Method for locating critical facilities in traffic network based on reinforcement learning and cascade dynamics

Abstract

The application discloses a traffic network key facility location method based on reinforcement learning and cascade kinetics, and belongs to the technical field of intelligent traffic network, and solves the problem that the traditional method is not accurate in identifying key nodes and is easy to fall into local optimization. The method comprises the following steps: a weighted complex network is constructed based on a road network, intersections are taken as nodes, road sections are taken as edges, multi-dimensional characteristics such as traffic capacity, location and risk are fused, and the entropy weight method is used to objectively distribute node weights; a cascade kinetics model based on an improved coupled map lattice is established to simulate the dynamic propagation and amplification process of disturbance in the network; a residual network is taken as a state, node removal is taken as an action, and network efficiency loss is taken as a reward, a reinforcement learning model is trained, and a node failure strategy that makes the whole network efficiency decrease the most is learned end to end; and key node sets are selected in sequence according to Q values, and the key node sets are taken as facility preferential deployment positions. The application realizes precise location of global optimization, and significantly improves road network resilience.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation network technology, and in particular relates to a method for site selection of key facilities in transportation networks based on reinforcement learning and cascade dynamics. Background Technology

[0002] With the rapid development of smart cities and the digital transformation of transportation infrastructure, the optimized deployment of critical transportation infrastructure has become an important means to enhance road network resilience and ensure the reliable operation of urban transportation systems. In complex urban road networks, the functional failure of local nodes (such as key intersections or road sections) may trigger large-scale network paralysis through cascading effects. Therefore, accurately identifying key weak links in the network and prioritizing the deployment of smart transportation infrastructure has significant engineering value.

[0003] In the prior art, Chinese patent CN116579112A discloses a method for selecting emergency bridging bus parking spots based on rail transit networks. This method identifies key stations in the rail transit network, analyzes the emergency bridging bus rescue capabilities, establishes and solves a parking spot effectiveness evaluation model and a parking spot selection model to obtain the optimal parking spot scheme. However, this scheme uses static indicators (node ​​degree, vulnerability, passenger flow) combined with grey relational analysis to identify key stations, without considering the dynamic propagation process of cascading failures in the rail transit network. Furthermore, it uses a genetic algorithm for static multi-objective optimization, which makes it difficult to break through the limitations of local optima and cannot accurately identify the key weak links that cause a significant drop in the efficiency of the entire network. It also lacks adaptability to the dynamic evolution of actual traffic flow.

[0004] Chinese patent CN119784097A discloses an emergency facility site selection and service optimization model under facility failure and traffic congestion conditions. This model uses a pre-disaster planning decision model and a post-disaster dynamic optimization model to form a two-level mixed integer nonlinear programming model to realize emergency shelter site selection, service allocation and route adjustment. However, this scheme only considers the independent failure of a single or partial shelter and does not simulate the cascading failure effect of the traffic network caused by facility failure. It simplifies the simulation of traffic congestion based on the user equilibrium model, which cannot capture the chaotic characteristics and cascading amplification law of traffic flow. Moreover, it is easy to get trapped in local optima when using traditional mathematical programming methods to solve the problem. It does not integrate the multi-modal traffic interaction characteristics and is difficult to adapt to complex dynamic scenarios.

[0005] Chinese patent CN112232599A discloses a method for selecting emergency medical facilities for public health events based on POI data. This method organizes POI data, establishes a traffic network model, and dynamically updates basic data. It uses an ArcGIS location allocation model to select emergency medical facilities at the community and center levels. However, this method relies on manually setting mandatory options and static time impedance constraints for site selection. It does not consider the cascading effects caused by the failure of medical facilities. It only updates the data periodically and then re-solves the static solution. It is not a real-time dynamic decision-making process. Furthermore, it does not identify the most critical nodes for the overall network service resilience through objective network analysis. It is prone to site selection deviations due to subjective settings.

