Dynamic road network collaborative expansion decision system based on federated learning-digital twinning

Abstract

The application relates to the technical field of intelligent traffic, in particular to a dynamic road network cooperative expansion decision system based on federated learning-digital twinning, which comprises the following steps: calibrating the weights of pulse type and periodic type data streams through a dynamic weight distributor to generate a normalized feature vector; fusing the normalized feature vectors of various domains through a privacy protection aggregation engine to obtain a passenger flow pressure distribution prediction matrix; a reconstruction module is used for constructing a digital twin based on the prediction matrix and initialization; based on the initialized twin environment, a blind area data reconstructor is used to identify the sensing missing area, and a space-time correlation algorithm is used to fuse historical features and real-time data streams of adjacent nodes to reconstruct the complete road network state. The application fuses the federated learning and digital twinning technologies to construct the dynamic road network cooperative expansion decision system, and improves the expansion decision efficiency of the traffic road network.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically, to a dynamic road network collaborative expansion decision system based on federated learning and digital twins. Background Technology

[0002] With urbanization and motorization, traffic congestion around large integrated transportation hubs is becoming increasingly serious. Sudden surges in passenger flow during peak hours and holidays overlap with regular traffic flow, affecting travel efficiency and urban operation.

[0003] Current road network expansion decisions rely on static plans, human experience, or data from a single department, which may struggle to cope with the dynamic changes in multi-source heterogeneous traffic data and the complexity of cross-departmental collaboration. Existing technologies may have the following shortcomings: First, data silos and privacy barriers may lead to biased decision-making, and data sharing between departments may be difficult, making it impossible to fully perceive cross-domain traffic flow relationships. Second, some static decision-making models are not suitable for dynamic traffic scenarios, and fixed plans may not consider real-time passenger flow fluctuations and lack dynamic adjustment mechanisms. Third, cross-departmental collaboration may face administrative barriers and execution delays, and differences in processes may lead to time differences in strategy execution, weakening the overall effect. Fourth, there is a lack of accurate simulation and risk prediction capabilities, relying on experience or simple models, resulting in high trial-and-error costs. Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a dynamic road network collaborative expansion decision system based on federated learning and digital twin. By integrating federated learning and digital twin technologies, a dynamic road network collaborative expansion decision system is constructed to improve the efficiency and reliability of traffic network expansion decisions.

[0005] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by the present invention is as follows:

[0006] The first aspect is a dynamic road network collaborative expansion decision-making system based on federated learning and digital twins, including:

[0007] The fusion module is used by various traffic management domains to input spatiotemporally heterogeneous traffic feature data into local clients, classifying it into pulse-type and periodic data streams; it calibrates the weights of pulse-type and periodic data streams through a dynamic weight allocator to generate normalized feature vectors; and it fuses the normalized feature vectors from each domain through a privacy-preserving aggregation engine to obtain a passenger flow pressure distribution prediction matrix.

[0008] The reconstruction module is used to construct and initialize a digital twin based on the prediction matrix; based on the initialized twin environment, the blind spot data reconstructor identifies the perception missing area, and the spatiotemporal correlation algorithm is used to fuse historical features with real-time data streams from neighboring nodes to reconstruct the complete road network status.

[0009] The expansion module is used to input the reconstructed complete road network state into the digital twin, execute multimodal expansion strategy simulation in parallel, and generate a set of expansion strategy simulation results.

[0010] The multi-objective module is used to input the simulation result set of the expanded strategy into the Pareto front optimization algorithm to obtain the multi-objective value set of the strategy; based on the multi-objective value set of the strategy, the baseline mode and the perturbation mode evaluation are performed to generate the Pareto optimal strategy combination;

[0011] The dynamic closed-loop module is used to input the Pareto optimal policy combination into the policy compilation engine, generate the execution instruction sequence, and parse the pre-set administrative delay feature matrix to dynamically trigger the instruction nodes that require approval; inject the execution status of the instruction nodes into the digital twin, and obtain the execution deviation through the deviation detector; when the execution deviation exceeds the adaptive tolerance boundary, trigger the perturbation mode re-optimization engine pre-set in the digital twin to generate a correction policy; call the update interface of the policy compilation engine to load the correction policy, forming a dynamic closed loop.

[0012] Secondly, the dynamic road network collaborative expansion decision-making method based on federated learning and digital twins includes the following steps:

[0013] Each traffic management domain inputs spatiotemporally heterogeneous traffic characteristic data into its local client, classifying it into pulse-type and periodic data streams; the weights of the pulse-type and periodic data streams are calibrated through a dynamic weight allocator to generate normalized feature vectors; and the normalized feature vectors from each domain are fused through a privacy-preserving aggregation engine to obtain a passenger flow pressure distribution prediction matrix.

[0014] A digital twin is constructed and initialized based on the prediction matrix; based on the initialized twin environment, a blind spot data reconstructor is used to identify the perception-missing areas, and a spatiotemporal correlation algorithm is used to fuse historical features with real-time data streams from neighboring nodes to reconstruct the complete road network status.

[0015] The complete road network state is reconstructed and input into the digital twin. Multimodal expansion strategy simulation is performed in parallel to generate a set of expansion strategy simulation results.

[0016] The simulation results set of the expanded strategy is input into the Pareto front optimization algorithm to obtain the multi-objective value set of the strategy; the baseline mode and perturbation mode evaluation are performed based on the multi-objective value set of the strategy to generate the Pareto optimal strategy combination;

[0017] The Pareto optimal policy combination is input into the policy compilation engine to generate an execution instruction sequence and parse the pre-set administrative delay feature matrix to dynamically trigger the instruction nodes that require approval. The execution status of the instruction nodes is injected into the digital twin, and the execution deviation is obtained through the deviation detector. When the execution deviation exceeds the adaptive tolerance boundary, the perturbation mode re-optimization engine pre-set in the digital twin is triggered to generate a correction policy. The update interface of the policy compilation engine is called to load the correction policy, forming a dynamic closed loop.

[0018] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art.

[0019] A federated learning architecture enables local processing and privacy-preserving aggregation of data from various traffic management domains, fusing multi-source heterogeneous traffic features without disclosing the original data, thus providing a more comprehensive basis for decision-making. By classifying pulse-type and periodic data streams and dynamically calibrating weights, normalized feature vectors are generated. Combined with spatial coordinate alignment and topological distance weighting, the consistency of cross-domain data in the spatiotemporal dimensions is ensured. A digital twin constructed based on the prediction matrix accurately reproduces the spatial structure and traffic characteristics of the road network. Virtual-real synchronization is achieved through initial state mirroring, providing a reliable virtual test field for subsequent simulation and decision-making. A blind spot data reconstructor identifies data missing areas and combines historical data from the same period with real-time information from neighboring nodes for spatiotemporal correlation reconstruction, ensuring the integrity and dynamic timeliness of the road network status.

