A cross-boundary road network congestion early warning and dynamic shunting guidance method

Abstract

This invention relates to the field of neural network traffic situation prediction and federated learning data processing technology, and discloses a method for cross-border road network congestion early warning and dynamic diversion guidance. This method constructs regional traffic map samples from checkpoint traffic flow, floating car speed sequences, road network topology connections, and signal phase summaries within different road management areas. Locally, each area generates internal node representations and boundary node representations using a regional private encoder and a boundary shared encoder, respectively. Under a federated learning architecture, the boundary node representation prototype, model update amount, and uncertainty weights are uploaded. Boundary consistency constraints are constructed based on the congestion level, speed decline trend, and transmission lag of boundary nodes in adjacent areas, and the resulting cross-border situation prediction model is aggregated. This method maintains consistency in boundary congestion transmission representations without cross-domain sharing of original traffic data, reducing the instability in judging the source direction and risk level of cross-border congestion.

Description

Technical Field

[0001] This invention relates to the field of neural network traffic situation prediction and federated learning data processing technology, and discloses a method for congestion early warning and dynamic diversion guidance for cross-border road networks. Background Technology

[0002] Cross-border road networks are typically composed of adjacent administrative regions, different road management entities, or different control platforms. Vehicles continuously travel between boundary passages, interchanges, arterial road connections, and toll station connections. Existing congestion warning methods mostly focus on single-region roads, with regional traffic platforms collecting data such as checkpoint vehicle records, floating car speeds, road segment flow, intersection queue lengths, and road topology relationships, forming a traffic state sequence at a fixed time granularity. The platform typically abstracts intersections, road segments, or cross-sections as nodes and road connections as edges, then uses statistical prediction models, time-series prediction models, or graph structure prediction models to infer future short-term traffic conditions. Under the conditions of complete data within the region, stable topological relationships, and unified control strategies, this type of method can generate congestion levels or warning information based on historical speed decreases, flow saturation, and reduced capacity. However, its modeling boundary is usually consistent with the data management boundary, and the model mainly relies on visible data within the region for training and inference.

[0003] In cross-regional traffic cooperation scenarios, a relatively common technical solution involves each regional platform training its own local prediction model and then exchanging some statistical results, such as boundary section traffic flow, average speed, congestion index, or warning level, through an interface. The receiving region uses the statistical values ​​provided by the neighboring region as external features to input into its local model, or, after the prediction results are formed in its own region, it makes rule-based corrections to the warning levels of the neighboring regions. This approach avoids the direct sharing of original vehicle trajectories, checkpoint details, and internal platform control data, and can meet certain data isolation requirements. However, since the exchanged content is mostly low-dimensional statistics, it usually does not include the temporal evolution characteristics, topological adjacency relationships, congestion propagation direction, and propagation time lag information within the boundary nodes. Therefore, the local model's understanding of the external regional state remains at the result level, making it difficult to transform the traffic state of the neighboring regions into a trainable and comparable cross-border congestion representation.

[0004] Some technical solutions attempt to address the problem of cross-regional data isolation using federated learning. Each region trains its model locally using raw traffic data, then uploads the model parameters or gradients to a server. The server then aggregates the data to form a global model and distributes it to each region for updates. While this invention reduces the cross-domain transmission of raw data, in traffic network scenarios, differences in road density, node granularity, sampling frequency, traffic flow composition, signal timing strategies, and the number of boundary channels across different regions lead to inconsistent sample distributions. Conventional federated aggregation typically merges model updates based on sample quantity, model loss, or fixed weights, without specifically addressing the semantic alignment of boundary nodes between adjacent regions or distinguishing between intra-regional traffic patterns and boundary propagation patterns. This can easily cause the global model to over-average intra-regional patterns and insufficiently represent boundary congestion spillover, backflow, and time-delay propagation.