[0006] Furthermore, existing technologies for identifying key nodes in transportation networks mainly rely on traditional network topology metrics, such as degree centrality, betweenness centrality, or eigenvector centrality, which are static structural measures. These metrics are used to rank the importance of nodes and thus determine the location of infrastructure deployment. However, such methods have the following drawbacks: (1) Static structural metrics ignore the dynamic propagation characteristics of traffic flow and the actual cascading failure process, and cannot accurately reflect the dynamic impact of node failure on the overall network performance; (2) Traditional methods often use a single metric for evaluation and do not comprehensively consider the coupling effect of basic road map elements (capacity), location characteristics, and multi-modal traffic interaction characteristics on the importance of nodes; (3) Existing location selection methods based on cascading failure often use greedy strategies or random attack simulations, making it difficult to find the optimal set of key nodes that maximize the impact of network failure in the solution space, and easily getting trapped in local optima.

[0007] Therefore, there is an urgent need for a technical solution that can integrate network topology, traffic flow dynamics, and cascading failure mechanisms, and globally optimize the location strategy through adaptive learning methods, so as to break through the limitations of traditional location methods and achieve accurate identification of the deployment location of intelligent transportation infrastructure. Summary of the Invention

[0008] The purpose of this invention is to provide a method for selecting key infrastructure locations in transportation networks based on reinforcement learning and cascade dynamics. By constructing a weighted complex network and a cascade dynamics model, and using reinforcement learning algorithms to learn node failure strategies end-to-end, this method solves the technical problems of inaccurate identification of key weak links in complex networks and local optima in infrastructure location selection.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by this invention is a method for site selection of critical facilities in transportation networks based on reinforcement learning and cascade dynamics, comprising:

[0010] A weighted complex network is constructed based on the road network, where road intersections are nodes and road segments between adjacent intersections are edges. The node weights are determined by fusing the basic features of the road map, location features, and other road features.

[0011] A network cascade dynamics model was established to simulate the propagation process of node failure in the network;

[0012] Using a reinforcement learning model, the agent is trained to learn a strategy for removing nodes, with the reduction in network efficiency as a reward; and

[0013] Based on the trained reinforcement learning model, the nodes that cause the greatest decrease in network efficiency are selected sequentially from the weighted complex network to obtain the set of key nodes, which are used as the sites for key facilities in the transportation network.

[0014] Furthermore, the construction of the weighted complex network includes using the Space L method, with intersections as nodes and road segments between adjacent intersections as edges; and defining node weighting functions. :

[0015]

[0016] in, As basic graph elements, As a location feature, For other features, These are the weighting coefficients.

[0017] Furthermore, the weighting coefficients The determination process includes:

[0018] Based on the basic graph elements of each node Location characteristics Other characteristics A judgment matrix is ​​constructed for the evaluation indicators, the range is standardized for the evaluation indicators, the information entropy and entropy weight of each indicator are calculated, and the obtained weights are normalized.

[0019] Furthermore, the network cascade dynamics model includes:

[0020] Nodes are defined based on coupled image lattice. exist The state at time is for:

[0021]

[0022] in, For the improved Logistic mapping function, Represents a node At any moment state, For chaotic parameters, For adjacency matrix elements, To improve the coupling coefficient, For nodes The degree; For nodes The set of neighbors; For external disturbance terms; For traffic interaction feature functions, when node The state exceeds the failure threshold The node is determined to be invalid and removed from the network at the next moment.

[0023] Furthermore, the improved coupling coefficient Specifically, it is a function related to the traffic capacity of adjacent road segments and the signal coordination coefficient:

[0024]

[0025] in They are nodes The traffic capacity of the road section; They are nodes saturation This is the signal coordination coefficient, which reflects the degree of coordination between signal timings at adjacent intersections.

[0026] Furthermore, the improved Logistic mapping function is a piecewise function, employing different evolution rules based on the traffic states of free flow, synchronous flow, and congested flow, including a free flow attenuation coefficient and a congestion amplification coefficient.

[0027] Furthermore, nodes The failure threshold is dynamically adjusted based on the node's historical state fluctuations, node degree, and node weight; when the node's state value exceeds its failure threshold, the node is determined to be failed and removed from the network.

