Smart City Emergency Management Methods and Systems Based on Digital Twins

Abstract

This application provides a smart city emergency management method and system based on digital twins, relating to the technical field of smart transportation emergency response. The method includes: generating vehicle state vectors by fusing real-time traffic flow speed, driving intention data, and corresponding historical vehicle behavior data collected through a memory-enhanced network; then inputting these vectors along with traffic state data into a city traffic digital twin model; subsequently, using a graph neural network to perform twin inference within the model to obtain an initial prediction map; then constructing an influence propagation model based on the embedded road network topology, and performing propagation simulation and correction using node state deviations as perturbations to obtain a target prediction map; finally, combining road capacity and employing an evolutionary game mechanism to generate emergency strategies. This application enables intelligent and precise traffic emergency response.

Description

Technical Field

[0001] This application relates to the technical field of intelligent transportation emergency response, and in particular to a smart city emergency management method and system based on digital twins. Background Technology

[0002] In the construction of smart cities, creating virtual models of urban transportation systems to achieve real-time state mapping and simulation analysis can improve the resilience and safety of urban operations, and has broad application prospects in scenarios such as traffic management and accident rescue.

[0003] Currently, existing digital twin emergency management methods typically rely on road sensors and monitoring equipment to collect macroscopic traffic flow data in order to construct a digital representation of the urban road network. These methods reflect the overall traffic situation by integrating historical and real-time information and formulate emergency plans based on preset rules or macroscopic statistical predictions to achieve an initial response to congestion or accidents.

[0004] However, most of these methods focus on macro-level traffic flow analysis and lack a detailed capture of individual vehicle behavior intentions. They also fail to effectively depict the dynamic transmission process of traffic state changes in the road network. This makes emergency decisions often lag behind the actual situation and adjustments lack specificity, making it difficult to cope with complex and ever-changing urban traffic scenarios. Summary of the Invention

[0005] The purpose of this application is to provide a smart city emergency management method and system based on digital twins, in order to solve the problems of low intelligence and poor accuracy in emergency management response in existing technologies.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a smart city emergency management method based on digital twins, comprising:

[0007] Acquire traffic status data and road capacity of urban roads, as well as real-time traffic speed and driving intention data of each vehicle;

[0008] Based on the real-time traffic flow speed, the driving intention data, and the historical behavior data of the corresponding vehicles, a memory-enhanced network is used to perform state backtracking and intention prediction to generate a vehicle state vector representing the future behavior of each vehicle.

[0009] The vehicle state vector is used as a micro-behavioral input and injected into a pre-built urban traffic digital twin model. At the same time, the traffic state data is used as a macro-system state to drive the operation of the urban traffic digital twin model.

[0010] In the digital twin model of urban traffic, a graph neural network is used to perform twin inference processing on the traffic state data and the vehicle state vector to obtain an initial prediction map of the traffic state of urban roads in the future.

[0011] Based on the road network topology and traffic flow patterns embedded in the urban traffic digital twin model, an impact propagation model based on a spatiotemporal graph neural network is constructed.

[0012] Using the road network topology as the transmission path and the node state deviation in the initial prediction map as the disturbance amount, the influence propagation model is used to simulate the propagation process of traffic disturbances along the transmission path to adjacent nodes, outputting a correction amount, and correcting the initial prediction map based on the correction amount to obtain the target prediction map.

[0013] Based on the target prediction map and the road capacity, an evolutionary game mechanism is used to solve the strategy and generate an emergency strategy.

[0014] Optionally, in the urban traffic digital twin model, a graph neural network is used to perform twin inference processing on the traffic state data and the vehicle state vector to obtain an initial prediction map of the urban road traffic state for a future period of time, including:

[0015] Based on the traffic state data and the vehicle state vector, a traffic topology map is constructed with road segments as environment nodes, vehicles as intelligent agent nodes, and traffic lights as control nodes.

[0016] The candidate control strategies associated with the control nodes and the future behavior strategies associated with the agent nodes are injected into the traffic topology as policy features. Through the policy and situation coupling layer in the graph neural network, the diffusion effect of the candidate control strategies along the road network connection relationship to the environmental nodes and agent nodes is simulated. At the same time, the travel demand implied by the future behavior strategies is integrated to generate a dynamic policy influence situation field that comprehensively reflects the control intention and travel demand, which serves as the macro pressure field in the cross-scale causal feedback constraint mechanism.

[0017] The influence of the strategy on the situation field is used as edge weights to dynamically reconstruct the strength of the interaction relationship between nodes in the traffic topology graph.

[0018] Through the multi-role interaction evolution layer in the graph neural network, based on the strength of the reconstructed interaction relationship, the accessibility of environmental nodes, the decision preferences of agent nodes, and the policy utility of control nodes are co-evolved and their states are updated to generate a spatiotemporal topology graph. Among them, the macroscopic pressure field is dynamically mapped as a moderating factor affecting the decision preferences of agent nodes, and the statistical behavior pattern of the evolved agent nodes is used to correct the macroscopic pressure field.

[0019] Based on the aforementioned spatiotemporal topology, attention-guided propagation simulation is used to model the game and state adjustment process between agent nodes and control nodes under preset emergency constraints, generating a simulation diagram.

[0020] Based on the projection diagram, the state of road network nodes over a future period is reconstructed through the decoder layer of the graph neural network to generate an initial prediction diagram.

[0021] Optionally, the step of simulating the game and state adjustment process between agent nodes and control nodes under preset emergency constraints based on the spatiotemporal topology graph through attention-guided propagation deduction, and generating a deduction graph, includes:

[0022] Based on the spatiotemporal topology graph, the association strength between the agent node and the current control node and the downstream control node is calculated through the attention matching mechanism to obtain the node association degree.

[0023] The node correlation degree and the spatiotemporal topology graph are input into the attention diffusion network. The correlation degree is then propagated and superimposed along the road network topology through the attention diffusion network to obtain the propagation field.

[0024] Combining the propagation field with preset emergency constraints, the strategy parameters of the control nodes are optimized and adjusted through dynamic strategies, and the expected paths of the agent nodes are replanned to generate a projection diagram.

[0025] Optionally, the step of generating emergency strategies by using an evolutionary game mechanism to solve strategies based on the target prediction map and the road capacity includes:

[0026] Potential traffic conflict zones and critical rescue routes that need to be ensured to pass through are identified from the target prediction map;

[0027] Based on the identification results, the road capacity, and the traffic state represented by the target prediction map, the target prediction map is processed by a game theory model construction method to construct a situation field representing traffic conflicts and traffic priorities, as well as a transmission field representing changes in road network state.

[0028] By integrating the situation field and the transmission field, a game field that integrates traffic situation and propagation laws is generated;

[0029] The various preset traffic light control schemes and path guidance schemes are combined and defined as the hybrid strategy space to be optimized;

[0030] Under the constraints of the game field, evolutionary game optimization is performed on the strategies in the mixed strategy space through multiple rounds of iteration. In each round of optimization, the fitness of the corresponding strategy is updated according to the simulation effect of the strategy in the game field to obtain the strategy ratio. When the strategy ratio reaches the preset equilibrium condition, the emergency strategy is obtained.

[0031] Optionally, under the constraints of the game field, multi-round iterative evolutionary game optimization is performed on the strategies in the mixed strategy space. Each round of optimization updates the corresponding fitness based on the simulated performance of the strategy in the game field to obtain the strategy proportion, including:

[0032] The strategies in the hybrid strategy space are allocated according to a preset initial ratio to form an initial strategy population;

[0033] Under the constraints of the game field, the fitness of the initial strategy population is evaluated by a scenario-based performance simulation and comprehensive evaluator to obtain the fitness distribution. The evaluation process of the scenario-based performance simulation and comprehensive evaluator includes: inputting the strategy into the game field to simulate traffic conditions, and extracting the conflict resolution amount of the strategy on potential conflict sections, the efficiency improvement value of key rescue channels, and the resulting road network load cost. The basic fitness of each strategy is calculated by weighting according to preset weights. The basic fitness is corrected by combining capacity penalty and inter-strategy competition adjustment to obtain the final fitness distribution.