[0005] The main technical problem with existing technologies is that, under the condition that raw traffic data cannot be shared across regions, the traffic state representations trained on local data by each region lack boundary consistency constraints. The congestion levels, speed decline trends, and propagation lags of boundary nodes in adjacent regional models are difficult to align, resulting in the fragmentation of the cross-border congestion propagation chain at the model level. Because the model cannot convert the boundary states of adjacent regions into latent space representations comparable to its own region, external congestion inputs are easily identified as local fluctuations or abnormal noise within the region, leading to instability in the cross-border situation prediction regarding the direction of congestion sources, risk levels, and propagation times. This problem cannot be solved simply by increasing data interfaces or model complexity; its root cause lies in the lack of a federated representation fusion and consistency calibration mechanism for boundary nodes in a data-isolated environment. Summary of the Invention

[0006] The purpose of this invention is to provide a method for early warning and dynamic diversion guidance of congestion in cross-border road networks, which can effectively solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for congestion early warning and dynamic diversion guidance in cross-border road networks includes: constructing regional traffic map samples from checkpoint traffic flow, floating car speed sequence, road network topology connectivity and signal phase summary in different road management areas; In each region, a local private encoder and a boundary shared encoder are used to generate internal node representations and boundary node representations. Upload the boundary node representation prototype, model update amount, and uncertainty weights under the federated learning architecture; Boundary consistency constraints are constructed based on the congestion level, speed decline trend, and transmission delay of adjacent regional boundary nodes; Based on the boundary consistency constraints, a cross-boundary situation prediction model is obtained, which outputs the congestion risk level of the cross-boundary road network and the direction of boundary transmission.

[0008] Preferably, the construction of the regional traffic map sample includes: determining road segment nodes, boundary connection nodes, and traversable edges between nodes according to the road management area; Write the traffic flow at the checkpoint into the node traffic sequence according to the road segment node affiliation, write the floating car speed sequence into the node speed sequence according to the trajectory matching result, and write the signal phase summary into the time slice attribute corresponding to the boundary connection node. A static topology matrix is ​​generated based on the node connection relationship, and a dynamic influence matrix is ​​generated based on the velocity change direction and flow migration direction within adjacent time slices; The static topology matrix, the dynamic influence matrix, and the node time series are collectively encapsulated into a local regional traffic map sample.

[0009] Preferably, the region private encoder and the boundary shared encoder are trained using a separate parameter update method; The regional private encoder receives internal road segment nodes, internal topological edges, and local historical time series to generate a representation of the traffic status within the region. The boundary shared encoder receives boundary connection nodes, cross-boundary adjacent edge identifiers, exchangeable statistical summaries on both sides of the boundary, and boundary time slice attributes to generate boundary node representations. In the local loss, an internal state reconstruction term, a boundary state prediction term, and a private representation isolation term are set respectively, so that the internal node representation and the boundary node representation are updated separately in the same training round.

[0010] Preferably, the construction of the boundary consistency constraint includes: for boundary connection nodes in adjacent road management areas pointing to the same cross-border passage, extracting the local congestion level distribution, speed decrease trend vector, queue growth trend vector, and transmission delay candidate values ​​for each area; Map the congestion level distribution to a unified level space, and map the speed decrease trend vector and the queue increase trend vector to a unified direction space; Based on the upstream and downstream directions of the cross-border channel, paired boundary samples are formed, and level consistency constraints, trend direction consistency constraints, and time lag reachability constraints are applied to the paired boundary samples.

[0011] Preferably, the generation of the dynamic influence matrix includes: within each regional traffic map sample, using the speed difference, flow difference, and node adjacency direction of consecutive time slices as input, calculating the local traffic influence value between node pairs; For node pairs containing boundary connection nodes, add cross-boundary direction markers and match the cross-boundary direction markers with the release direction in the signal phase summary; When the velocity difference direction and the flow difference direction of a node pair are inconsistent, the influential edge in the reachable direction is retained according to the topological reachability relationship; Write the influence values ​​of internal node pairs and boundary node pairs into the same dynamic influence matrix.

[0012] Preferably, the separation parameter update method includes: in each federated training round, first updating the parameters of the regional private encoder with the regional internal traffic status representation, and then freezing the parameters related to internal nodes in the regional private encoder; The parameters of the boundary shared encoder are then updated using the boundary node representation, boundary time slice attributes, and the boundary node representation prototype returned by the adjacent region. Before uploading the parameters, the gradient of the boundary shared encoder is desensitized and compressed at the node level, and the compressed gradient and the boundary node representation prototype are written into the federated update package.