[0028] Furthermore, the use of reinforcement learning models includes:

[0029] Graph embedding is performed on the residual network to generate state embedding vectors. and action embedding vector ;

[0030] Calculate the score of the state-action pair using a scoring function. :

[0031]

[0032] in, and These are the weight parameters learned during the model training phase; and It is an embedded vector. It is the sigmoid function;

[0033] The reward used to quantify the traffic efficiency loss in complex network paths is determined based on the difference in network efficiency before and after the action. ,in This represents the immediate reward obtained after taking an action at time t. Based on the network efficiency of the previous moment, For the network efficiency at the current moment; and

[0034] The Q-value is updated using the n-step DQN algorithm, and the action is selected using an ε-greedy strategy.

[0035] Furthermore, the Q-value update strategy is specifically expressed as follows:

[0036]

[0037] in, This is the current state; It is in state The following actions were taken; It is to take action The immediate reward obtained is used to quantify the traffic efficiency loss in complex network paths. It is to take action The new state that was subsequently transferred to; In the new state All possible actions below; It is the learning rate, used to control the degree to which new information is accepted in each iteration; It is a discount factor.

[0038] Furthermore, the improved Logistic mapping function Combining free flow, synchronous flow, and congested flow, different evolution rules are adopted in different state intervals, specifically:

[0039]

[0040] in, For the state threshold, This is the free-flow attenuation coefficient, reflecting the relatively small fluctuations in the node state under free-flow conditions. This is the congestion amplification factor, reflecting the rapid amplification of disturbances under congested conditions. Represents a node At any moment The state.

[0041] Compared with existing technologies, the beneficial effects of this invention include: In assessing node importance, traditional methods typically rely solely on single topological indicators such as node degree or betweenness centrality, resulting in a one-sided and highly subjective evaluation basis that fails to fully utilize multidimensional information. This invention overcomes this limitation by integrating multiple qualitative and quantitative factors such as road capacity, location characteristics, and accident risk, and employing the entropy weight method to objectively assign weights to each indicator. This makes the evaluation process more scientific, and the evaluation results more comprehensively and accurately reflect the true importance of nodes in the actual road network.

[0042] Existing technologies often oversimplify the portrayal of congestion and cascading failures when simulating the propagation of traffic disturbances, making it difficult to realistically reproduce the dynamic chain reaction of disturbances spreading and amplifying within the network. This invention, based on an improved coupled image lattice model, fully considers the specific evolutionary patterns under different traffic states, from free flow to congested flow. This allows for the construction of a simulation process that more closely resembles the characteristics of real-world traffic flow, achieving a more realistic depiction of the impact of node failures on the entire network dynamics.

[0043] In optimizing the location of critical nodes, conventional methods often employ greedy strategies or static optimization algorithms, which can easily lead to solutions getting trapped in local optima. This invention innovatively introduces a reinforcement learning framework, enabling the agent to explore and learn through continuous interaction with the environment. This allows it to autonomously search for the combination of critical nodes that has the most significant impact on network robustness from a global and long-term perspective. This method effectively overcomes the limitations of local optima, resulting in a significant improvement in both the overall efficiency and effectiveness of the final location selection scheme. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of the method for selecting key infrastructure locations in transportation networks based on reinforcement learning and cascade dynamics in this implementation.

[0046] Figure 2 It is a complex road network diagram;

[0047] Figure 3 It is a key element identification map based on reinforcement learning. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] like Figure 1 As shown, this embodiment provides a method for selecting key infrastructure locations in transportation networks based on reinforcement learning and cascade dynamics. This method integrates network topology, traffic flow dynamics, and cascade failure mechanisms. Through adaptive learning, it globally optimizes the location strategy, accurately identifying the optimal set of key nodes for intelligent transportation infrastructure deployment. This overcomes the limitations of traditional static index evaluation, single feature considerations, and local optimum solutions, improving road network resilience and the reliability of the transportation system. In some specific embodiments, the method for selecting key infrastructure locations in transportation networks based on reinforcement learning and cascade dynamics includes the following steps:

[0050] S1. Constructing a weighted complex road network based on the SpaceL method: Using road intersections as network nodes and road segments between adjacent intersections as edges, a weighting function is constructed based on basic road map elements, location features, and other road features. The entropy weight method is used to determine the weight coefficients of each feature, generating a weighted network triplet containing a set of nodes, a set of edges, and a set of weights.