[0034] Based on the fitness distribution, the proportion of each strategy in the initial strategy population is dynamically updated through the replicator dynamic equation to obtain the proportion change vector;

[0035] The proportional change vector and the game field are input into the equilibrium solver to obtain the strategy proportion.

[0036] Secondly, this application provides a smart city emergency management system based on digital twins, comprising:

[0037] The acquisition module is used to acquire traffic status data and road capacity of urban roads, as well as real-time traffic speed and driving intention data of each vehicle.

[0038] The prediction module is used to generate vehicle state vectors representing the future behavior of each vehicle by performing state backtracking and intention prediction through a memory augmentation network based on the real-time traffic flow speed, the driving intention data, and the historical behavior data of the corresponding vehicles.

[0039] The driving module is used to inject the vehicle state vector as a micro-behavior input into the pre-built urban traffic digital twin model, and at the same time, use the traffic state data as a macro-system state to drive the operation of the urban traffic digital twin model.

[0040] The extrapolation module is used to perform twin extrapolation processing on the traffic state data and the vehicle state vector in the urban traffic digital twin model using a graph neural network to obtain an initial prediction map of the traffic state of urban roads for a period of time in the future.

[0041] The construction module is used to build an impact propagation model based on a spatiotemporal graph neural network, based on the road network topology and traffic flow patterns embedded in the urban traffic digital twin model.

[0042] The correction module is used to simulate the propagation process of traffic disturbances along the propagation path to adjacent nodes through the influence propagation model, using the road network topology as the propagation path and the node state deviation in the initial prediction map as the disturbance amount, outputting a correction amount, and correcting the initial prediction map according to the correction amount to obtain the target prediction map.

[0043] The solution module is used to perform strategy solving based on the target prediction map and the road capacity, and generate emergency strategies using an evolutionary game mechanism.

[0044] Thirdly, this application provides an electronic device, comprising:

[0045] Memory, used to store computer programs;

[0046] A processor is configured to execute the computer program to implement the steps of the digital twin-based smart city emergency management method described in the first aspect above.

[0047] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the digital twin-based smart city emergency management method described in the first aspect above.

[0048] The digital twin-based smart city emergency management method provided in this application has the following beneficial effects: This application provides a comprehensive and real-time information foundation for emergency management by acquiring traffic status, traffic capacity, and real-time speed and intention data of vehicles; then, by fusing real-time and historical data through a memory-enhanced network, it generates a fine vector representing the future behavior of vehicles through state backtracking and intention prediction, which can improve the accuracy of micro-behavior prediction; next, the vehicle state vector is injected into the digital twin model as a micro-input, while the model is driven by macro-traffic status data, which can achieve deep fusion and collaborative simulation of macro-situation and micro-behavior; then, a graph neural network is used in the model to perform twin inference processing on the fused data, which can generate an initial prediction map of traffic status that is closer to the real dynamics in the future.

[0049] Subsequently, an impact propagation model is constructed based on the road network topology and flow patterns embedded in the model to simulate the dynamic diffusion process of traffic disturbances on multi-hop paths. Then, using this model, the state deviation propagation simulation and correction of the initial prediction map are performed, and the resulting target prediction map improves the reliability and anti-interference capability of the prediction. Finally, based on the target prediction map and traffic capacity, and using an evolutionary game mechanism to solve the strategy, the generated emergency strategy can dynamically weigh multiple factors, thereby achieving a more accurate and adaptive emergency management response.

[0050] Furthermore, in the digital twin model, this application constructs a traffic topology map that integrates the environment, vehicles, and control nodes, and injects policy features to generate a dynamic policy influence situation field, thereby dynamically reconstructing the node interaction relationships. Then, based on the reconstructed relationships, the node states undergo collaborative evolution, and the game and adjustment process is simulated through attention-guided propagation deduction, ultimately reconstructing the initial prediction map. This process achieves a comprehensive simulation of control intentions, travel demands, and dynamic interactions among multiple roles, making the initial prediction map more reflective of the real evolutionary trends under complex systems, thus providing a high-quality input foundation for subsequent correction and decision-making. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart illustrating a smart city emergency management method based on digital twins, provided for an embodiment of this application;

[0053] Figure 2 A schematic diagram illustrating a specific implementation of a smart city emergency management method based on digital twins, provided in this application embodiment;

[0054] Figure 3 This is a schematic diagram of the structure of a smart city emergency management system based on digital twins, provided as an embodiment of this application.

[0055] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0056] The core reasons for the lack of intelligence and precision in existing emergency management methods are as follows: on the one hand, these methods rely heavily on macro-level statistics, making it difficult to capture and understand the behavioral intentions and future trends of individual vehicles in detail; on the other hand, there is a lack of effective means to simulate how congestion or accidents at an intersection dynamically affect and spread to the surrounding road network. This results in emergency measures that are often out of sync with actual developments and have limited effectiveness.

[0057] To address this, this application proposes a smart city emergency management method based on digital twins. The basic concept is as follows: First, predict the microscopic future behavior of vehicles by combining real-time and historical data. Next, input these detailed microscopic predictions along with macroscopic traffic conditions into the digital twin system, and use network relationship analysis technology to deduce a preliminary map of future traffic conditions. Then, construct a dedicated analysis model to simulate the propagation process of abnormal traffic conditions along the road network and correct the preliminary predictions, thereby obtaining a more reliable future situation prediction map. Finally, based on this corrected prediction map and road capacity, generate the final emergency control strategy through a dynamic game optimization method. This solution, by integrating microscopic behavior prediction and macroscopic disturbance propagation analysis, makes the assessment of traffic conditions more forward-looking and accurate, thereby supporting the generation of more timely and precise emergency strategies, effectively solving the shortcomings of existing methods in terms of response intelligence and accuracy.

[0058] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] The core of this application is to provide a smart city emergency management method based on digital twins, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0060] S101. Obtain traffic status data and road capacity of urban roads, as well as real-time traffic speed and driving intention data of each vehicle.

[0061] Traffic status data includes sudden abnormal events, abnormal location data, and historical traffic data.

[0062] In step S101, in an emergency scenario of sudden traffic congestion on urban roads, traffic status data covering the road network is first collected and aggregated in real time from sensors, surveillance cameras, and floating cars deployed on the roads to grasp the macro-level congestion situation and the location of the event. At the same time, road capacity data of relevant roads in the affected area are extracted from the traffic infrastructure database to serve as a benchmark for subsequent assessment of the road network's carrying capacity. Then, in order to implement precise traffic management, this application also obtains the real-time traffic speed of each vehicle in the event-affected area through vehicle-to-everything (V2X) or mobile signaling to characterize its instantaneous movement state.

[0063] In addition, by analyzing the signals emitted by the vehicles or the interaction information with the navigation platform, the driving intention data of these vehicles can be obtained synchronously to infer their possible route selection or driving operations, thereby providing a basis for micro-behavioral analysis.

[0064] S102. Based on the real-time traffic flow speed, the driving intention data, and the historical behavior data of the corresponding vehicles, a memory-enhanced network is used to perform state backtracking and intention prediction to generate a vehicle state vector representing the future behavior of each vehicle.

[0065] Among them, memory-enhanced network refers to an intelligent computing model based on neural network, which is specially designed to store and retrieve historical experience information. In this application, it may contain two core units: a memory retrieval unit for quickly finding relevant historical scenarios from the memory bank; and a reasoning unit for logical deduction based on the retrieved historical information.