[0013] Preferably, the determination of the candidate conduction time delay values ​​includes: performing sliding matching on the velocity decrease trend vectors of the boundary connection nodes on both sides of the same cross-border channel to obtain multiple candidate time offsets; Perform directional consistency screening on the queue growth trend vector under each candidate time offset; Candidate time offsets that satisfy the topological reachability of cross-boundary channels are retained as candidate values ​​for conduction delays; In the time-delay reachability constraint, the candidate values ​​of the conduction time delay are jointly encoded with the historical passage time distribution of the cross-border channel and the boundary signal release window.

[0014] Preferably, the aggregation under the federated learning architecture includes: receiving model update data, boundary node representation prototypes, and uncertainty weights uploaded by each road management area; The boundary deviation is calculated based on the level consistency constraint, the trend direction consistency constraint, and the time delay reachability constraint. Adjust the weight of model update amounts in each region during aggregation based on the boundary deviation. When a region lacks a boundary node representation prototype with continuous time slices, an alternative representation is generated using the boundary node representation prototype with the same cross-boundary direction in adjacent regions. The alternative representation and the current model update are then used together in the cross-boundary situation prediction model aggregation.

[0015] Preferably, the node-level desensitization compression includes: grouping the gradient of the boundary shared encoder according to the cross-boundary channel to which the boundary connection node belongs; For each set of gradients, extract the directional component, level component, and temporal component, and delete the fields corresponding to vehicle identification, original trajectory points, and checkpoint original records; The gradient components after deleting fields are prototyped and merged to form a compressed gradient that only contains cross-boundary channel identifiers, time slice identifiers, congestion level distributions, and trend direction summaries. After establishing a round index between the compression gradient and the corresponding boundary node representation prototype, the data is uploaded.

[0016] Preferably, the generation of the alternative representation includes: selecting candidate prototypes from the boundary node representation prototypes uploaded from adjacent regions that are consistent with the cross-boundary channel direction of the missing region, adjacent in time slice, and have the same level of constraint deviation less than a preset consistency range. The candidate prototypes are time-aligned based on the proposed conduction delay values; The time-aligned candidate prototypes are fused with the boundary node representation prototypes of the missing regions from the previous training round. The fused representation is written with uncertainty weights, and the uncertainty weights are corrected in the next federated training round based on the actual retransmitted boundary node representation prototype.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs regional traffic map samples from checkpoint traffic flow, floating car speed sequences, road network topology connections, and signal phase summaries in each region. Locally, it uses a region-specific encoder and a boundary-shared encoder to generate internal node representations and boundary node representations, respectively, allowing for the separation and modeling of internal traffic patterns and cross-boundary transmission patterns during training. Under the federated learning architecture, each region only uploads the boundary node representation prototype, model update amount, and uncertainty weights, without transmitting original vehicle trajectories, checkpoint details, or complete platform data. The aggregation end constructs boundary consistency constraints based on the congestion level, speed decline trend, and transmission delay of boundary nodes in adjacent regions, ensuring that the boundary states on both sides of the same cross-boundary channel are calibrated within a unified level and direction space. The resulting cross-boundary situation prediction model can retain boundary congestion transmission relationships without data leaving the domain, treating speed decline, queue growth, and time delay propagation in external regions as continuous situation inputs. This reduces the possibility of misjudging external congestion as random fluctuations within the local area, providing a more stable model basis for congestion risk levels and boundary transmission directions.