[0051] In some specific implementations, the S1 process is as follows:

[0052] S101. Based on actual road geographic information data, the SpaceL network construction method is adopted, with road intersections as network nodes. The road segment between two adjacent intersections is a connecting edge. Construct an undirected graph ,in For a set of nodes, It is a set of connected edges.

[0053] S102, Define the node weighting function Integrating three types of features: basic graph elements (Reflecting traffic supply capacity characteristics such as road segment capacity and saturation) Location characteristics (Reflecting the importance of a node's topological location in the network, such as centrality and reachability) and other characteristics (Reflecting external risk factors such as accident rate and construction impact coefficient):

[0054]

[0055] S103. Determine the weighting coefficients using the entropy weighting method. Constructing an evaluation index judgment matrix ,in The number of nodes to be evaluated (i.e., the total number of network nodes) ), For the number of evaluation indicators (here) These correspond to three dimensions: basic map elements, location features, and other features, respectively. Indicates the first The node at the th The original observations for the evaluation index were used. The range standardization method was employed to analyze the original observations for the evaluation index. The node at the th The original observations of each evaluation indicator were normalized:

[0056]

[0057] Obtain the normalized matrix Calculate the proportion of features. Calculate the first Information entropy of the indicator , among which when Time definition Then calculate the entropy weight. ,in Corresponding basic graph element weights , Corresponding location feature weights , Corresponding to other feature weights The evaluation indicators are weighted based on entropy weights to obtain the final weighting function. This allows for the objective allocation of weights.

[0058] S104. Normalize the weights. ,in The maximum weight value in the network is used to obtain the final weighted network triple. ,in .

[0059] like Figure 2 As shown, Figure 2 This is a schematic diagram of the complex road network topology constructed in this embodiment. The diagram illustrates the process of constructing a weighted complex network based on the Space L method in step S1 above. The actual urban road network is abstracted as a graph structure in the diagram, where circular nodes represent actual road intersections, and the edges connecting the nodes represent road segments between adjacent intersections.

[0060] S2. Establish a network cascade dynamics model: Based on the coupled image lattice model, construct a cascade failure model that considers the spatiotemporal discrete characteristics, introduce an improved coupling coefficient, Logistic mapping function and node failure threshold, establish a node state evolution equation, and simulate the cascade failure propagation process under initial disturbance.

[0061] In some specific implementations, the S2 process is as follows:

[0062] S201. Construct a network cascade dynamics model based on the Coupled Image Lattice (CML) model, considering the spatiotemporal discrete characteristics and chaotic behavior of traffic flow. Define nodes. At any moment The state is Its evolution equation ( The time is:

[0063]

[0064] in, For the improved Logistic mapping function, Represents a node At any moment state, This is the chaos parameter (usually set to 4). For adjacency matrix elements, The improved coupling coefficient has a value range of (0,1). For nodes The degree; For nodes The set of neighbors; For external disturbance terms; This is a traffic interaction feature function used to characterize the coupled influence of multimodal traffic flow on node states.

[0065] S202. The traffic significance of the coupling coefficient is reconstructed by redefining it as the traffic flow influence intensity factor between road segments. Its value is related to the road segment capacity, saturation, and signal control parameters, making the state transmission between nodes not only dependent on the topological connection strength but also reflecting the dynamic coupling relationship of actual traffic flow. By introducing real-time traffic flow data for dynamic calibration, the model's ability to fit actual traffic conditions is improved. Its traffic significance is that the coupling strength is positively correlated with the road segment capacity and negatively correlated with the saturation difference.