[0066] In one specific implementation, step S102 includes:

[0067] Step 1021: Perform data fusion and data completion processing on the real-time traffic flow speed and the driving intention to obtain spatiotemporal correlation data.

[0068] Spatiotemporal correlation data refers to an integrated dataset formed by aligning, splicing, and completing real-time traffic speed and driving intention data according to time sequence and spatial location. It uniformly describes "when, where, at what speed, and what the vehicle intends to do".

[0069] In step 1021, the real-time traffic flow speed and driving intention data from the same vehicle are first fused. This process aligns the two types of data with the vehicle position at the same time point and fills in any possible transient signal gaps by using data from nearby times, thereby generating spatiotemporal correlation data that describes the vehicle's current complete dynamics.

[0070] Step 1022: Input the spatiotemporal correlation data and the corresponding vehicle's historical behavior data into the memory enhancement network, and perform matching and retrieval through the memory retrieval unit of the memory enhancement network to obtain the memory index.

[0071] In step 1022, the spatiotemporal correlation data and the vehicle's historical behavior data over a period of time, such as speed and lane-changing habits when passing through the same road segment in the past, are input into the memory enhancement network. Then, the network's memory retrieval unit quickly matches and compares the current situation with a large number of behavioral patterns in the historical database to find the most similar historical scenarios and output memory indexes pointing to these scenarios.

[0072] Furthermore, the embodiments of this application do not limit the structural design, parameter design, and training process of the memory enhancement network, and can be set accordingly based on the actual situation.

[0073] Step 1023: Based on the memory index, memory playback and state deduction are performed through the inference unit of the memory enhancement network to obtain risk prediction data for each vehicle.

[0074] In step 1023, the inference unit of the memory-enhanced network "replays" the subsequent behavior sequence of the vehicle in similar scenarios from historical memory based on these memory indexes. Then, by performing trend analysis and state deduction on these historical behavior patterns, the unit predicts various risk prediction data that the current vehicle may encounter in the near future, such as the probability of rapid acceleration, abnormal lane change, or driving into a congested area.

[0075] Step 1024: Using a risk diffusion model based on cross-attention and the risk prediction data, calculate the attention weights among vehicles in the current urban road network to obtain the risk impact value of each vehicle.

[0076] In step 1024, the overall framework of the risk diffusion model can be a graph neural network, a cellular automata, or a risk diffusion model based on information entropy. The risk diffusion model based on cross-attention is introduced to quantify the mutual influence between each pair of vehicles by evaluating the spatial topological relationship, motion state correlation and risk transmission intensity between vehicles. For example, the risk of a vehicle braking suddenly in front will generate an "attention weight" for the vehicle immediately behind it. The risk impact value borne by a specific vehicle is calculated by summarizing the weighted influence of all other vehicles on a specific vehicle.

[0077] Step 1025: Integrate the risk prediction data and the risk impact value to generate a vehicle state vector for each vehicle.

[0078] In step 1025, the risk prediction data derived by each vehicle itself is fused with the risk impact value generated by environmental interaction. This fusion process is not a simple addition, but rather a weighting and integration of the two types of information according to preset rules, ultimately generating a unified vehicle state vector for each vehicle that includes its individual predicted behavior and its interaction with the environment. This embodiment does not limit the content of the preset rules.

[0079] For example, the risk prediction data for a freight vehicle traveling on an urban expressway may include two main dimensions: "low probability of frequent lane changes" and "high risk of following too closely". At the same time, based on the risk prediction assumptions of surrounding vehicles, the risk impact value borne by the freight vehicle is calculated to be 0.6, and the preset rules are set as follows: for freight vehicles, the weight of their own driving risk prediction data is 0.7, and the weight of the risk impact value affected by the environment is 0.3; while for passenger cars, the weights may be set to 0.5 and 0.5 respectively.

[0080] This application correlates the real-time dynamics, historical habits, and risks of surrounding traffic flow of vehicles by simulating the cognitive processes of human "experience review" and "environmental impact assessment." This enables a more accurate and in-depth prediction of the future micro-behavior and inherent risks of each vehicle. The resulting vehicle state vector provides crucial and reliable micro-input for the subsequent construction of high-fidelity traffic digital twins and the formulation of precise emergency diversion strategies.

[0081] S103. The vehicle state vector is used as a micro-behavioral input and injected into the pre-built urban traffic digital twin model. At the same time, the traffic state data is used as a macro-system state to drive the operation of the urban traffic digital twin model.

[0082] Among them, the urban traffic digital twin model refers to a virtual simulation system built in a computer that forms a high-fidelity mirror image of the real urban road network in the physical world. The model includes digital copies of static infrastructure such as roads, intersections, and traffic lights. Its core is a computing engine trained with historical traffic big data that can simulate vehicle interaction, signal control logic, and macroscopic traffic flow evolution.

[0083] In step S103, firstly, the vehicle state vector generated for each vehicle is associated with and imported into the corresponding virtual vehicle agent in the urban traffic digital twin model based on the vehicle identification code or real-time location information. This process completes the "assignment" of the future intentions and potential risks of all traffic individuals in the model, that is, the injection of micro-behavioral input is completed. Then, the model learns the statistical laws of driver behavior under different vehicle state vectors by analyzing massive historical trajectory data, thereby simulating the real reaction of individuals.

[0084] At the same time, the acquired traffic status data reflecting the overall real-time condition of the road network is loaded into the system parameter layer of the digital twin model. This data sets macroscopic initial conditions and operating baselines for the model that are synchronized with the real world, thereby driving the entire virtual traffic system to start simulating according to the learned rules. Furthermore, the model's macroscopic dynamic module has been pre-trained by integrating historical traffic patterns and real-time data, enabling it to accurately respond to such macroscopic state inputs.

[0085] This application activates the urban traffic digital twin model by combining "micro-injection" and "macro-driven" approaches. When the model is started, it can not only restore the overall real-time status of the road network based on macro data, but also endow each vehicle intelligent agent with predictive behavioral logic at the micro level. This lays a key foundation for subsequent high-precision and realistic traffic evolution simulations within the model.

[0086] S104. In the urban traffic digital twin model, a graph neural network is used to perform twin inference processing on the traffic state data and the vehicle state vector to obtain an initial prediction map of the traffic state of urban roads in the future.

[0087] It should be noted that the structural design of the graph neural network, the design of each layer within the network, the parameter design, and the training process can all be set according to the actual situation, and will not be elaborated here.

[0088] In one specific implementation, such as Figure 2 As shown, step S104 includes:

[0089] Step 1041: Based on the traffic state data and the vehicle state vector, construct a traffic topology map with road segments as environment nodes, vehicles as intelligent agent nodes, and traffic lights as control nodes.

[0090] In step 1041, firstly, the road network is discretized into a series of interconnected road segments based on the city's electronic map, and each road segment is created as an environment node, and its real-time traffic status data and static traffic capacity are assigned to the node; at the same time, an intelligent agent node is created for each vehicle, and its initial attributes are set to its corresponding vehicle state vector, which encodes the vehicle's predicted behavior and risks, and the spatial location of the node is determined by its current road segment and linked to the corresponding environment node.

[0091] Finally, all traffic signal control nodes in the road network are instantiated, and directed connections between all environmental nodes are established based on the actual physical connections of the road network. Then, based on the real-time location of the vehicle, the affiliation edges between the agent node and its environmental node, as well as the control relationship edges between the control node and the environmental nodes it manages, are established, thereby completing the construction of the traffic topology map.