[0018] 2. This invention further enhances the usability and consistency of boundary state data during federated training by employing techniques such as dynamic influence matrices, transmission delay candidate values, node-level desensitization and compression, and alternative representation generation. The dynamic influence matrix incorporates speed differences, flow differences, node adjacency directions, cross-boundary direction markers, and signal phase summaries into the same graph sample, enabling the passage direction and time-slice attributes of boundary nodes to participate in model training. Transmission delay candidate values ​​are jointly determined by sliding matching of speed decrease trends on both sides of the cross-boundary channel, selection of queue growth trend directions, and topological reachability, avoiding judgments of congestion propagation based solely on fixed time windows. Node-level desensitization and compression remove fields related to vehicle identifiers, original trajectory points, and checkpoint records, retaining only cross-boundary channel identifiers, time-slice identifiers, congestion level distributions, and trend direction summaries, reducing irrelevant original information entering the federated update package. For regions lacking continuous time-slice boundary representations, alternative representations are generated using candidate prototypes with consistent directions in adjacent regions, adjacent time slices, and limited consistency deviations, and are incorporated into uncertainty weights, ensuring continuous updates to the cross-boundary situation prediction model even when boundary data is temporarily missing. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the cross-border road network congestion early warning and dynamic diversion guidance method of the present invention; Figure 2 This is a flowchart illustrating the construction of regional traffic map samples and the generation of dynamic influence matrices according to the present invention. Figure 3 This is a flowchart of the separate training and node-level desensitization compression process for the regional private encoder and the boundary shared encoder of the present invention. Figure 4 This is a flowchart illustrating the boundary consistency constraints, federated aggregation, and alternative representation compensation of the present invention. Detailed Implementation

[0020] 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, not all, of the embodiments of the present invention. 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.

[0021] Please refer to Figure 1This embodiment provides a method for cross-border road network congestion early warning and dynamic diversion guidance, including: constructing regional traffic map samples from checkpoint traffic flow, floating car speed sequences, road network topology connections, and signal phase summaries in different road management areas; generating internal node representations and boundary node representations locally in each area using a regional private encoder and a boundary shared encoder; uploading boundary node representation prototypes, model update amounts, and uncertainty weights under a federated learning architecture; constructing boundary consistency constraints based on the congestion level, speed decline trend, and transmission delay of boundary nodes in adjacent areas; and aggregating the cross-border situation prediction model according to the boundary consistency constraints to output the cross-border road network congestion risk level and boundary transmission direction. Specifically, different road management areas maintain their own local storage status for their original traffic data. Within each area, the platform first maps the collectable checkpoint traffic flow, floating car speed sequences, road network topology connections, and signal phase summaries to a unified time slice, and then sorts them by road segment nodes and boundary connection nodes. The system generates regional traffic map samples by using the travel edges between nodes. The regional private encoder is used to learn the normal traffic relationships and local congestion changes between road segments within the region. The boundary shared encoder is used to learn the comparable congestion levels, speed decline trends, and transmission delays of boundary connection nodes between adjacent regions. After local training, each region only outputs the prototype representation of boundary nodes, model update amount, and uncertainty weights. The federated aggregation end does not receive vehicle identifiers, original trajectory points, or original checkpoint records. The aggregation end places the boundary node representations pointing to the same cross-border channel in adjacent regions into a unified level space and a unified direction space. It adjusts the participation weight of the model update amount according to the boundary consistency constraint, and enables the aggregated cross-border situation prediction model to simultaneously provide the congestion risk level, congestion transmission direction, and situation source corresponding to the boundary channel during inference. This embodiment maintains the continuity of the cross-border congestion transmission chain under data isolation conditions at the model level through continuous processing of local modeling, boundary shared representation, federated aggregation, and consistency calibration.

[0022] In this embodiment, the node features in the regional traffic map sample can be constructed using the following formula: ; Where r represents the road management area number, i represents the road segment node or boundary connection node, and t represents the unified time slice. Let represent the node feature vector of node i within region r at time slice t. This represents the node flow value calculated based on checkpoint flow. This represents the node velocity value obtained by matching the floating car trajectory. This indicates the time slice attribute corresponding to the signal phase summary. Represents node topology attributes and boundary markers. , , , These represent transformation functions that perform same-scale mapping on different fields. For example, when the flow mapping value of a certain boundary node is 0.8, the speed mapping value is 0.3, the release direction summary mapping value is 1, and the boundary marker mapping value is 1, the node feature vector in the model is represented as a combination of high flow, low speed, current release, and boundary-related inputs, enabling the boundary shared encoder to include it in the candidate state of congestion spillover or recharge for subsequent processing.