[0020] Parallel simulation of multimodal expansion strategies in a digital twin enables the coverage of more solutions in a short time, improving decision-making efficiency. Virtual simulation is used to evaluate strategy effectiveness, avoiding congestion or resource waste caused by testing immature strategies in real road networks. Pareto front optimization algorithms are used to collaboratively optimize the three objectives of "congestion mitigation, cost control, and risk reduction," ensuring optimal overall strategy benefits. Dual evaluation using baseline and perturbation modes enhances the adaptability of decisions to complex scenarios. Pareto optimal combinations satisfying multiple constraints are generated, providing traffic management departments with clear and executable strategy options. Dynamic triggering of approval instruction nodes is achieved by analyzing the administrative delay feature matrix, adapting to process differences across departments. Real-time monitoring of instruction execution status and calculation of deviation rates automatically triggers a re-optimization engine to generate corrective strategies when deviations exceed tolerance boundaries. A closed-loop update mechanism ensures continuous adaptation of strategies to road network changes. This closed-loop process, from strategy generation and execution to deviation correction, enables the system to continuously respond to road network dynamics.

[0021] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Some specific embodiments of this application will be described in detail below with reference to the accompanying drawings in an exemplary and non-limiting manner. The same reference numerals in the drawings designate the same or similar parts or components. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings:

[0023] Figure 1This is a schematic diagram of the dynamic road network collaborative expansion decision system based on federated learning and digital twins of the present invention.

[0024] Figure 2 This is a schematic diagram of the dynamic road network collaborative expansion decision-making method based on federated learning and digital twins of the present invention.

[0025] Figure reference numerals: 11. Fusion module; 12. Reconstruction module; 13. Expansion module; 14. Multi-objective module; 15. Dynamic closed-loop module. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort should fall within the scope of protection of the present application.

[0027] The following embodiments of this application use a dynamic road network collaborative expansion decision system based on federated learning and digital twins as an example to illustrate the solution of this application in detail. However, this embodiment does not limit the scope of protection of this application.

[0028] like Figure 1 As shown, this invention provides a dynamic road network collaborative expansion decision system based on federated learning and digital twins, comprising:

[0029] The fusion module 11 is used for each traffic management domain to input spatiotemporally heterogeneous traffic feature data into the local client, classifying it into pulse-type and periodic data streams; calibrating the weights of pulse-type and periodic data streams through a dynamic weight allocator to generate normalized feature vectors; and fusing the normalized feature vectors of each domain through a privacy-preserving aggregation engine to obtain a passenger flow pressure distribution prediction matrix.

[0030] The reconstruction module 12 is used to construct and initialize a digital twin based on the prediction matrix; based on the initialized twin environment, the blind spot data reconstructor identifies the perception missing area, and the spatiotemporal correlation algorithm is used to fuse historical features with real-time data streams from neighboring nodes to reconstruct the complete road network status.

[0031] The expansion module 13 is used to input the reconstructed complete road network state into the digital twin, execute multimodal expansion strategy simulation in parallel, and generate a set of expansion strategy simulation results.

[0032] The multi-objective module 14 is used to input the simulation result set of the expanded strategy into the Pareto front optimization algorithm to obtain the multi-objective value set of the strategy; and to perform benchmark mode and perturbation mode evaluation based on the multi-objective value set of the strategy to generate the Pareto optimal strategy combination.

[0033] The dynamic closed-loop module 15 is used to input the Pareto optimal policy combination into the policy compilation engine, generate the execution instruction sequence, and parse the preset administrative delay feature matrix to dynamically trigger the instruction node that needs approval; inject the execution status of the instruction node into the digital twin, and obtain the execution deviation through the deviation detector; when the execution deviation exceeds the adaptive tolerance boundary, trigger the perturbation mode re-optimization engine preset in the digital twin to generate a correction policy; call the update interface of the policy compilation engine to load the correction policy, forming a dynamic closed loop.

[0034] In this embodiment of the invention, by classifying pulse-type and periodic data streams and combining them with a dynamic weight allocator to calibrate the influence weights of different types of data, fusion bias caused by differences in data characteristics is avoided, making the normalized feature vector more closely match the actual state of the road network; data privacy protection is strengthened, meeting the needs of cross-departmental collaboration and complying with data security compliance requirements; the digital twin is initialized based on the prediction matrix to ensure that the twin environment is highly consistent with the initial state of the real road network, laying the foundation for "virtual-real synchronization"; data gaps in perception blind spots are filled, improving the twin's ability to fully map the physical world; and a multimodal expansion strategy is executed in parallel by the digital twin. Simulation improves decision-making efficiency, reduces the cost of trial and error in the physical world, and avoids problems caused by immature strategies. Pareto front optimization algorithms are used to process the multi-objective value set of expansion strategies, avoiding the unintended consequences of single-objective optimization and ensuring the strategy achieves optimal overall benefits. By analyzing a pre-set administrative delay feature matrix, instruction nodes requiring approval are dynamically triggered, taking into account differences in administrative processes across departments, avoiding strategy failures due to asynchronous execution, and improving cross-departmental collaboration efficiency. Injecting instruction execution status into a digital twin and using a deviation detector to monitor execution deviation in real time ensures the strategy can be dynamically adjusted based on actual execution results.

[0035] In this embodiment of the invention, the aforementioned fusion module 11 receives spatiotemporally heterogeneous traffic feature data from local clients for each traffic management domain, classifying it into pulse-type and periodic data streams; it calibrates the weights of the pulse-type and periodic data streams using a dynamic weight allocator to generate normalized feature vectors; and it fuses the normalized feature vectors from each domain using a privacy-preserving aggregation engine to obtain a passenger flow pressure distribution prediction matrix, including:

[0036] Step 111: The local client of each traffic management domain receives spatiotemporally heterogeneous traffic feature data streams and identifies the dynamic patterns of the data streams, including pulse-type data streams and periodic data streams. Specifically: First, the local client corresponding to each traffic management domain receives spatiotemporally heterogeneous traffic feature data streams from its management scope. These data streams contain various dynamic information related to traffic operations. Next, the local client analyzes the received spatiotemporally heterogeneous traffic feature data streams to identify the dynamic patterns presented by the data streams. Specifically, it distinguishes them into pulse-type data streams and periodic data streams. Pulse-type data streams mainly exhibit sudden, high-intensity data fluctuations that occur within a short period of time, while periodic data streams exhibit regular and repetitive change characteristics.