[0092] For example, assuming a traffic accident occurs at intersection X in city A, firstly, the affected intersection X and the surrounding road segments Y and Z are created as environment nodes Ex, Ey, and Ez, respectively, and their current congestion traffic status data is loaded. Then, intelligent agent nodes Ab1 and Ab2 are created for vehicles B1 and B2 traveling on road segment Y, with their node attributes being their respective vehicle state vectors. Finally, traffic lights S1 and S2 controlling this area are created as corresponding control nodes Cs1 and Cs2, and directed edges are established between Ey, Ex, and Ez based on map connectivity. Ab1 and Ab2 are associated with Ey, Cs1 with Ex, and Cs2 with Ez, thus forming a traffic topology map describing this local emergency scenario.

[0093] Step 1042: Inject the candidate control strategies associated with the control nodes and the future behavior strategies associated with the agent nodes as policy features into the traffic topology graph. Through the policy and situation coupling layer in the graph neural network, simulate the diffusion effect of the candidate control strategies along the road network connection relationship to the environment nodes and agent nodes. At the same time, integrate the travel demand implied by the future behavior strategies to generate a dynamic policy influence situation field that comprehensively reflects the control intention and travel demand, as a macro pressure field in the cross-scale causal feedback constraint mechanism.

[0094] Among them, candidate control strategies refer to multiple alternative control schemes pre-set for traffic light control nodes to cope with the current traffic conditions, such as "extending the green light time for east-west traffic"; future behavior strategies refer to the expected driving path selection or driving behavior mode of the vehicle decoded from the vehicle state vector of the intelligent agent node.

[0095] Furthermore, the policy and situation coupling layer can include three feedforward neural networks and a multi-head graph attention network, where each feedforward neural network encodes candidate control policies, future behavior policies, and the real-time traffic status and static attributes of environmental nodes, respectively. Then, information propagation is achieved through two parallel attention diffusion sub-layers in the multi-head graph attention network. Each environmental node fuses the received policy influence features and demand influence features with its own environmental features. The fusion process can be achieved through concatenation, weighted summation, or gating mechanisms to generate node-level fused features.

[0096] The fused features of each environmental node are mapped to a scalar pressure value through a multilayer perceptron. The set of pressure values ​​of all environmental nodes constitutes the dynamic policy influence situation field. This situation field reflects the expected congestion pressure or traffic attraction of each area of ​​the road network under the combined effect of candidate control strategies and predicted travel demand. Based on the generated situation field, the weights of the connecting edges between environmental nodes in the traffic topology are dynamically reconstructed. The adjustment of edge weights takes into account the pressure values ​​of connected nodes and the inherent attributes of road segments, so that the pressure propagation path is explicitly represented in the network structure, providing causal constraints for subsequent traffic flow inference.

[0097] In step 1042, firstly, a set of candidate control policies is loaded for each control node as an extended feature of that node; simultaneously, the future behavior policy is parsed from the vehicle state vector of each agent node, which is also used as a feature of that node; then, through the policy and situation coupling layer in the graph neural network, the control node diffuses its policy features along the control relationship edge to the environment node to simulate the issuance of control commands; at the same time, the future travel needs of the agent node are also projected along the association edge between it and the environment node to simulate the vehicle's intention to choose a path.

[0098] Then, when an environmental node receives control signals from multiple control nodes and path selection intentions from numerous agent nodes, the core aggregation function of this layer performs weighted fusion and superposition calculations on these heterogeneous information. Finally, through the calculation of this layer, a value representing the expected pressure or attraction is generated for each environmental node in the graph, and all these values ​​are set together to form a dynamic policy influence situation field covering the entire domain.

[0099] For example, in the traffic topology map of an accident scenario, the candidate strategy of "extending the north-south green light" is injected into the control node Cs1, and the future behavior strategy of "planning to turn left at the next intersection" is parsed from the vehicle state vectors of agent nodes Ab1 and Ab2. Then, the strategy and situation coupling layer simulates the diffusion of these strategies. For example, Cs1's strategy of extending the green light will improve the expected passage of the Ex node it controls; the left-turning intentions of Ab1 and Ab2 will be projected onto their target road segment Ez to express travel demand. After comprehensive calculation, this layer may conclude that the Ex node is still under high pressure due to the accident, while the Ez node is also under increasing pressure due to the increased demand for detours, thus generating a macroscopic pressure field that reflects the situation.

[0100] Step 1043: Using the situation field affected by the strategy as edge weights, dynamically reconstruct the strength of the interaction relationship between nodes in the traffic topology graph.

[0101] In step 1043, based on the generated macro-level situation, the strength of connections in the traffic topology graph is dynamically adjusted. In the initial traffic topology graph, the edges between nodes only represent physical connections, and the policy influence situation field reveals the expected pressure distribution of different road segments. At this time, the weights of the edges in the graph are dynamically updated according to the value of the policy influence situation field.

[0102] Specifically, for an edge connecting two environmental nodes, its new weights are reconstructed based on the pressure values ​​of the two endpoints. A common reconstruction method is to correlate the weights with the pressure gradient and simulate the tendency of traffic flow to flow from high-pressure areas to low-pressure areas. For example, the new edge weights can be proportional to the sum of the pressures of the two endpoints, mapped using a Sigmoid function. This transforms the abstract macroscopic pressure into concrete connection strength, enabling subsequent evolutionary deductions to respond more sensitively to real-time situational changes. The Sigmoid function can be: ,in, This refers to the edge weights from node i to node j after reconstruction. For the Sigmoid function, and Let i represent the macroscopic pressure values ​​for environmental nodes i and j. For pressure threshold parameters, such as 1 is given, and e is the natural constant.

[0103] Step 1044: Through the multi-role interaction evolution layer in the graph neural network, based on the strength of the reconstructed interaction relationship, the accessibility of environmental nodes, the decision preferences of agent nodes, and the policy utility of control nodes are co-evolved and their states are updated to generate a spatiotemporal topology graph; wherein, the macroscopic pressure field is dynamically mapped as a moderating factor affecting the decision preferences of agent nodes, and the statistical behavior pattern of the evolved agent nodes is used to correct the macroscopic pressure field.

[0104] The structure of the multi-role interaction evolution layer can be a graph messaging network. The structural design of the graph messaging network can refer to relevant technologies, and the structural design of the network will not be described in detail in this embodiment.

[0105] In step 1044, the multi-role interaction evolution layer in the graph neural network performs graph convolution and other operations based on the dynamically reconstructed edge weights to collaboratively update the states of the three types of nodes: the environment node receives messages from connected nodes and control nodes and updates its "effective passage capability" state; the agent node receives pressure information from its own and target environment nodes and dynamically maps the macroscopic pressure field into factors that adjust its decision preferences; the control node receives state feedback from the environment nodes it manages and updates its "utility evaluation" of different candidate control strategies.

[0106] This evolutionary process is then iterated multiple times. Crucially, the updated decision preferences of numerous agent nodes after evolution form a new collective behavior pattern. This new pattern serves as a feedback signal to correct the initial macroscopic pressure field, achieving cross-scale causal feedback between individual behavior and the macroscopic situation. Iteration stops when any preset convergence condition is met, ultimately yielding a spatiotemporal topology graph containing the latest node states. The convergence condition can be that the average change in the state vectors of all environmental nodes, such as effective passage capacity, over two consecutive iterations is less than a set minimum threshold. At times, such as The value can be 0.0001, or it can be the value that terminates the iteration when the number of iterations reaches a preset number of rounds, such as 50 rounds.

[0107] Step 1045: Based on the spatiotemporal topology graph, through attention-guided propagation deduction, simulate the game and state adjustment process between agent nodes and control nodes under preset emergency constraints, and generate a deduction graph.

[0108] Step 1045 may specifically include the following steps:

[0109] Step a1: Based on the spatiotemporal topology graph, calculate the association strength between the agent node and the current control node and the downstream control node through the attention matching mechanism to obtain the node association degree.