[0023] In a preferred embodiment, reference Figure 2 The construction of the regional traffic map sample includes: determining road segment nodes, boundary connection nodes, and passable edges between nodes according to the road management area; writing checkpoint traffic flow into the node traffic sequence according to the road segment node affiliation, writing floating car speed sequence into the node speed sequence according to the trajectory matching result, and writing signal phase summary into the time slice attribute corresponding to the boundary connection node; generating a static topology matrix based on the node connection relationship, and generating a dynamic influence matrix based on the speed change direction and traffic migration direction within adjacent time slices; and encapsulating the static topology matrix, the dynamic influence matrix, and the node time sequence together into a local regional traffic map sample. In specific implementation, the road management area can be divided according to the platform data affiliation boundary, road segment nodes correspond to road sections, intersection entrance lanes, or road segment segments that can be identified by the topology model and can carry traffic state sequences, and boundary connection nodes correspond to cross-regional... Nodes on both sides of the domain channel have continuous vehicle traffic relationships with adjacent areas. The passable edges between nodes are jointly determined by road connections, turning permissions, and cross-border channel directions. Checkpoint traffic flow is assigned to the node traffic flow sequence according to the mapping relationship between checkpoint location and road segment nodes. Floating car speed sequences are aggregated in time slices according to the road segment where the trajectory point is located and the node affiliation relationship. The signal phase summary does not record the complete control scheme, but writes the release direction, the time slice to which the phase belongs, and the correspondence with the boundary channel. The static topology matrix represents the physical connection relationship of the road, and the dynamic influence matrix represents the influence intensity of the traffic state on adjacent nodes in the current time window. The node time sequence carries traffic flow, speed, phase summary, and boundary markers. This embodiment writes the static road network structure and dynamic traffic changes into the same map sample, so that the internal state of the region and the boundary transmission state have the same model input basis.

[0024] The sample fields and processing methods for the regional traffic map are shown in Table 1: Table 1. Configuration Instructions for Basic Road Network Fields in the Traffic Congestion Transmission Model ; Furthermore, the generation of the dynamic influence matrix includes: within each regional traffic map sample, using the speed difference, flow difference, and node adjacency direction of consecutive time slices as input, calculating the local traffic influence value between node pairs; adding cross-boundary direction markers to node pairs containing boundary connection nodes, and matching the cross-boundary direction markers with the release direction in the signal phase summary; when the speed difference direction of a node pair is inconsistent with the flow difference direction, retaining the influence edge in the reachable direction according to the topological reachability relationship; writing the influence values ​​of internal node pairs and boundary node pairs into the same dynamic influence matrix. Specifically, the speed difference of consecutive time slices is used to identify the decrease or recovery of node capacity, and the flow difference is used to identify the vehicle migration direction or flow aggregation state. The adjacency direction of a point is used to limit the propagation of traffic states to the permissible direction. Node pairs containing boundary connection nodes are also marked with cross-boundary direction tags, enabling the model to identify whether vehicles are flowing from the local area to the adjacent area or flowing back from the adjacent area to the local area. The release direction in the signal phase summary is used to eliminate instantaneous impact edges that do not match the current release direction. When the speed decreases but the flow does not increase synchronously, the system determines whether the state belongs to upstream queue overflow through topological reachability. When the flow increases but the speed change is small, the system treats it as a candidate pressure input instead of directly determining it as congestion. This embodiment incorporates time changes, spatial reachability, and boundary release directions into the same adjacency expression through a dynamic influence matrix, reducing the problem that static topology cannot characterize the cross-boundary traffic influence direction.

[0025] In this embodiment, the dynamic influence value between node pairs can be calculated using the following formula: ; in, This represents the dynamic impact value of node i to node j within region r during time slice t. This represents the velocity difference between node i and node j calculated along the propagation direction. This represents the flow difference between node i and node j calculated along the propagation direction. This indicates the direction encoding for whether node i is topologically reachable from node j. The release code indicates whether the signal phase summary matches the node pair's direction of travel. to This represents the field weights learned during training. This represents a function that maps the input to the interval between 0 and 1, for example, when 0.7 0.6 For 1, When the value is 1, the node pair is written with a higher dynamic influence value, and the subsequent graph neural network aggregates more boundary propagation information along this direction; when... 0 or When the value is 0, even if the speed difference is large, the node pair will not be treated as the main propagation edge.