[0037] Step 112: Based on the pulse-type and periodic data streams, adaptive weight calibration is performed using a dynamic weight allocator to generate a normalized feature vector. Specifically, after identifying the pulse-type and periodic data streams in Step 111, the local client invokes the dynamic weight allocator. This dynamic weight allocator assigns corresponding weights to the pulse-type and periodic data streams according to their respective characteristics and the actual needs of the current traffic scenario, and performs adaptive adjustment and calibration to balance the impact of different types of data streams in subsequent processing. Specifically, the dynamic weight allocator... and Weights are assigned to pulse-type data streams (I) and periodic data streams (P), where, It is the weight of the pulse data stream (I). It is the weight of the periodic data stream (P); B(I) represents the burst intensity of the pulse data stream within the recent time window, which is the ratio of the peak value to the mean value; The stability index for periodic data streams is the inverse of the data's variance. α and β are adjustable parameters that are dynamically adjusted based on the current traffic scenario. For example, during morning rush hour, when there is a sudden increase in passenger flow, the system automatically increases the value of α to enhance the weight of pulse-type data; while during off-peak hours, the value of β is increased to strengthen the contribution of periodic data. Through this mechanism, the system can adaptively balance the influence of the two types of data streams, generating a more representative normalized feature vector. Through this weight calibration process, pulse-type and periodic data streams are converted into parameters with uniform dimensions and comparable scales, thereby generating a normalized feature vector. This allows data streams of different types and magnitudes to be compared and integrated on the same dimension.

[0038] Step 113: Receive normalized feature vectors from each domain, inject random noise that satisfies differential privacy requirements into each feature value to generate a noise-perturbed feature vector; based on the noise-perturbed feature vector, map the features of the heterogeneous traffic domains to a unified road network coordinate system; calculate the road network topology distance between the source domain nodes and the target domain nodes; allocate coordinate transformation weights according to the road network topology distance to generate a coordinate system-aligned feature matrix. Specifically: the system receives normalized feature vectors from each traffic management domain; to ensure data privacy and security, the system injects random noise into the feature values ​​of each of the above normalized feature vectors, and the injected random noise satisfies differential privacy requirements to generate a noise-perturbed feature vector.

[0039] Based on the noise-perturbed feature vector, the system maps the feature information of various heterogeneous traffic domains (such as railways, subways, municipal transportation, etc.) to a preset unified road network coordinate system to eliminate the differences in spatial coordinate systems between different traffic management domains. On this basis, the system calculates the road network topology distance between the source domain nodes (i.e., the road network nodes corresponding to the data source) and the target domain nodes (i.e., the target road network nodes that need to be fused with data) in the unified road network coordinate system. This distance can reflect the degree of correlation between different nodes in the road network structure.

[0040] Based on the calculated road network topology distance, corresponding coordinate transformation weights are assigned to each node. Nodes that are closer usually have a higher degree of correlation, and their corresponding coordinate transformation weights are also adaptively adjusted. The feature information is adjusted through these coordinate transformation weights, and finally a coordinate-aligned feature matrix is ​​generated.

[0041] Step 114: Based on the coordinate-aligned feature matrix, the system allocates aggregation weights according to the proportion of data volume in each domain to the total data volume of the entire domain, and performs weighted summation calculation to obtain the road network-level passenger flow pressure distribution prediction matrix. Each element of the prediction matrix represents the passenger flow pressure index of a specific grid area in the road network. Specifically, after obtaining the coordinate-aligned feature matrix in step 113, the system further aggregates the feature information of each traffic management domain. Specifically, the system counts the data volume of each traffic management domain and calculates its proportion to the total data volume of all traffic management domains. Based on this proportion, the system allocates corresponding aggregation weights to the feature matrix of each traffic management domain. The traffic management domain with a larger proportion of data volume usually has a higher influence weight in the aggregation process to reflect the representativeness of its data.

[0042] Next, the system performs a weighted summation calculation on the feature matrix aligned to the above coordinate system according to the assigned aggregation weights. Through this calculation process, the feature information of each traffic management domain is fused to obtain a road network-level passenger flow pressure distribution prediction matrix. Each element in this passenger flow pressure distribution prediction matrix corresponds to the passenger flow pressure index of a specific grid area in the road network, which can intuitively reflect the passenger flow load status of different areas.

[0043] In this embodiment of the invention, by having local clients in each traffic management domain perform dynamic pattern recognition on the received spatiotemporally heterogeneous traffic feature data streams, classifying them into pulse-type and periodic data streams, this classification method lays the foundation for subsequent differentiated processing strategies for data streams with different characteristics. An adaptive weight calibration of the pulse-type and periodic data streams using a dynamic weight allocator allows for flexible adjustment of the influence weights of the two types of data streams based on real-time changes in the traffic scenario. This dynamic adjustment mechanism ensures that the generated normalized feature vectors highlight the role of key data while balancing the contributions of different types of data. Injecting random noise that meets differential privacy requirements into the normalized feature vectors effectively protects the original data privacy of each traffic management domain without affecting data availability, preventing the leakage of sensitive information. Mapping heterogeneous traffic domain features to a unified road network coordinate system and allocating coordinate transformation weights based on road network topology distance eliminates the differences in spatial coordinate systems between different traffic management domains, enabling precise alignment of originally scattered feature information in the spatial dimension. Allocating aggregation weights based on the proportion of data volume in each domain and performing weighted summation calculations enriches the data volume, thereby improving the reliability of the fusion results.

[0044] In this embodiment of the invention, the reconstruction module 12 constructs and initializes a digital twin based on a prediction matrix; based on the initialized twin environment, it identifies areas with missing perception through a blind spot data reconstructor, and uses a spatiotemporal correlation algorithm to fuse historical features with real-time data streams from neighboring nodes to reconstruct the complete road network state, including:

[0045] Step 121: Based on the passenger flow pressure distribution prediction matrix, construct a high-fidelity digital twin of the target road network; map the passenger flow pressure index of each grid area in the prediction matrix to the corresponding spatial node position of the digital twin to generate an initial road network state mirror. Specifically, based on the passenger flow pressure distribution prediction matrix, construct a high-fidelity digital twin of the target road network. The construction process of the high-fidelity digital twin of the target road network is as follows:

[0046] Based on the actual physical space of the target road network, basic geographic information data of the road network is collected, including the geometric shape of the road (such as the radius of curvature of straight and curved sections), road level (such as arterial roads, secondary arterial roads, and branch roads), number and width of lanes, intersection layout (such as cross intersections and T-junctions), traffic signs and markings (such as stop lines, directional arrows, and no-overtaking signs), and other static spatial features. Based on this, a spatial skeleton model of the digital twin is built to ensure that the skeleton model is completely consistent with the physical road network in terms of spatial topology.

[0047] The dynamic traffic characteristic parameters of the road network are integrated, including the design speed, traffic capacity (such as the maximum number of vehicles allowed to pass per unit time), road material (such as asphalt pavement and cement pavement, used to simulate the difference in vehicle driving resistance), and surrounding traffic ancillary facilities (such as bus stop locations, taxi stands, and parking lot entrances and exits). These parameters are then associated with the corresponding locations in the spatial skeleton model, so that the digital twin not only has a mapping of physical form, but also reflects the traffic function attributes of the road network.