[0110] In step a1, based on the state characteristics of nodes in the spatiotemporal topology graph, an attention matching mechanism is initiated. This mechanism uses the characteristics of each agent node, such as its location and future behavior strategy, as the "query" and the characteristics of each control node, such as its control range and candidate strategy, as the "key," to calculate the similarity between them. Specifically, it calculates the correlation between the agent node and the control node at its current intersection, which depends on whether the vehicle is in the traffic flow affected by the traffic light. At the same time, it also calculates the correlation between the agent node and the control node at the downstream intersection, which depends on whether its future behavior strategy plans to drive to and pass through that intersection, and uses this similarity as the node correlation.

[0111] Step a2: Input the node correlation degree and the spatiotemporal topology graph into the attention diffusion network, and perform the transmission and superposition process of the correlation degree along the road network topology through the attention diffusion network to obtain the propagation field.

[0112] In step a2, the node associativity is compared with the spatiotemporal topology containing all connections. Figure 1The input attention diffusion network is used to simulate the multi-hop propagation and superposition of node correlation along the connection edges between environmental nodes. For example, the correlation between a control node and vehicles on an upstream road segment will propagate along the road network topology to adjacent downstream road segments, thus affecting the assessment of the indirect impact on vehicles on downstream road segments. This process is usually achieved through multi-layer graph convolution or random walk algorithms, which smooths and diffuses the initial point-to-point correlation into a continuous field, and finally outputs a propagation field. The propagation field provides an "influence map" for each control node and reveals the potential impact range and intensity distribution of its strategy changes.

[0113] Step a3: Combining the propagation field with the preset emergency constraints, optimize and adjust the strategy parameters of the control node through dynamic strategies, and replan the expected path of the agent node to generate a projection diagram.

[0114] In step a3, the propagation field and preset emergency constraints, such as "critical channel speed > 30km / h", are used as core inputs to initiate the dynamic strategy optimization process. This process is usually formalized as a constrained optimization problem: under the relationship model between strategy adjustment and network-wide impact described by the propagation field, a set of strategy parameters for control nodes, such as the green light duration of each phase, are found through the gradient descent algorithm to make the optimization objective optimal, such as minimizing the total network travel time, while strictly satisfying all emergency constraints.

[0115] Then, the state of all control nodes is updated according to the new strategy parameters obtained from the optimization; at the same time, the expected paths of all agent nodes are synchronously replanned based on the new signal control scheme and the latest road network prediction state; finally, the updated control node strategies and agent node paths are integrated into the graph to generate the final projection graph. It should be noted that the specific implementation process of the gradient descent algorithm to find strategy parameters can refer to relevant technologies, and will not be elaborated in the embodiments of this application.

[0116] Step 1046: Based on the projection map, reconstruct the state of road network nodes in the future through the decoder layer of the graph neural network to generate an initial prediction map.

[0117] In step 1046, the decoder layer of the graph neural network is a pre-trained network module whose function is to represent the nodes containing high-level strategies and plans in the projection graph and decode them back into specific and understandable sequences of future traffic states. In this embodiment, the type of decoder in the decoder layer is not limited.

[0118] Specifically, the decoder layer reads the final state vector of each environmental node in the projection map. This vector encodes the expected traffic load and control results of the road segment after all the aforementioned steps. Then, the decoder uses forward propagation to map this vector into a set of time series data, such as the predicted traffic flow and average speed for multiple future time slices. Finally, the predicted state sequences of all environmental nodes are integrated according to their geographic spatial location and output in the form of heatmaps or state matrices, thus generating the initial prediction map.

[0119] This application not only simulates the complex process of the interaction between control strategies and travel demand, but also introduces game optimization based on emergency objectives. This makes the generated initial prediction map not a simple trend extrapolation, but a high-fidelity inference result obtained by integrating multi-party dynamic game and feedback adjustment, thus providing a reliable and forward-looking basis for emergency decision-making.

[0120] S105. Based on the road network topology and traffic flow patterns embedded in the urban traffic digital twin model, construct an influence propagation model based on a spatiotemporal graph neural network.

[0121] In step S105, the embedded road network topology is first extracted from the urban traffic digital twin model, which constitutes the spatial skeleton of influence propagation, with road segments as nodes and connection relationships as edges. At the same time, the traffic flow patterns learned or built into the model are extracted, such as the relationship between speed and flow under different densities, and the turning probability at intersections. These patterns provide dynamic constraints for the transmission and transformation of the simulated influence on nodes and edges, and together define the "physical path" and "transmission rules" of disturbance propagation.

[0122] Based on this, a spatiotemporal graph neural network is designed and instantiated as the core computing engine of the model. The network architecture is designed to capture both spatial dependence and temporal evolution. Its spatial graph convolution module runs on the road network topology graph and simulates the lateral diffusion of disturbances in space by aggregating state information from adjacent nodes in the topology. Its temporal modeling module, such as the gated recurrent unit, is responsible for capturing the continuous changes in the state of each node over time to simulate the dynamic processes such as the continuation, decay, or enhancement of disturbances.

[0123] Then, the spatiotemporal graph neural network is trained through supervised learning. Its training data comes from a large number of historical or simulated traffic disturbance cases stored or generated by the urban traffic digital twin model. Each training sample contains an initial local disturbance state and its subsequent evolution over a period of time across the entire road network, representing a true state sequence. The training objective is to enable the network to learn a mapping function that minimizes the error between the predicted state sequence and the actual evolution sequence based on the initial input disturbance. The parameters in the network are continuously optimized through backpropagation, ultimately resulting in a trained impact propagation model that accurately fits the disturbance propagation effect of traffic flow patterns on a specific road network topology. The mapping function is not the focus of this application, and this embodiment does not limit the expression of this function.

[0124] This application enables accurate quantitative simulation of the dynamic diffusion process of abnormal traffic conditions, and can transform the state deviation of local nodes into a reliable prediction of the state changes of the entire network over a period of time. This provides key mechanistic basis and calculation methods for the scientific correction of the initial prediction map in subsequent steps, thereby improving the predictability of the emergency management system for complex chain effects.

[0125] S106. Using the road network topology as the transmission path and the node state deviation in the initial prediction map as the disturbance amount, the traffic disturbance is simulated to propagate to adjacent nodes along the transmission path using the influence propagation model. A correction amount is output, and the initial prediction map is corrected based on the correction amount to obtain the target prediction map.

[0126] In one specific implementation, step S106 includes:

[0127] Step 1061: Calculate the node state deviation of each node in the initial prediction graph.

[0128] In step 1061, for each future prediction time slice covered by the initial prediction map, the node state deviation of each environmental node in the map is calculated. Specifically, the predicted future state value Vpred of a node in the initial prediction map is compared with a predefined or statistically derived expected baseline value Vbase of that node at the same time. The node state deviation Δ is typically the difference between the two: Δ = Vpred Vbase, where a positive Δ indicates that the predicted traffic is higher than the historical normal, implying a potential risk of oversaturation; a negative Δ indicates that the predicted speed is lower than the ideal level, implying an unconsidered risk of congestion.

[0129] Step 1062: Based on the connection relationship of the road network topology, determine the transmission path between nodes, and calculate the sensitivity coefficient of each node to disturbance based on the node attributes.

[0130] The connection relationship of the road network topology can be simply referred to as the topological connection relationship, and the two have the same interpretation.

[0131] In step 1062, firstly, based on the road network topology extracted from the digital twin model, the connection relationship between all nodes is clarified. For example, if road segment F is connected to road segment G, it constitutes a propagation path from node Ef to node Eg. All such connection relationships together form a network describing how disturbances can propagate step by step.