[0066] As shown in the following formula:

[0067]

[0068] in ; are nodes respectively The traffic capacity of the road section ; They are nodes Saturation (flow / capacity). The signal coordination coefficient reflects the degree of coordination in signal timing between adjacent intersections. .

[0069] S203. The improvement of the Logistic mapping function: Traditional Logistic mapping is used to describe chaotic behavior. This invention extends it to a piecewise mapping function, combining it with the three-stage theory of traffic flow (free flow, synchronous flow, congested flow). Different evolution rules are adopted in different state intervals, making the node state evolution more closely match the actual traffic flow change pattern, as shown in the following equation:

[0070]

[0071] in, To determine the state threshold, distinguish between three stages (recommendation). ), This is the free-flow attenuation coefficient, reflecting the relatively small fluctuations in the node state under free-flow conditions. This is the congestion amplification factor, reflecting the rapid amplification of disturbances under congested conditions.

[0072] S204. Adaptive setting of failure threshold: The node failure threshold is no longer uniformly set, but dynamically adjusted based on the fluctuations in the node's historical state, traffic flow density, and road segment importance, improving the model's adaptability to heterogeneous networks. Setting the node failure threshold... ,when Time-determined node Failure, the node will fail at the next time step. Remove a node from the network and delete its edges. The updated network recalculates the degree of each node and its neighbor set, and enters the next time step of evolution until the network reaches a steady state or completely collapses.

[0073]

[0074] in, The baseline threshold is 0.85 (recommended).

[0075] For nodes In recent The standard deviation of the state fluctuation at each time step Represents historical state values, represents nodes In the past The specific state values ​​at each time step. For nodes The degree is the number of connected edges. The normalized weights of the nodes calculated for S104. To adjust the parameters to meet the requirements To ensure the validity of the threshold.

[0076] S3, Reinforcement Learning Model Training:

[0077] The reinforcement learning model consists of four steps, including (1) graph weighting: assigning traffic-related weights to each road network node; (2) graph embedding: encoding the structural information and traffic flow information features of the nodes; (3) Q function learning: the function learns to evaluate the Q value of the node after performing the node failure operation in a given state; and (4) ε-greedy strategy selection: the key node is selected based on its Q value through this selection.

[0078] S301, Graph Weighting: To better integrate the structural and functional information of the complex road network, the training graph needs to be preprocessed before the graph embedding process. The basic road graph elements (F), location (L), and other road features (Other, O) are used as the basis for node weighting, and the resulting weighting function is shown in S103.

[0079] S302. Graph Embedding: The goal is to process the basic graph elements, location, and other road features of road networks in a low-dimensional manner. However, real-world networks can be very large and contain complex graph information. Considering the data complexity, Graph Neural Networks (GNNs), similar to GraphSAGE, are used to encode the state (S) (residual network) and actions (A) (nodes that should be invalidated) into state embedding vectors. and action embedding vector Using these embedding vectors, we can compute the maximum expected reward for performing an action (A) in a given state (S). Embedded vectors are needed to represent node features, including their positions on the graph (nodes). In the figure (relative positions in the equation) and their failure probabilities.

[0080] It is important to note that 𝐾 determines the depth of the embedding process, i.e., the embedding vector. It includes information about K-hop neighbors. Furthermore, informational features of complex road networks are included in the input features. In this context, the characteristics of a network to be built, which maximize the impact of network failures, can be evaluated using a series of topological features. Taking network efficiency as an example, it can measure the ease of connection between node pairs and demonstrate network connectivity and performance. The average network efficiency... As shown in the following formula:

[0081]

[0082] in This represents the efficiency value of the road conflict network; Represents a node Distance to node j; This represents the number of nodes in the road conflict network.