[0048] Next, based on the spatial skeleton model and associated traffic characteristic parameters built above, and combined with the grid division rules of the passenger flow pressure distribution prediction matrix, the spatial range of the digital twin is divided according to the same grid scale, so that each grid in the twin and the grid area in the prediction matrix form a one-to-one spatial mapping relationship, ensuring the accuracy of subsequent data mapping.

[0049] Subsequently, the spatial skeleton model, traffic feature parameters, grid segmentation information and association rules are integrated through a preset interface to form a complete digital model containing a spatial coordinate system, attribute database and dynamic interaction rules. This model can respond to external data input and update its own state in real time, thereby achieving accurate reproduction of the target road network physical structure and traffic features, and finally constructing a high-fidelity digital twin that can accurately reflect the target road network spatial structure and traffic features.

[0050] The passenger flow pressure index corresponding to each grid area in the passenger flow pressure distribution prediction matrix is ​​mapped one by one to the corresponding spatial node position in the digital twin. Through this mapping process, an initial road network state mirror corresponding to the initial state of the target road network is generated.

[0051] Step 122: Based on the initial road network state image, the blind spot data reconstructor detects node areas in the road network where no real-time sensors are deployed, identifies nodes that have not received data reports for multiple consecutive time periods, and obtains a list of nodes with missing perception areas. Specifically: Based on the initial road network state image generated above, the blind spot data reconstructor is activated to detect node areas in the target road network to identify node areas where no real-time sensors are deployed; at the same time, the blind spot data reconstructor also monitors the data reporting status of each node and identifies nodes that have not received data reports for multiple consecutive time periods; through the above detection and identification operations, a list of nodes with missing perception areas is finally obtained, which clearly records all nodes with missing data perception problems.

[0052] Step 123: For each target node in the node list of the missing perception area, dynamically calculate the real-time data nodes within a set radius based on the road network topology to obtain the set of neighboring nodes of the target node; based on the target node currently being processed, calculate its average value of data in the same historical period, and assign historical feature weight values ​​in combination with the time decay law. Specifically: for each target node in the node list of the missing perception area, firstly, based on the road network topology of the target road network, dynamically calculate all nodes that can provide real-time data within a set radius centered on the target node, and take these nodes as the neighboring nodes of the target node, thus forming the set of neighboring nodes of the target node; at the same time, for the target node currently being processed, retrieve its historical traffic data, calculate the average traffic data of the node in the same historical period, and assign corresponding historical feature weight values ​​to the historical feature in combination with the preset time decay law. The setting of the time decay law can reflect the influence of recent historical data on the current node status.

[0053] Step 124: Based on the set of neighboring nodes, calculate the spatial correlation weight value of the actual road network distance between each neighboring node and the target node. Specifically, based on the set of neighboring nodes obtained in step 123, calculate the actual road network distance between each neighboring node and the target node being processed, and assign a corresponding spatial correlation weight value to each neighboring node according to the distance of the actual road network distance. In particular, the closer the neighboring node is to the target node, the higher its corresponding spatial correlation weight value is usually, so as to reflect the stronger spatial correlation between the two.

[0054] Step 125: Multiply the historical feature mean by the historical feature weight value to obtain the historical contribution; multiply the real-time data of each neighboring node by the corresponding spatial correlation weight value and sum them to obtain the real-time data contribution; add the historical contribution and the real-time data contribution and divide by the sum of all weight values ​​to normalize and generate the target node state. Specifically: multiply the historical feature mean obtained in step 123 by the corresponding historical feature weight value to obtain the historical contribution of the target node; then, multiply the real-time data of each neighboring node by the corresponding spatial correlation weight value obtained in step 124 and sum these products to obtain the real-time data contribution of the target node; finally, add the historical contribution and the real-time data contribution and divide the sum by the sum of the historical feature weight value and all spatial correlation weight values. Through this normalization process, the state data of the target node is finally generated.

[0055] Step 126: After all target node states have been reconstructed, the reconstructed node states are merged with the real-time data node states to generate a reconstructed complete road network state. Specifically, after all target node states in the node list of the missing perception area have been reconstructed, these reconstructed node states are merged with the node states in the target road network that were originally able to obtain real-time data normally. Through this merging operation, the node states in the missing perception area are completed, thereby generating a reconstructed complete road network state covering all nodes of the target road network.

[0056] In this embodiment of the invention, constructing a high-fidelity digital twin based on a prediction matrix and generating an initial road network state mirror allows the digital twin environment to accurately correspond to the initial passenger flow pressure distribution of the target road network, providing a realistic and reliable virtual foundation for subsequent dynamic simulation and analysis of the road network state. By identifying areas with missing perception through a blind spot data reconstructor, gaps in data monitoring within the road network can be accurately located, avoiding information omissions caused by insufficient sensor deployment or data transmission problems. Calculating the average historical data for nodes with missing perception and assigning weights based on time decay patterns fully utilizes the patterns inherent in historical data, while time decay reflects recent historical data. This provides higher reference value; calculating spatial correlation weights based on the actual distance between neighboring nodes and the target node reflects the characteristic that the closer the nodes are geographically, the stronger their data correlation, providing a weight basis for subsequently using real-time data from neighboring nodes to fill in the target node status; by integrating historical contribution and real-time data contribution and normalizing them to generate the target node status, it can comprehensively consider historical patterns and surrounding real-time conditions, making the reconstructed perception-missing node status closer to reality and effectively making up for the lack of information caused by data loss; merging the reconstructed node status with the real-time data node status can form a complete road network status, ensuring that the information of all nodes in the road network can be accurately reflected.

[0057] In this embodiment of the invention, the aforementioned expansion module 13 reconstructs the complete road network state input into the digital twin, executes multimodal expansion strategy simulation in parallel, and generates a set of expansion strategy simulation results, including:

[0058] Step 131: Based on the reconstructed complete road network state, a multimodal expansion strategy library is preloaded into the digital twin, including infrastructure expansion strategies, traffic control strategies, and resource scheduling strategies, to obtain an initial strategy set. Specifically, based on the reconstructed complete road network state obtained in step 126, a multimodal expansion strategy library is preloaded into the digital twin. This strategy library contains various types of expansion strategies. Among them, infrastructure expansion strategies may include temporarily widening lanes and opening emergency lanes, traffic control strategies may include adjusting traffic light timings and restricting the passage of specific vehicles, and resource scheduling strategies may include increasing shuttle vehicles and allocating police force for traffic control. By incorporating the above-mentioned strategies into the strategy library, an initial strategy set is obtained.