[0132] Secondly, the sensitivity coefficient Si is calculated for each environmental node. This coefficient is typically based on the node's inherent properties and real-time state, and a simplified calculation formula is provided. We can consider the ratio of a node's remaining capacity to its current traffic, where qi is the current or predicted traffic of the i-th node and Ci is its capacity. When the traffic qi approaches the capacity Ci, its sensitivity coefficient Si increases, indicating that the node is on the verge of saturation and even a small upstream disturbance can cause its state to deteriorate; conversely, the sensitivity is low.

[0133] Step 1063: Using the node state deviation as traffic disturbance, and combining it with the sensitivity coefficient, simulate the multi-hop propagation of traffic disturbance along the transmission path through the spatiotemporal graph convolutional layer of the influence propagation model to generate the cumulative influence value of each node.

[0134] In step 1063, the node state deviation of each node is used as the initial traffic disturbance input. At the same time, the determined transmission path and the sensitivity coefficient of each node are used as constraints and modulation factors for model operation and input to the influence propagation model that has been constructed and trained in step S105. Then, the spatiotemporal graph convolutional layer of the model uses the graph defined by the aforementioned transmission path as its structure to perform multiple rounds of information transmission. In one round of transmission, the disturbance of a node not only affects its direct downstream neighbors, but also continues to spread downstream through its neighbors, thereby achieving multi-hop propagation.

[0135] Then, during propagation, when a disturbance Δj propagates from a node to its downstream node, its influence intensity is modulated by the node's sensitivity coefficient Si. For example, the cumulative disturbance signal propagating to the node is... ,in, Let j be the set of neighboring nodes of node i, and j represent an upstream node belonging to this set. The state deviation of upstream node j, These are the basic propagation weights from node j to node i that the propagation model learns from historical data.

[0136] Then, through the stacking and iteration of multiple spatiotemporal graph convolutional layers within the model, the initial local node state deviation is continuously diffused and superimposed along the propagation path, and amplified or attenuated by the sensitivity coefficients of the nodes along the way. The stacking and iteration terminate when the iteration convergence condition is reached. Finally, the model calculates a cumulative influence value I for each node. The convergence condition can be defined as: when all nodes that can be reached from the initial perturbation node along the propagation path within a finite number of hops, such as 3-5 hops, or when the increment of the cumulative influence value I received by them is lower than a pre-set small threshold after one forward propagation calculation.

[0137] Step 1064: Map the cumulative influence value to the correction amount of each node through the output layer of the influence propagation model.

[0138] In step 1064, the cumulative influence value I of each node is received through the output layer of the influence propagation model. The output layer uses the learned mapping function to convert I into a correction amount Ui for the original predicted state of node i. For example, the output layer can perform the following transformation: ,in, and These are the trainable weights and biases of the output layer, and have been learned during the model training phase through a large number of sample pairs of "initial bias and final actual error" to ensure that Ui can accurately compensate for the prediction error caused by the propagation of initial bias.

[0139] Step 1065: Correct the state of the corresponding node in the initial prediction map based on the correction amount to generate the target prediction map.

[0140] In step 1065, for each environmental node in the initial prediction map at each future time slice, the corresponding correction amount is applied to correct its predicted state value. This correction operation can be achieved through linear superposition, such as: Vtarget = Vpred + λ × In this context, Vpred is the original predicted value in the initial prediction graph, λ is an adjustable gain coefficient, usually 1, and Vtarget is the corrected state value. After performing this operation on all nodes at all prediction time slices, a completely new prediction graph that has compensated for potential network propagation effects is obtained, namely the target prediction graph.

[0141] This application effectively simulates and compensates for the chain diffusion effect of traffic disturbances in the spatiotemporal dimensions, so that the final target prediction map not only includes the prediction of direct impacts, but also the estimation of indirect and secondary impacts, thereby improving the completeness and reliability of the overall traffic state prediction.

[0142] S107. Based on the target prediction map and the road capacity, an evolutionary game mechanism is used to solve the strategy and generate an emergency strategy.

[0143] In one specific implementation, step S107 includes:

[0144] Step 1071: Identify potential traffic conflict sections and key rescue channels that need to be guaranteed to pass through from the target prediction map.

[0145] Among them, key rescue channels refer to a set of specific road sections that must be kept open in order to ensure that rescue forces can quickly reach the scene of the incident or evacuate people in emergency scenarios. They are usually determined in advance based on emergency plans and road network structure.

[0146] In step 1071, potential traffic conflict sections in the target prediction map are first analyzed. This is done by comparing the predicted traffic flow of each section in the target prediction map with its own road capacity in real time. When the ratio of the predicted flow to its capacity of a certain section, i.e., the saturation, continuously exceeds a preset conflict threshold, such as 90%, the section is marked as a potential traffic conflict section. At the same time, a series of road sections connecting the event point, emergency unit and evacuation area are automatically mapped and locked according to the preset emergency geographic information database and event location, and these road sections are defined as key rescue channels that need to be guaranteed.

[0147] Step 1072: Based on the identification results, the road capacity, and the traffic state represented by the target prediction map, the target prediction map is processed by a game theory model construction method to construct a situation field representing traffic conflicts and traffic priorities, as well as a transmission field representing changes in road network state.

[0148] In step 1072, based on the original data of the identification results and the target prediction map, a situation field representing traffic conflict and traffic priority is first constructed. Specifically, for each road segment in the road network, a negative value representing conflict pressure is assigned according to whether it is identified as a traffic conflict road segment and its predicted saturation. Then, a positive value representing guarantee priority is assigned according to whether it belongs to a critical rescue channel and its order in the rescue path. The two values ​​are then weighted and summed to obtain the situation value of the road segment in the situation field. The situation values ​​of all road segments in the entire road network constitute a spatial distribution field.

[0149] Secondly, a transmission field characterizing changes in road network status is constructed using traffic flow propagation models learned from historical data or derived from road capacity and topology. This field defines dynamic rules such as the proportion of traffic transfer from one road segment to its downstream segments and the scope of congestion backtracking. In essence, the transmission field is a kernel function or transition probability matrix that describes how status changes propagate in space and time.

[0150] Step 1073: Merge the situation field and the transmission field to generate a game field that integrates traffic situation and propagation laws.

[0151] In step 1073, the method of constructing the game field is to deeply couple the situation field and the transmission field, rather than simply superimposing them. The fusion operation is achieved by establishing a joint computing framework. This framework first embeds the conflict pressure and guarantee priority of each road segment in the situation field as basic attributes. At the same time, it embeds the dynamic rules defined in the transmission field, that is, the quantitative relationship of how the state change of a road segment triggers the chain reaction of its downstream road segments, as a causal relationship network.

[0152] Then, through this framework, the static value distribution of the situation field and the dynamic propagation law of the transmission field are organically integrated, thus jointly defining a new and complete computing environment, which is the game field. In the game field, the state of any node is simultaneously associated with its inherent value in the situation field, as well as the dynamic influence link established with other nodes through the transmission field.

[0153] Step 1074: Combine the preset traffic light control schemes and path guidance schemes to define a hybrid strategy space to be optimized.

[0154] In step 1074, all traffic light control schemes and path guidance schemes related to the current emergency scenario are selected from a preset strategy library, and these schemes are pre-designed feasible specific actions; then these different types of schemes are combined, such as a complete candidate strategy may be composed of controlling traffic lights and guiding paths P1 and P2, and all such possible and meaningful combinations constitute a mixed strategy space to be screened.

[0155] Step 1075: Under the constraints of the game field, perform multi-round iterative evolutionary game optimization on the strategies in the mixed strategy space. In each round of optimization, the fitness of the corresponding strategy is updated according to the simulation effect of the strategy in the game field to obtain the strategy ratio. When the strategy ratio reaches the preset equilibrium condition, the emergency strategy is obtained.