[0083] Node contribution can be reflected by the change in network efficiency before and after a node failure. When calculating this metric, a node removal method is used to remove failed nodes when cascading failures are simulated. Then, the network efficiency before and after the node failure is recorded, and the degree of change is compared to measure the node's importance to the network. If the change in network efficiency is small, it means that the node's failure has little impact on the network, indicating that the node contributes little to network connectivity. Conversely, if the difference in network efficiency before and after failure is large, with a significant decrease in network efficiency after failure, then the node performs important connectivity functions and is therefore considered an important node in the network. To make the contribution value proportional to the degree of change—that is, the greater the degree of change, the higher the contribution value—a network efficiency contribution metric can be introduced.

[0084]

[0085] in, For nodes The network efficiency value after failure. Node contribution is an input feature. One of the important pieces of information in it.

[0086] S303 and Q function learning

[0087] During the decoding phase, the goal is to use a scoring function. Pairing Scoring is performed (scoring the state-action pairs). The Q function is designed as follows:

[0088]

[0089] in and These are the weight parameters learned during the model training phase; and It is an embedded vector. These are sigmoid functions, used to represent action a and state s, respectively.

[0090] To optimize weight parameters during the encoding and decoding stages, this implementation introduces an n-step DQN mechanism. This reinforcement learning process includes key components such as state space, actions, rewards, policies, and loss functions, defined as follows:

[0091] 1) Environment: Complex road network topology.

[0092] 2) State space: The state is defined as the residual network after the cascading failure of the network at any time, taking into account critical nodes.

[0093] 3) Action: The action is to remove identified nodes that are considered important.

[0094] 4) Reward: The reward for an action is a quantified value of its impact on the traffic efficiency loss due to the complex road network, as shown in the formula below:

[0095] (i.e., the reduction in network efficiency)

[0096] in This represents the immediate reward obtained after taking an action at time t; The network efficiency at the previous moment; This represents the network efficiency at the current moment.

[0097] 5) Update Strategy: In each iteration, this implementation method updates the strategy based on the current reward. and new status Update the current state-action pair using the maximum Q value. The Q value is calculated using the following formula:

[0098]

[0099] in, This is the current state; It is in state The following actions were taken; It is to take action The immediate reward received afterward; It is to take action The new state that was subsequently transferred to; In the new state All possible actions below; It is the learning rate, which controls the degree to which new information is accepted in each iteration; It is a discount factor used to balance immediate rewards and future rewards.

[0100] It should be noted that in this implementation method, the Q-value update strategy is based on probability. Select the node with the highest Q value With probability Randomly select nodes for exploration. Use n-step DQN to update parameters:

[0101]

[0102] Where the target value , For the target network parameters, periodically from Replication update, i.e., the aforementioned formula for updating the Q value.

[0103] like Figure 3 As shown, Figure 3 This embodiment demonstrates the identification of key elements (nodes) and the dynamic evolution of cascading failures based on reinforcement learning. Figure 3 This corresponds to the interaction process between the cascaded dynamics model (step S2) and the reinforcement learning model training (step S3) in this embodiment.

[0104] Figure 3 The road network section on the left illustrates the graph weighting and graph embedding process. The multi-layered stacked network structure represents the integration of information such as basic road network graph elements, location features, and other features (i.e., the graph weighting step), and then the complex network structure and features are encoded into state embedding vectors and action embedding vectors.

[0105] Figure 3 The main body of the text demonstrates the iterative optimization process of reinforcement learning: In Figure 3 In Step 1, the network is in a complete state; the agent selects and removes the first critical node according to the policy (represented in the diagram as a change in the state of a specific node), which then triggers a cascading failure dynamics evolution based on the improved Coupled Image Lattice (CML) model. Figure 3 Starting with Step 2, the network evolves into a residual network, and the agent continues to remove the next node based on the current residual network state. For each action performed, the environment provides an immediate reward based on the reduction in network efficiency. Through multiple iterations, the agent eventually learns the node removal strategy that maximizes the reduction in network efficiency. Its output action sequence corresponds to the set of critical nodes in the transportation network, thereby guiding the priority location of critical infrastructure.