[0059] Step 132: For each strategy in the initialization strategy set, inject real-time traffic flow data from the reconstructed road network state into the digital twin. Iterate and deduce the strategy execution effect according to the simulation clock step size to obtain the strategy spatiotemporal evolution trajectory. Specifically, for each strategy in the initialization strategy set, inject the real-time traffic flow data (such as vehicle speed, traffic volume, and traffic density) contained in the reconstructed complete road network state into the digital twin so that the digital twin can simulate the traffic flow operation state of the real road network. Then, according to the preset simulation clock step size (such as 30 seconds per step size), iterate and deduce the execution process of each strategy in the digital twin. That is, by simulating the changes in traffic flow after the implementation of the strategy step by step, record the state parameters (such as congestion level and vehicle distribution) of each area of ​​the road network at different time nodes, and thus obtain the strategy spatiotemporal evolution trajectory corresponding to each strategy.

[0060] Step 133: Based on the spatiotemporal evolution trajectory of the strategy, calculate the congestion reduction rate, resource consumption coefficient, and risk exposure to obtain a three-dimensional indicator set for the strategy. Specifically, after obtaining the spatiotemporal evolution trajectory of the strategy, calculate the three-dimensional performance indicators corresponding to each strategy based on the information recorded in the trajectory: First, calculate the congestion reduction rate, which is obtained by comparing the average vehicle speed of the road network after the implementation of the strategy with the baseline vehicle speed to measure the effect of the strategy on alleviating congestion; Second, calculate the resource consumption coefficient, which integrates the labor costs, equipment depreciation costs, and energy consumption involved in the implementation of the strategy, and generates a comprehensive cost index through weighted processing to reflect the resource input of the strategy; Third, calculate the risk exposure, which combines accident records of similar scenarios in the historical accident database to quantify the probability that the strategy may cause secondary risks; Through the above calculations, obtain the three-dimensional indicator set corresponding to each strategy.

[0061] Step 134: Aggregate the three-dimensional indicator groups of all strategies and construct a structured database containing strategy IDs, indicator values, and timestamps to obtain the expansion strategy simulation result set. Specifically, this involves aggregating the three-dimensional indicator groups of all strategies to construct a structured database. This database contains a unique strategy ID corresponding to each strategy, the values ​​of each indicator in the three-dimensional indicator group of that strategy (i.e., the specific values ​​of congestion decay rate, resource consumption coefficient, and risk exposure), and a timestamp corresponding to the simulation process of that strategy (used to identify the time information of the strategy simulation). By constructing the above structured database, the expansion strategy simulation result set is finally obtained.

[0062] In this embodiment of the invention, a pre-loaded multimodal expansion strategy library, covering various strategies such as infrastructure, traffic control, and resource scheduling, provides rich alternatives for simulation, ensuring the comprehensiveness of strategy evaluation and avoiding the limitations of a single type of strategy. Injecting real-time traffic flow data into the twin and extrapolating strategy effects according to clock steps realistically simulates the execution process of strategies in actual road networks, accurately capturing the evolution of strategies over time and space, and providing reliable trajectory data for subsequent indicator calculations. By calculating congestion attenuation rate, resource consumption coefficient, and risk exposure to form a three-dimensional indicator set, the merits of strategies can be quantified from three core dimensions: mitigation effect, cost, and risk, providing clear and comparable standards for strategy evaluation. Aggregating all the three-dimensional indicator sets of strategies and constructing a structured database systematizes and organizes the scattered simulation results, facilitating rapid access and analysis of subsequent multi-objective optimization algorithms and improving decision-making efficiency.

[0063] In this embodiment of the invention, the multi-objective module 14 inputs the expanded policy simulation result set into the Pareto front optimization algorithm to obtain a policy multi-objective value set; based on the policy multi-objective value set, it performs baseline mode and perturbation mode evaluation to generate a Pareto optimal policy combination, including:

[0064] Step 141: Based on the simulation result set of the expansion strategy, the Pareto front optimization algorithm is used to calculate the objectives of maximizing the congestion decay rate, minimizing the resource consumption coefficient, and minimizing the risk exposure for each strategy, resulting in a multi-objective value set for the strategy. Specifically, based on the aforementioned simulation result set of the expansion strategy, the Pareto front optimization algorithm is called to perform multi-objective calculations on each strategy in the result set. For each strategy, maximizing the congestion decay rate is the primary objective, while minimizing the resource consumption coefficient and minimizing the risk exposure are secondary objectives. The algorithm is used to perform synergistic optimization on the above three objectives, thereby obtaining a multi-objective value set that reflects the performance of each strategy in the above three objective dimensions.

[0065] Step 142: Based on the multi-objective value set of the strategy, perform baseline modal and disturbance modal evaluations to obtain a dual-modal evaluation result. Specifically, after obtaining the multi-objective value set of the strategy, perform baseline modal evaluation and disturbance modal evaluation based on the data in the value set. Baseline modal evaluation refers to analyzing the multi-objective values ​​of each strategy under normal traffic scenarios to evaluate its performance under normal operating conditions. Disturbance modal evaluation refers to analyzing the multi-objective values ​​of each strategy under preset extreme traffic scenarios (such as sudden accidents, severe weather, etc.) to evaluate its adaptability under abnormal conditions. Through the above two modal evaluations, the final dual-modal evaluation result is obtained.

[0066] Baseline Modal Evaluation: Under typical traffic scenarios, the system uses historical average traffic flow and standard vehicle speed distribution as inputs to calculate the congestion reduction rate for each strategy. Resource consumption coefficient Risk exposure Performance on three indicators. For example ,in, and Representing the average vehicle speed before and after the strategy implementation; and the resource consumption coefficient. It is a comprehensive cost indicator used to quantify the total amount of various resources required to implement a certain expansion strategy. The resource consumption coefficient is equal to the labor cost C1 multiplied by its weight coefficient W1, plus the equipment cost C2 multiplied by its weight coefficient W2, plus the energy cost C3 multiplied by its weight coefficient W3. Here, the labor cost C1 refers to the labor cost incurred in implementing the strategy, the equipment cost C2 refers to the cost incurred in using, depreciating or leasing related equipment, and the energy cost C3 refers to the energy cost such as electricity and fuel consumed in the process of implementing the strategy. W1, W2 and W3 are the weight coefficients of the above three costs, which are used to reflect the relative importance of different cost dimensions in the overall evaluation, and the sum of the three weight coefficients is 1. They can be set according to historical data.

[0067] Disturbance Modal Assessment: The system presets extreme scenarios (such as a sudden accident on a main road, heavy rain, etc.), and reruns the digital twin simulation under these scenarios to calculate the index values ​​of each strategy. For example, in the heavy rain scenario, traffic capacity decreases, and the system adjusts the simulation parameters and recalculates, using the following formula: ,in, It is the degree of risk exposure. It is the accident type number or category index. This represents the probability of a certain type of accident occurring under perturbation conditions. Indicating the severity of an accident, it is a dimensionless quantitative value. A benchmark value is set through standards established by the relevant authorities, transforming the consequences of the accident into a calculable numerical value. It is typically assessed comprehensively from three dimensions: personal injury, economic loss, and duration of traffic disruption. For example, a common definition is to set a benchmark value: property damage only (minor scratches). = 10; resulting in minor personal injury or moderate property damage: = 30; resulting in serious personal injury or significant property damage, or traffic disruption for 1 hour: = 60; resulting in death or prolonged traffic disruption (>3 hours): = 100.