[0156] Step 1075 may specifically include the following steps:

[0157] Step b1: Allocate the strategies in the hybrid strategy space according to a preset initial ratio to form an initial strategy population.

[0158] In step b1, in the optimization for the accident at intersection A, strategy combinations S1 and T1 and strategy combinations S2 and T2 are selected from the mixed strategy space and allocated in an initial ratio of half each to form an initial strategy population containing these two strategies.

[0159] Step b2: Under the constraints of the game field, the fitness of the initial strategy population is evaluated by a scenario-based performance simulation and comprehensive evaluator to obtain the fitness distribution. The evaluation process of the scenario-based performance simulation and comprehensive evaluator includes: inputting the strategy into the game field to simulate traffic conditions, and extracting the conflict resolution amount of the strategy on potential conflict sections, the efficiency improvement value of key rescue channels, and the resulting road network load cost. The basic fitness of each strategy is calculated by weighting according to preset weights. The basic fitness is corrected by combining capacity penalty and inter-strategy competition adjustment to obtain the final fitness distribution.

[0160] In step b2, the comprehensive evaluator inputs each strategy into the game field to deduce the changes in traffic status after its execution and extracts key performance indicators: the amount of conflict resolution on conflict sections, the efficiency improvement value of rescue channels, and the resulting road network load cost. Subsequently, these indicators are weighted and summed according to the preset emergency target weights to calculate the basic fitness. Afterward, the evaluator will also apply two corrections: if the strategy causes the traffic flow on any section to exceed the road capacity, a capacity penalty will be applied to deduct points; at the same time, considering the competition and adjustment between strategies, the fitness of strategies that weaken each other will be fine-tuned. Finally, after the above simulation and calculation process, the comprehensive score of each strategy is output to form a fitness distribution.

[0161] Step b3: Based on the fitness distribution, dynamically update the proportion of each strategy in the initial strategy population using the reciprocator dynamic equation to obtain the proportion change vector.

[0162] In step b3, the policy proportions are updated according to the replicator dynamic equation. The core idea is that the proportion of policies with fitness above average in the current round will increase in the next round of the population, while the proportion of policies with fitness below average will decrease, thus obtaining the proportion change vector.

[0163] This embodiment does not limit the specific expression of the reproducer dynamic equation; it can be set according to the actual situation.

[0164] For example, since the fitness of strategy combination S2 and T2 is higher than the average fitness of the current population, their proportion in the next generation of strategy population will be increased, for example, from 50% to 70%, according to the replier dynamics equation, while the proportion of strategy combination S1 and T1 will be reduced accordingly. This process produces a proportion change vector indicating how the strategy proportions should change.

[0165] Step b4: Input the proportional change vector and the game field into the equilibrium solver to obtain the strategy proportion.

[0166] In step b4, the proportional change vector and the constraints of the game field are input into the equilibrium solver. The solver then determines whether the current proportional distribution has reached the preset equilibrium condition. The equilibrium condition usually means that the update change of the strategy proportion in the population approaches zero, that is, no strategy can obtain higher returns or reach the preset number of iterations by changing itself alone. If the equilibrium is not reached, the process returns to step b2 based on the new proportion to start the next round of evaluation and update. If the equilibrium is reached, the final strategy proportion represents the evolutionary stable state.

[0167] This application can not only effectively coordinate multiple means such as signal control and route guidance, but also automatically weigh multiple complex objectives such as local conflict resolution and global road network load, regular traffic and emergency rescue priority, and finally generate an emergency strategy with high comprehensive benefits and strong robustness in the dynamic traffic system, thereby improving the effectiveness of decision-making.

[0168] Figure 3 This application provides a schematic diagram of a specific implementation of a smart city emergency management system based on digital twins, referring to... Figure 3 The system may include:

[0169] The acquisition module 31 is used to acquire traffic status data and road capacity of urban roads, as well as real-time traffic speed and driving intention data of each vehicle.

[0170] The prediction module 32 is used to generate a vehicle state vector representing the future behavior of each vehicle by performing state backtracking and intention prediction through a memory augmentation network based on the real-time traffic flow speed, the driving intention data and the historical behavior data of the corresponding vehicles.

[0171] The driving module 33 is used to inject the vehicle state vector as a micro-behavior input into the pre-built urban traffic digital twin model, and at the same time, use the traffic state data as a macro-system state to drive the operation of the urban traffic digital twin model.

[0172] The inference module 34 is used to perform twin inference processing on the traffic state data and the vehicle state vector in the urban traffic digital twin model using a graph neural network to obtain an initial prediction map of the traffic state of urban roads in the future.

[0173] Module 35 is used to construct an influence propagation model based on a spatiotemporal graph neural network, based on the road network topology and traffic flow patterns embedded in the urban traffic digital twin model.

[0174] The correction module 36 is used to simulate the propagation process of traffic disturbances along the propagation path to adjacent nodes through the influence propagation model, using the road network topology as the propagation path and the node state deviation in the initial prediction map as the disturbance amount, outputting a correction amount, and correcting the initial prediction map according to the correction amount to obtain the target prediction map.

[0175] The solution module 37 is used to perform strategy solving based on the target prediction map and the road capacity, using an evolutionary game mechanism to generate emergency strategies.

[0176] The digital twin-based smart city emergency management system of this application is used to implement the aforementioned digital twin-based smart city emergency management method. Therefore, the specific implementation of the digital twin-based smart city emergency management system can be found in the embodiment section of the digital twin-based smart city emergency management method above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0177] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described digital twin-based smart city emergency management methods.

[0178] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described digital twin-based smart city emergency management methods.

[0179] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0180] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the smart city emergency management method based on digital twins described above.

[0181] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0182] The above provides a detailed description of a smart city emergency management method and system based on digital twins provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A smart city emergency management method based on digital twins, characterized in that, include: Acquire traffic status data and road capacity of urban roads, as well as real-time traffic speed and driving intention data of each vehicle; Based on the real-time traffic flow speed, the driving intention data, and the historical behavior data of the corresponding vehicles, a memory-enhanced network is used to perform state backtracking and intention prediction to generate a vehicle state vector representing the future behavior of each vehicle. The vehicle state vector is used as a micro-behavioral input and injected into a pre-built urban traffic digital twin model. At the same time, the traffic state data is used as a macro-system state to drive the operation of the urban traffic digital twin model. In the digital twin model of urban traffic, a graph neural network is used to perform twin inference processing on the traffic state data and the vehicle state vector to obtain an initial prediction map of the traffic state of urban roads in the future. Based on the road network topology and traffic flow patterns embedded in the urban traffic digital twin model, an impact propagation model based on a spatiotemporal graph neural network is constructed. Using the road network topology as the transmission path and the node state deviation in the initial prediction map as the disturbance amount, the influence propagation model is used to simulate the propagation process of traffic disturbances along the transmission path to adjacent nodes, outputting a correction amount, and correcting the initial prediction map based on the correction amount to obtain the target prediction map. Based on the target prediction map and the road capacity, an evolutionary game mechanism is used to solve the strategy and generate an emergency strategy.