[0106] S4. Critical Infrastructure Site Selection Decisions:

[0107] A well-trained reinforcement learning model is applied to the target road network. The Q-value of each node is calculated through forward propagation. The set of key nodes is selected in descending order of Q-value. The node with the largest Q-value is selected as the current failed node. The network state is updated and the process is repeated until the preset failure ratio or network collapse threshold is reached. The resulting node failure sequence is the optimal set of locations for the deployment of critical infrastructure.

[0108] The well-trained reinforcement learning model is deployed to the target road network. First, graph weighting and feature extraction are performed on the real-world network, and then the trained encoder is input to generate low-dimensional embedding vectors.

[0109] The Q-values ​​of each candidate node in the current state are calculated by forward propagation, and the node with the largest Q-value is selected as the first critical node. The failure of this node is simulated and the cascading process is triggered. After updating the network state, the Q-values ​​of the remaining nodes are recalculated, and the next critical node is selected.

[0110] Repeat the above process until the preset infrastructure deployment budget ratio or number of key nodes is reached. The resulting sequence of failed nodes is the optimal set of key nodes that maximizes the decline in network efficiency. These locations are prioritized for deployment and key protection of smart transportation infrastructure (such as intelligent traffic signals, roadside sensing units, and emergency control equipment).

[0111] This embodiment also provides a location selection system for critical infrastructure in transportation networks based on reinforcement learning and cascade dynamics. The system includes: a data acquisition module, a network modeling module, a cascade simulation module, a reinforcement learning training module, and a location decision module.

[0112] The data acquisition module is used to acquire basic road data, including road network topology, road segment capacity, traffic flow status, and signal timing information.

[0113] The network modeling module abstracts the road network into a weighted complex network based on the SpaceL method, and generates node weights by fusing multi-source heterogeneous features through the entropy weight method.

[0114] The cascaded simulation module constructs a cascaded dynamics model based on the coupled image lattice model to simulate the network failure propagation process under different attack strategies and calculate the network efficiency loss.

[0115] The reinforcement learning training module includes a graph embedding layer, a Q-network, and a policy optimizer, which are used to train the agent to learn a node attack strategy that maximizes the impact of cascading failures.

[0116] The site selection decision module is used to deploy the trained model to the actual road network, output the ranking of key nodes through forward inference, and generate a priority deployment plan for intelligent transportation infrastructure.

[0117] In some specific implementations, the network modeling module includes: a feature extraction unit for extracting basic graph elements, location features, and traffic interaction features; a weight calculation unit for calculating the weights of each feature using the entropy weight method and generating a weighted network; and a matrix generation unit for constructing the adjacency matrix, degree matrix, and Laplacian matrix of the network.

[0118] As a preferred technical solution, the cascaded simulation module includes: a state initialization unit for setting the initial state of nodes and chaotic parameters; a dynamic evolution unit for performing state updates of the coupled image lattice model; a failure determination unit for determining node failure based on a threshold and updating the network topology; and a performance evaluation unit for calculating network efficiency, connectivity, and connectivity performance indicators.

[0119] Compared with traditional traffic flow propagation models based on fluid dynamics or cellular automata, the improved CML model used in this embodiment has the following significant advantages:

[0120]

[0121] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for site selection of critical infrastructure in transportation networks based on reinforcement learning and cascade dynamics, characterized in that, include: A weighted complex network is constructed based on the road network, where road intersections are nodes and road segments between adjacent intersections are edges. The node weights are determined by fusing the basic features of the road map, location features, and other road features. A network cascade dynamics model was established to simulate the propagation process of node failure in the network; Using a reinforcement learning model, the agent is trained to learn a strategy for removing nodes, with the amount of reduction in network efficiency as a reward. as well as Based on the trained reinforcement learning model, the nodes that cause the greatest decrease in network efficiency are selected sequentially from the weighted complex network to obtain the set of key nodes, which are used as the sites for key facilities in the transportation network.

2. The method for site selection of critical infrastructure in transportation networks based on reinforcement learning and cascade dynamics as described in claim 1, characterized in that, The construction of the weighted complex network includes using the Space L method, with intersections as nodes and road segments between adjacent intersections as edges; and defining node weighting functions. : ； in, As basic graph elements, As a location feature, For other features, These are the weighting coefficients.

