central processing unit, various functions defined in the device of the present application are executed.

[0128] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart and / or block diagram in the variation of the present application can be implemented by computer program instructions. These program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing apparatus, or other programmable data processing apparatus to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, generate a means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 The flowchart and / or block diagram in the variation of the present application can be implemented by computer program instructions. These program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing apparatus, or other programmable data processing apparatus to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, generate a means for implementing the functions specified in the flowchart and / or block diagram block or blocks.

[0129] The above only is the preferred embodiment of the present application, and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A traffic signal control method based on physical prior constraints and time-series graph attention, characterized in that, include: Based on the traffic flow model based on inventory conservation and real-time traffic data, the physical storage capacity of the road segment is calculated. Then, combined with the current number of vehicles, the remaining storage space of the downstream road segment is obtained, which is the overflow safety buffer. Based on preset safety criteria, overflow safety buffer is transformed into overflow risk cost, and a constrained Markov decision process model is constructed with traffic efficiency as the reward and overflow risk as the cost; wherein, the state of the constrained Markov decision process model is the state information of the intersection node obtained based on the road network topology and real-time traffic data; the action is to select the next signal phase; By using a pre-trained temporal graph attention network to solve the constrained Markov decision process model, the optimal signal control action that balances traffic efficiency and spillover risk is obtained. During the training process, the temporal graph attention network adopts the Lagrange dual method optimization algorithm to adaptively adjust the penalty weights for spillover risk cost. Based on the preset safety criteria, the cost of converting the overflow safety buffer into an overflow risk is as follows: ； in, It is a safety buffer threshold; It is a penalty curvature; for The cost of spillover risks at any given moment; It is a section of road exist A constant overflow safety buffer; For road segments; This indicates taking the maximum value; Lagrange multipliers are introduced during the training of the sequence graph attention network. The constrained Markov decision process model described above is transformed into an unconstrained dual problem, with the corresponding Lagrangian function set as follows: : ； in, These are Lagrange multipliers, representing the dynamic price of risk; For policy network parameters; In strategy The following expectations; A discount factor for future returns; for The cost of spillover risks at any given moment; For traffic efficiency; The time-series graph attention network comprises a multi-head graph attention network and a gated recurrent unit. The multi-head graph attention network captures the attention weights between intersection nodes in the road network topology, while the gated recurrent unit propagates states along the time axis and learns the temporal dynamics of congestion wave propagation. The multi-head graph attention network captures attention weights from intersection nodes... To the intersection node attention weights for: ； in, and It is node embedding. and It is a learnable parameter matrix. It is the coefficient of the time penalty term; It is a safety buffer threshold; for Transpose of; This represents an activation function with a leakage slope; It is the intersection node Relevant timestamps / update times; This represents a vector concatenation operation; Indicates direct proportion; During the training of the temporal graph attention network, the proximal policy optimization algorithm is used to update the policy network parameters. Its alternative objective function also includes the advantage of return. and cost advantage : ； Using Lagrange multipliers The algorithm is updated using a subgradient ascent method, enabling it to dynamically adjust the penalty for risk. ； in It's the learning rate; when actual risk exceeds the budget, Increase the punishment; conversely, decrease it. Optimize the objective function for the PPO strategy; This indicates a batch of time steps sampled. Take the average; It represents the probability ratio; The shear threshold of PPO; The updated Lagrange multipliers will be used for the next round of training / optimization; For the first Lagrange multipliers in the next iteration; In the current strategy Below, the price Expectations; Cost-based upper limit / risk budget; It is a data processing function used to limit numerical values ​​to a specified range; values ​​that exceed the range will be truncated to boundary values. It takes the minimum value; It takes the maximum value.

2. The traffic signal control method based on physical prior constraints and time-series graph attention as described in claim 1, characterized in that, The status information of the intersection node is characterized by the intersection node status information, lane occupancy, queue length, average speed, signal phase, and overflow safety buffer.

3. The traffic signal control method based on physical prior constraints and time-series graph attention as described in claim 1, characterized in that, The calculation process for the overflow safety buffer is as follows: ； ； in, It is a section of road exist A constant overflow safety buffer; It is a section of road The effective length, It is the upper limit of the blocking density. It refers to the number of lanes; For road section exist Number of vehicles at any given time; It refers to storage capacity, i.e., the maximum number of vehicles that can be accommodated.

4. A traffic signal control system based on physical prior constraints and time-series graph attention, characterized in that, The method for traffic signal control based on physical prior constraints and time-series graph attention, as described in any one of claims 1-3, includes: The overflow safety buffer calculation module is used to calculate the physical storage capacity of a road segment based on the traffic flow model based on inventory conservation and real-time traffic data, and then combine it with the current number of vehicles to obtain the remaining storage space of the downstream road segment, i.e. the overflow safety buffer. The decision process model construction module is used to transform overflow safety buffer into overflow risk cost based on preset safety criteria, and construct a constrained Markov decision process model with traffic efficiency as the reward and overflow risk as the cost; wherein, the state of the constrained Markov decision process model is the state information of the intersection node obtained based on the road network topology and real-time traffic data; the action is to select the next signal phase; The optimal signal control action determination module is used to solve the constrained Markov decision process model using a pre-trained temporal graph attention network to obtain the optimal signal control action that balances traffic efficiency and spillover risk. During the training process, the temporal graph attention network adopts the Lagrange dual method optimization algorithm to adaptively adjust the penalty weights for spillover risk costs.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the physical prior constraint timing graph attention traffic signal control method as described in any one of claims 1-3.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the physical prior constraint timing graph attention traffic signal control method as described in any one of claims 1-3.

Images