2. The method according to claim 1, characterized in that, In the urban traffic digital twin model, a graph neural network is used to perform twin inference processing on the traffic state data and the vehicle state vector to obtain an initial prediction map of the urban road traffic state over a future period, including: Based on the traffic state data and the vehicle state vector, a traffic topology map is constructed with road segments as environment nodes, vehicles as intelligent agent nodes, and traffic lights as control nodes. The candidate control strategies associated with the control nodes and the future behavior strategies associated with the agent nodes are injected into the traffic topology as policy features. Through the policy and situation coupling layer in the graph neural network, the diffusion effect of the candidate control strategies along the road network connection relationship to the environmental nodes and agent nodes is simulated. At the same time, the travel demand implied by the future behavior strategies is integrated to generate a dynamic policy influence situation field that comprehensively reflects the control intention and travel demand, which serves as the macro pressure field in the cross-scale causal feedback constraint mechanism. The influence of the strategy on the situation field is used as edge weights to dynamically reconstruct the strength of the interaction relationship between nodes in the traffic topology graph. Through the multi-role interaction evolution layer in the graph neural network, based on the strength of the reconstructed interaction relationship, the accessibility of environmental nodes, the decision preferences of agent nodes, and the policy utility of control nodes are co-evolved and their states are updated to generate a spatiotemporal topology graph. Among them, the macroscopic pressure field is dynamically mapped as a moderating factor affecting the decision preferences of agent nodes, and the statistical behavior pattern of the evolved agent nodes is used to correct the macroscopic pressure field. Based on the aforementioned spatiotemporal topology, attention-guided propagation simulation is used to model the game and state adjustment process between agent nodes and control nodes under preset emergency constraints, generating a simulation diagram. Based on the projection diagram, the state of road network nodes over a future period is reconstructed through the decoder layer of the graph neural network to generate an initial prediction diagram.

3. The method according to claim 2, characterized in that, Based on the spatiotemporal topology graph, attention-guided propagation simulation is used to model the game and state adjustment process between agent nodes and control nodes under preset emergency constraints, generating a simulation graph, including: Based on the spatiotemporal topology graph, the association strength between the agent node and the current control node and the downstream control node is calculated through the attention matching mechanism to obtain the node association degree. The node correlation degree and the spatiotemporal topology graph are input into the attention diffusion network. The correlation degree is then propagated and superimposed along the road network topology through the attention diffusion network to obtain the propagation field. Combining the propagation field with preset emergency constraints, the strategy parameters of the control nodes are optimized and adjusted through dynamic strategies, and the expected paths of the agent nodes are replanned to generate a projection diagram.

4. The method according to claim 1, characterized in that, The process of generating emergency strategies based on the target prediction map and the road capacity, using an evolutionary game mechanism, includes: Potential traffic conflict zones and critical rescue routes that need to be ensured to pass through are identified from the target prediction map; Based on the identification results, the road capacity, and the traffic state represented by the target prediction map, the target prediction map is processed by a game theory model construction method to construct a situation field representing traffic conflicts and traffic priorities, as well as a transmission field representing changes in road network state. By integrating the situation field and the transmission field, a game field that integrates traffic situation and propagation laws is generated; The various preset traffic light control schemes and path guidance schemes are combined and defined as the hybrid strategy space to be optimized; Under the constraints of the game field, evolutionary game optimization is performed on the strategies in the mixed strategy space through multiple rounds of iteration. In each round of optimization, the fitness of the corresponding strategy is updated according to the simulation effect of the strategy in the game field to obtain the strategy ratio. When the strategy ratio reaches the preset equilibrium condition, the emergency strategy is obtained.

5. The method according to claim 4, characterized in that, Under the constraints of the game field, multi-round iterative evolutionary game optimization is performed on the strategies in the mixed strategy space. Each round of optimization updates the corresponding fitness based on the simulated performance of the strategy in the game field to obtain the strategy proportion, including: The strategies in the hybrid strategy space are allocated according to a preset initial ratio to form an initial strategy population; Under the constraints of the game field, the fitness of the initial strategy population is evaluated by a scenario-based performance simulation and comprehensive evaluator to obtain the fitness distribution. The evaluation process of the scenario-based performance simulation and comprehensive evaluator includes: inputting the strategy into the game field to simulate traffic conditions, and extracting the conflict resolution amount of the strategy on potential conflict sections, the efficiency improvement value of key rescue channels, and the resulting road network load cost. The basic fitness of each strategy is calculated by weighting according to preset weights. The basic fitness is corrected by combining capacity penalty and inter-strategy competition adjustment to obtain the final fitness distribution. Based on the fitness distribution, the proportion of each strategy in the initial strategy population is dynamically updated through the replicator dynamic equation to obtain the proportion change vector; The proportional change vector and the game field are input into the equilibrium solver to obtain the strategy proportion.

6. The method according to claim 1, characterized in that, Based on the real-time traffic flow speed, the driving intention data, and the corresponding vehicle's historical behavior data, a memory-enhanced network is used for state backtracking and intention prediction to generate a vehicle state vector representing the future behavior of each vehicle, including: The real-time traffic flow speed and the driving intention are fused and data completion processes are performed to obtain spatiotemporal correlation data; The spatiotemporal correlation data and the corresponding vehicle's historical behavior data are input into the memory enhancement network, and the memory retrieval unit of the memory enhancement network performs matching and retrieval to obtain the memory index; Based on the memory index, the inference unit of the memory enhancement network performs memory playback and state deduction to obtain risk prediction data for each vehicle; By using a risk diffusion model based on cross-attention and combining the risk prediction data, the attention weights between vehicles in the current urban road network are calculated to obtain the risk impact value of each vehicle. By integrating the risk prediction data and the risk impact value, a vehicle state vector is generated for each vehicle.

7. The method according to claim 1, characterized in that, The method uses the road network topology as the propagation path, the node state deviation in the initial prediction map as the disturbance, and simulates the propagation process of traffic disturbances along the propagation path to adjacent nodes using the influence propagation model. It outputs a correction value and corrects the initial prediction map based on the correction value to obtain the target prediction map, including: Calculate the node state deviation for each node in the initial prediction graph; Based on the connection relationship of the road network topology, the transmission path between nodes is determined, and the sensitivity coefficient of each node to disturbance is calculated based on the node attributes. Using the node state deviation as traffic disturbance, and combining it with the sensitivity coefficient, the spatiotemporal graph convolutional layer of the influence propagation model is used to simulate the multi-hop propagation of traffic disturbance along the transmission path, generating the cumulative influence value of each node. The cumulative impact value is mapped to the correction amount of each node through the output layer of the impact propagation model; The state of the corresponding node in the initial prediction map is corrected based on the correction amount to generate the target prediction map.

8. A smart city emergency management system based on digital twins, characterized in that, include: The acquisition module is used to acquire traffic status data and road capacity of urban roads, as well as real-time traffic speed and driving intention data of each vehicle. The prediction module is used to generate vehicle state vectors representing the future behavior of each vehicle by performing state backtracking and intention prediction through a memory augmentation network based on the real-time traffic flow speed, the driving intention data, and the historical behavior data of the corresponding vehicles. The driving module is used to inject the vehicle state vector as a micro-behavior input into the pre-built urban traffic digital twin model, and at the same time, use the traffic state data as a macro-system state to drive the operation of the urban traffic digital twin model. The extrapolation module is used to perform twin extrapolation processing on the traffic state data and the vehicle state vector in the urban traffic digital twin model using a graph neural network to obtain an initial prediction map of the traffic state of urban roads for a period of time in the future. The construction module is used to build an impact propagation model based on a spatiotemporal graph neural network, based on the road network topology and traffic flow patterns embedded in the urban traffic digital twin model. The correction module is used to simulate the propagation process of traffic disturbances along the propagation path to adjacent nodes through the influence propagation model, using the road network topology as the propagation path and the node state deviation in the initial prediction map as the disturbance amount, outputting a correction amount, and correcting the initial prediction map according to the correction amount to obtain the target prediction map. The solution module is used to perform strategy solving based on the target prediction map and the road capacity, and generate emergency strategies using an evolutionary game mechanism.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the digital twin-based smart city emergency management method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the smart city emergency management method based on digital twins as described in any one of claims 1 to 7.

Images