3. The method for selecting the location of critical infrastructure in a transportation network based on reinforcement learning and cascade dynamics as described in claim 2, characterized in that, The weighting coefficients The determination process includes: Based on the basic graph elements of each node Location characteristics Other characteristics A judgment matrix is ​​constructed for the evaluation indicators, the range is standardized for the evaluation indicators, the information entropy and entropy weight of each indicator are calculated, and the obtained weights are normalized.

4. The method for site selection of critical infrastructure in transportation networks based on reinforcement learning and cascade dynamics as described in claim 1, characterized in that, The network cascade dynamics model includes: Nodes are defined based on coupled image lattice. exist The state at time is for: ； in, For the improved Logistic mapping function, Represents a node At any moment state, For chaotic parameters, For adjacency matrix elements, To improve the coupling coefficient, For nodes The degree; For nodes The set of neighbors; For external disturbance terms; For traffic interaction feature functions, when node The state exceeds the failure threshold The node is determined to be invalid and removed from the network at the next moment.

5. The method for selecting the location of critical infrastructure in a transportation network based on reinforcement learning and cascade dynamics as described in claim 4, characterized in that, The improved coupling coefficient Specifically, it is a function related to the traffic capacity of adjacent road segments and the signal coordination coefficient: ； in They are nodes The traffic capacity of the road section; They are nodes saturation This is the signal coordination coefficient, which reflects the degree of coordination between signal timings at adjacent intersections.

6. The method for selecting the location of critical infrastructure in a transportation network based on reinforcement learning and cascade dynamics according to claim 4, characterized in that, The improved Logistic mapping function is a piecewise function that adopts different evolution rules according to the traffic states of free flow, synchronous flow, and congested flow, including a free flow attenuation coefficient and a congestion amplification coefficient.

7. The method for site selection of critical infrastructure in transportation networks based on reinforcement learning and cascade dynamics according to claim 4, characterized in that, node The failure threshold is dynamically adjusted based on the node's historical state fluctuations, node degree, and node weight. When the state value of a node exceeds its failure threshold, the node is deemed to have failed and is removed from the network.

8. The method for site selection of critical infrastructure in transportation networks based on reinforcement learning and cascade dynamics according to claim 1, characterized in that, in, The use of reinforcement learning models includes: Graph embedding is performed on the residual network to generate state embedding vectors. and action embedding vector ; Calculate the score of the state-action pair using a scoring function. : ； in, and These are the weight parameters learned during the model training phase; and It is an embedded vector. It is the sigmoid function; The reward used to quantify the traffic efficiency loss in complex network paths is determined based on the difference in network efficiency before and after the action. ,in This represents the immediate reward obtained after taking an action at time t. Based on the network efficiency of the previous moment, For the network efficiency at the current moment; and The Q-value is updated using the n-step DQN algorithm, and the action is selected using an ε-greedy strategy.

9. The method for site selection of critical infrastructure in a transportation network based on reinforcement learning and cascade dynamics as described in claim 8, characterized in that, The Q-value update strategy is specifically expressed as follows: ； in, This is the current state; It is in state The following actions were taken; It is to take action The immediate reward obtained is used to quantify the traffic efficiency loss in complex network paths. It is to take action The new state that was subsequently transferred to; In the new state All possible actions below; It is the learning rate, used to control the degree to which new information is accepted in each iteration; It is a discount factor.

10. The method for site selection of critical infrastructure in a transportation network based on reinforcement learning and cascade dynamics as described in claim 4, characterized in that, The improved Logistic mapping function Combining free flow, synchronous flow, and congested flow, different evolution rules are adopted in different state intervals, specifically: ； in, For the state threshold, This is the free-flow attenuation coefficient, reflecting the relatively small fluctuations in the node state under free-flow conditions. This is the congestion amplification factor, reflecting the rapid amplification of disturbances under congested conditions. Represents a node At any moment The state.

Images