[0068] Step 143: Based on the bimodal evaluation results, generate a Pareto optimal strategy combination that satisfies multiple constraints, including resource consumption not exceeding a preset budget threshold, risk exposure not exceeding a safety threshold, and congestion decay rate not lower than the minimum improvement requirement. Specifically, based on the aforementioned bimodal evaluation results, further select strategies that satisfy multiple constraints. These multiple constraints specifically include: resource consumption coefficient not exceeding a preset budget threshold, risk exposure not exceeding a preset safety threshold, and congestion decay rate not lower than the preset minimum improvement requirement. Combining strategies that simultaneously satisfy all the above constraints generates a Pareto optimal strategy combination.

[0069] In this embodiment of the invention, the Pareto front optimization algorithm is used to perform multi-objective calculations on the strategies in the simulation result set of capacity expansion strategies, focusing on maximizing congestion decay rate, minimizing resource consumption coefficient, and minimizing risk exposure. This comprehensively considers the performance of the strategies in the three core dimensions of effectiveness, cost, and risk, avoiding decision bias caused by single-objective optimization. It provides a multi-objective value set of strategies that takes into account multiple dimensions for subsequent evaluation, ensuring the comprehensiveness of strategy selection. Based on the multi-objective value set of strategies, baseline mode (normal scenario) and disturbance mode (extreme scenario) evaluations are performed to verify the adaptability and stability of the strategies under different traffic scenarios. This dual-modal evaluation method can effectively identify strategies that perform well under normal conditions but whose effectiveness drops sharply under sudden disturbances, ensuring that the selected strategies are not only applicable to daily situations but also able to cope with extreme scenarios, improving the universality and reliability of the strategies. Based on the dual-modal evaluation results, a Pareto optimal strategy combination that meets multiple constraints (resource consumption, risk exposure, and congestion decay rate are all within reasonable thresholds) is generated, which can ensure that the strategies achieve optimal comprehensive benefits while strictly controlling costs and risks within an acceptable range.

[0070] In this embodiment of the invention, the dynamic closed-loop module 15 inputs the Pareto optimal strategy combination into the strategy compilation engine to generate an execution instruction sequence, and parses the preset administrative delay feature matrix to dynamically trigger the instruction node requiring approval; it injects the execution state of the instruction node into the digital twin and obtains the execution deviation through a deviation detector; when the execution deviation exceeds the adaptive tolerance boundary, it triggers the perturbation mode re-optimization engine preset in the digital twin to generate a correction strategy; it calls the update interface of the strategy compilation engine to load the correction strategy, forming a dynamic closed loop, including:

[0071] Step 151: Input the Pareto optimal policy combination into the policy compilation engine to generate a spatiotemporally aligned execution instruction sequence. Specifically, the Pareto optimal policy combination obtained in step 143 is input into the policy compilation engine, which parses and transforms the Pareto optimal policy combination to generate a spatiotemporally aligned execution instruction sequence. The spatiotemporal alignment means that each execution instruction maintains coordination and consistency in both time nodes and spatial regions to ensure the orderliness of instruction execution.

[0072] Step 152: Based on the spatiotemporally aligned execution instruction sequence, the administrative delay feature matrix is ​​parsed, and instruction nodes requiring approval are dynamically triggered to obtain an instruction node flow. Specifically, based on the spatiotemporally aligned execution instruction sequence generated in Step 151, a preset administrative delay feature matrix is ​​parsed. This administrative delay feature matrix contains the approval process and duration information required for various types of instructions in different traffic management domains. According to the parsing results, instruction nodes requiring approval are dynamically identified and triggered in the execution instruction sequence. Simultaneously, these instruction nodes requiring approval are integrated with those not requiring approval according to their execution sequence to obtain an instruction node flow, adapting to the actual process of cross-departmental approval. Specifically, the administrative delay feature matrix, D, is an m×n matrix, where m represents the instruction type and n represents the traffic management domain. Matrix elements This represents the approval time required for the i-th type of instruction in the j-th field. The system generates the instruction node stream according to the following steps: For each instruction k in the execution instruction sequence, based on its type t... k and target domain d k The approval delay is obtained by querying matrix D. ;like If the value is greater than 0, the instruction is marked as a node requiring approval, and its earliest executable time is set to the current time + 0. The system sorts all instructions by execution time, merges nodes requiring and not requiring approval, and forms a consistent instruction node flow. For example, if an instruction requires approval from the traffic police department with a 5-minute delay, the system will postpone the execution time of the instruction by 5 minutes and automatically trigger the approval process at that time. The instruction can only be executed after approval is granted.

[0073] Step 153: Inject the real-time status of the instruction node flow into the digital twin. Calculate the deviation between the execution trajectory and the expected trajectory using a deviation detector to obtain the execution deviation. Specifically, the real-time status of the instruction node flow during actual execution (such as instruction start time, execution progress, current effect, etc.) is injected into the digital twin. The digital twin generates the expected execution trajectory of each instruction node based on a preset expected execution trajectory model. Simultaneously, the deviation detector compares and analyzes the actual execution trajectory of the instruction node flow with the expected execution trajectory, calculates the degree of deviation between the two, and obtains the execution deviation to quantitatively reflect the difference between the actual execution and the expected target.

[0074] Step 154: When the execution deviation exceeds the adaptive tolerance boundary, the pre-installed perturbation mode re-optimization engine in the digital twin is triggered to generate a correction strategy. Specifically, the execution deviation obtained in step 153 is compared with the system's pre-installed adaptive tolerance boundary, which is an acceptable deviation range dynamically adjusted according to the actual operation of the road network. When the execution deviation exceeds the adaptive tolerance boundary, the pre-installed perturbation mode re-optimization engine in the digital twin is automatically triggered. The perturbation mode re-optimization engine generates a targeted correction strategy based on the current road network status and the cause of the deviation to compensate for the execution deviation.

[0075] Step 155: Call the update interface of the strategy compilation engine to load the correction strategy and form a dynamic closed loop. Specifically, call the update interface of the strategy compilation engine to load the correction strategy generated in step 154 ​​into the execution instruction sequence, dynamically update and adjust the original instruction sequence so that the corrected instruction sequence can adapt to changes in actual execution, thereby forming a dynamic closed loop from strategy generation, execution, monitoring to correction.

[0076] In this embodiment of the invention, the Pareto optimal strategy combination is input into the strategy compilation engine to generate a spatiotemporally aligned execution instruction sequence. This ensures that the instructions are matched and coordinated in terms of timing and spatial deployment, avoiding execution chaos caused by instruction timing or spatial conflicts, and providing a clear and orderly operational basis for subsequent cross-departmental collaborative execution. Based on the spatiotemporally aligned execution instruction sequence, the administrative delay feature matrix is ​​analyzed, and the nodes requiring approval are dynamically triggered. This fully considers the differences in administrative process time consumption among different departments, ensuring that the approval process for nodes requiring approval is initiated at the appropriate time, reducing strategy execution delays caused by approval delays, and improving the timeliness and collaborative efficiency of instruction execution. By injecting the real-time status of node flows into the digital twin and calculating the execution deviation through a deviation detector, the deviation between the strategy execution process and the expected goal can be monitored in real time. This allows for the timely detection of anomalies during execution (such as instructions not proceeding as planned or results not meeting expectations), providing a basis for subsequent adjustments. When the execution deviation exceeds the adaptive tolerance boundary, the perturbation mode re-optimization engine is triggered to generate a correction strategy. This can quickly respond to the execution deviation and adjust the strategy accordingly to compensate for the deviation, avoiding strategy failure caused by the accumulation of small deviations. The update interface of the strategy compilation engine is called to load the correction strategy, forming a dynamic closed loop. This enables adaptive adjustment of the strategy throughout the entire process from generation, execution, monitoring to correction.

[0077] like Figure 2 As shown, a dynamic road network collaborative expansion decision-making method based on federated learning and digital twins is described, the method comprising:

[0078] Each traffic management domain inputs spatiotemporally heterogeneous traffic characteristic data into its local client, classifying it into pulse-type and periodic data streams; the weights of the pulse-type and periodic data streams are calibrated through a dynamic weight allocator to generate normalized feature vectors; and the normalized feature vectors from each domain are fused through a privacy-preserving aggregation engine to obtain a passenger flow pressure distribution prediction matrix.

[0079] A digital twin is constructed and initialized based on the prediction matrix; based on the initialized twin environment, a blind spot data reconstructor is used to identify the perception-missing areas, and a spatiotemporal correlation algorithm is used to fuse historical features with real-time data streams from neighboring nodes to reconstruct the complete road network status.

[0080] The complete road network state is reconstructed and input into the digital twin. Multimodal expansion strategy simulation is performed in parallel to generate a set of expansion strategy simulation results.

[0081] The simulation results set of the expanded strategy is input into the Pareto front optimization algorithm to obtain the multi-objective value set of the strategy; the baseline mode and perturbation mode evaluation are performed based on the multi-objective value set of the strategy to generate the Pareto optimal strategy combination;

[0082] The Pareto optimal policy combination is input into the policy compilation engine to generate an execution instruction sequence and parse the pre-set administrative delay feature matrix to dynamically trigger the instruction nodes that require approval. The execution status of the instruction nodes is injected into the digital twin, and the execution deviation is obtained through the deviation detector. When the execution deviation exceeds the adaptive tolerance boundary, the perturbation mode re-optimization engine pre-set in the digital twin is triggered to generate a correction policy. The update interface of the policy compilation engine is called to load the correction policy, forming a dynamic closed loop.

[0083] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the system as described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0084] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the system as described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A dynamic road network collaborative expansion decision-making system based on federated learning and digital twins, characterized in that, include: The fusion module is used by various traffic management domains to input spatiotemporally heterogeneous traffic characteristic data into local clients, classifying it into pulse-type and periodic data streams; The pulse and periodic data stream weights are calibrated by a dynamic weight allocator to generate normalized feature vectors; the normalized feature vectors from each domain are then fused by a privacy-preserving aggregation engine to obtain a passenger flow pressure distribution prediction matrix. The reconstruction module is used to construct and initialize a digital twin based on the prediction matrix; based on the initialized twin environment, the blind spot data reconstructor identifies the perception missing area, and the spatiotemporal correlation algorithm is used to fuse historical features with real-time data streams from neighboring nodes to reconstruct the complete road network status. The expansion module is used to input the reconstructed complete road network state into the digital twin, execute multimodal expansion strategy simulation in parallel, and generate a set of expansion strategy simulation results. The multi-objective module is used to input the simulation result set of the expansion strategy into the Pareto front optimization algorithm to obtain the multi-objective value set of the strategy. Based on the policy multi-objective value set, perform baseline mode and perturbation mode evaluation to generate Pareto optimal policy combination; The dynamic closed-loop module is used to input the Pareto optimal strategy combination into the strategy compilation engine, generate the execution instruction sequence, and parse the preset administrative delay feature matrix to dynamically trigger the instruction nodes that require approval. The execution state of the instruction node is injected into the digital twin, and the execution deviation is obtained through the deviation detector. When the execution deviation exceeds the adaptive tolerance boundary, the perturbation mode re-optimization engine pre-placed in the digital twin is triggered to generate a correction strategy. The update interface of the strategy compilation engine is called to load the correction strategy, forming a dynamic closed loop.

2. The dynamic road network collaborative expansion decision-making system based on federated learning and digital twins as described in claim 1, characterized in that, Each traffic management domain inputs spatiotemporally heterogeneous traffic characteristic data into its local client, classifying it into pulse-type and periodic data streams; the weights of the pulse-type and periodic data streams are calibrated through a dynamic weight allocator to generate normalized feature vectors; By fusing normalized feature vectors from various domains using a privacy-preserving aggregation engine, a passenger flow pressure distribution prediction matrix is ​​obtained, including: Local clients in each traffic management domain receive spatiotemporally heterogeneous traffic characteristic data streams and identify the dynamic patterns of the data streams, including pulse-type data streams and periodic data streams. Based on pulse-type and periodic data streams, a normalized feature vector is generated by adaptive weight calibration through a dynamic weight allocator. The system receives normalized feature vectors from each domain, injects random noise that satisfies differential privacy requirements into each feature value, and generates a noise-perturbed feature vector. Based on the noise-perturbed feature vector, the features of the heterogeneous traffic domains are mapped to a unified road network coordinate system. The road network topology distance between the source domain nodes and the target domain nodes is calculated. Coordinate transformation weights are assigned according to the road network topology distance to generate a coordinate system aligned feature matrix. Based on the feature matrix aligned to the coordinate system, aggregation weights are allocated according to the proportion of data volume in each domain to the total data volume of the entire domain, and weighted summation calculation is performed to obtain the road network-level passenger flow pressure distribution prediction matrix; each element of the prediction matrix represents the passenger flow pressure index of the grid area in the road network.

3. The dynamic road network collaborative expansion decision-making system based on federated learning and digital twins according to claim 2, characterized in that, A digital twin is constructed and initialized based on the prediction matrix; based on the initialized twin environment, a blind spot data reconstructor identifies perceptually missing areas, including: Based on the passenger flow pressure distribution prediction matrix, a high-fidelity digital twin of the target road network is constructed; the passenger flow pressure index of each grid area in the prediction matrix is ​​mapped to the corresponding spatial node position of the digital twin to generate an initial road network state mirror. Based on the initial road network state image, the blind spot data reconstructor detects the node areas in the road network where no real-time sensors are deployed, identifies nodes that have not received data reports for multiple consecutive time periods, and obtains a list of nodes in the perception missing area.

4. The dynamic road network collaborative expansion decision-making system based on federated learning and digital twins according to claim 3, characterized in that, It employs a spatiotemporal correlation algorithm to fuse historical features with real-time data streams from neighboring nodes, reconstructing the complete road network status, including: For each target node in the node list of the missing perception area, the real-time data nodes within a set radius are dynamically calculated according to the road network topology to obtain the set of neighboring nodes of the target node; based on the target node being processed, the average value of its data in the same historical period is calculated, and historical feature weight values ​​are assigned in combination with the time decay law. Based on the set of neighboring nodes, calculate the spatial correlation weight value of the actual road network distance between each neighboring node and the target node; The historical contribution is obtained by multiplying the historical feature mean by the historical feature weight; the real-time data of each neighboring node is multiplied by the corresponding spatial correlation weight and summed to obtain the real-time data contribution; the historical contribution and the real-time data contribution are added together and divided by the sum of all weight values ​​to normalize and generate the target node state. Once the states of all target nodes have been reconstructed, the reconstructed node states will be merged with the real-time data node states to generate the reconstructed complete road network state.

5. The dynamic road network collaborative expansion decision-making system based on federated learning and digital twins according to claim 4, characterized in that, The complete road network state is reconstructed and input into the digital twin. Multimodal expansion strategy simulations are then executed in parallel, generating a set of expansion strategy simulation results, including: Based on the reconstructed complete road network status, a multimodal expansion strategy library is preloaded into the digital twin, including: infrastructure expansion strategy, traffic control strategy and resource scheduling strategy, to obtain an initial strategy set; For each strategy in the initial strategy set, real-time traffic flow data from the reconstructed road network state is injected into the twin, and the strategy execution effect is iteratively deduced according to the simulation clock step to obtain the strategy spatiotemporal evolution trajectory. Based on the spatiotemporal evolution trajectory of the strategy, the congestion decay rate, resource consumption coefficient and risk exposure are calculated to obtain a three-dimensional index set of the strategy. Aggregate the three-dimensional indicator groups of all strategies, construct a structured database containing strategy ID, indicator value, and timestamp, and obtain the simulation result set of the expansion strategy.

6. The dynamic road network collaborative expansion decision-making system based on federated learning and digital twins according to claim 5, characterized in that, The simulation results set of the expansion strategy is input into the Pareto front optimization algorithm to obtain the multi-objective value set of the strategy; Based on the policy multi-objective value set, baseline mode and perturbation mode evaluations are performed to generate Pareto optimal policy combinations, including: Based on the simulation results set of the expansion strategy, the Pareto front optimization algorithm is used to calculate the objectives of maximizing the congestion decay rate, minimizing the resource consumption coefficient, and minimizing the risk exposure of each strategy, so as to obtain the multi-objective value set of the strategy. Based on the policy multi-objective value set, benchmark mode and perturbation mode evaluation are performed to obtain dual-mode evaluation results; Based on the bimodal evaluation results, a Pareto optimal strategy combination that satisfies multiple constraints is generated, including resource consumption not exceeding a preset budget threshold, risk exposure not exceeding a safety threshold, and congestion decay rate not lower than the minimum improvement requirement.

7. The dynamic road network collaborative expansion decision-making system based on federated learning and digital twins according to claim 6, characterized in that, The Pareto optimal strategy combination is input into the strategy compilation engine to generate an execution instruction sequence, and the preset administrative delay feature matrix is ​​parsed to dynamically trigger the instruction nodes that require approval. The execution state of the instruction node is injected into the digital twin, and the execution deviation is obtained through the deviation detector. When the execution deviation exceeds the adaptive tolerance boundary, the perturbation mode re-optimization engine pre-set in the digital twin is triggered to generate a correction strategy. The update interface of the strategy compilation engine is called to load the correction strategy, forming a dynamic closed loop, including: The Pareto optimal policy combination is input into the policy compilation engine to generate a time-space aligned sequence of execution instructions; Based on the spatiotemporally aligned execution instruction sequence, the administrative delay feature matrix is ​​analyzed, and the instruction nodes requiring approval are dynamically triggered to obtain the instruction node flow; The real-time status of the instruction node stream is injected into the digital twin, and the deviation between the execution trajectory and the expected trajectory is calculated by the deviation detector to obtain the execution deviation. When the deviation exceeds the adaptive tolerance boundary, the perturbation mode re-optimization engine pre-set in the digital twin is triggered to generate a correction strategy; The update interface of the strategy compilation engine is called to load the correction strategy, forming a dynamic closed loop.

8. A dynamic road network collaborative expansion decision-making method based on federated learning and digital twins, characterized in that, Applied to the system as described in any one of claims 1 to 7, the method comprises: Each traffic management domain inputs spatiotemporally heterogeneous traffic characteristic data into its local client, classifying it into pulse-type and periodic data streams; the weights of the pulse-type and periodic data streams are calibrated through a dynamic weight allocator to generate normalized feature vectors; and the normalized feature vectors from each domain are fused through a privacy-preserving aggregation engine to obtain a passenger flow pressure distribution prediction matrix. A digital twin is constructed and initialized based on the prediction matrix; based on the initialized twin environment, a blind spot data reconstructor is used to identify the perception-missing areas, and a spatiotemporal correlation algorithm is used to fuse historical features with real-time data streams from neighboring nodes to reconstruct the complete road network status. The complete road network state is reconstructed and input into the digital twin. Multimodal expansion strategy simulation is performed in parallel to generate a set of expansion strategy simulation results. The simulation results set of the expanded strategy is input into the Pareto front optimization algorithm to obtain the multi-objective value set of the strategy; the baseline mode and perturbation mode evaluation are performed based on the multi-objective value set of the strategy to generate the Pareto optimal strategy combination; The Pareto optimal policy combination is input into the policy compilation engine to generate an execution instruction sequence and parse the pre-set administrative delay feature matrix to dynamically trigger the instruction nodes that require approval. The execution status of the instruction nodes is injected into the digital twin, and the execution deviation is obtained through the deviation detector. When the execution deviation exceeds the adaptive tolerance boundary, the perturbation mode re-optimization engine pre-set in the digital twin is triggered to generate a correction policy. The update interface of the policy compilation engine is called to load the correction policy, forming a dynamic closed loop.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the system as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1-7.

Images