[0026] In one embodiment, reference Figure 3 The regional private encoder and the boundary shared encoder are trained using a separate parameter update method. The regional private encoder receives internal road segment nodes, internal topological edges, and local historical time series to generate a regional traffic state representation. The boundary shared encoder receives boundary connection nodes, cross-border adjacency edge identifiers, exchangeable statistical summaries on both sides of the boundary, and boundary time slice attributes to generate a boundary node representation. Internal state reconstruction, boundary state prediction, and private representation isolation terms are set in the local loss, allowing the internal node representation and the boundary node representation to be updated separately in the same training round. In specific implementation, the regional private encoder can use a graph sequence network that spatiotemporally aggregates the adjacency relationships of internal nodes. The input content is limited to the internal node features, internal topological matrix, and historical time series of the region. The sequence outputs a representation of the traffic state within the region. The boundary-shared encoder can employ a boundary graph coding network with shared parameters. Inputs include boundary connection node features, cross-boundary adjacent edge identifiers, congestion level distribution summaries exchangeable between adjacent regions, speed trend summaries, and boundary time slice attributes. Outputs a representation of the boundary nodes. When the separation parameters are updated, the internal state reconstruction term constrains the region's private encoder to restore the local historical state, the boundary state prediction term constrains the boundary-shared encoder to predict the boundary state of the next time slice, and the private representation isolation term constrains the region's internal traffic state representation and boundary node representation to maintain a distinguishable feature space, avoiding the direct writing of special traffic patterns within the region into the boundary-shared part. This embodiment uses separate coding and simultaneous training to ensure that internal regional differences and cross-boundary sharing patterns are carried separately in the model structure.

[0027] In this embodiment, the local training loss can be constructed using the following formula: ; in, This represents the total local training loss for region r; Represents the set of nodes within a region. Represents the set of boundary connection nodes. Represents the correlation quantity of node features; This represents the node features reconstructed by the region-private encoder. This represents the next time-slice boundary state predicted by the boundary-shared encoder. This indicates the boundary state observed in the next time slice. Represents the internal node representation. Representing the boundary node characterization, , , Indicates the weight of the loss term. Describing the L2 norm, for example, when When a higher value is taken, the cosine similarity between the internal node representation and the boundary node representation is suppressed, and the boundary shared encoder tends to preserve cross-boundary channel exchangeable trend information rather than local internal noise; when When a higher value is taken, the prediction error of the next time slice state of the boundary node accounts for a larger proportion of the constraints during training, making the boundary node representation more suitable for federated consistency calibration.

[0028] In a preferred embodiment, the separation parameter update method includes: within each federated training round, first updating the parameters of the regional private encoder using the internal traffic state representation, then freezing the parameters related to internal nodes in the regional private encoder; subsequently updating the parameters of the boundary shared encoder using the boundary node representation, boundary time slice attributes, and boundary node representation prototypes returned by adjacent regions; before uploading the parameters, performing node-level desensitization and compression on the gradient of the boundary shared encoder, and writing the compressed gradient and boundary node representation prototypes together into the federated update package. Specifically, each road management region receives the shared parameters obtained from the previous round of aggregation and the exchangeable boundary node representation prototypes from adjacent regions at the start of the federated training round on the local platform. First, the time series of nodes within the region are read and the region's private encoder is updated to ensure that the representation of internal nodes matches the real-time traffic status of the region. Then, the parameters related to the internal nodes are frozen to retain their ability to represent the internal state. Next, the current time slice attributes of the boundary connection nodes, the cross-border adjacent edge identifiers, and the boundary node representation prototypes returned by the adjacent regions are read to update the boundary shared encoder. After the boundary shared encoder is updated, it outputs the representation prototype, model update amount, and uncertainty weights corresponding to each boundary connection node. The federated update package only contains gradient summaries and boundary prototype indexes after node-level desensitization and compression. This embodiment reduces the interference of differences in the distribution of samples within the region on the boundary shared parameters by adopting the training order of internal nodes first and then boundary nodes and the freezing mechanism.

[0029] Further, the node-level desensitization and compression includes: grouping the gradients of the boundary shared encoder according to the cross-border channel to which the boundary connection node belongs; extracting the directional component, level component, and temporal component for each group of gradients, and deleting fields corresponding to vehicle identifiers, original trajectory points, and checkpoint original records; merging the gradient components after deleting fields to form a compressed gradient containing only cross-border channel identifiers, time slice identifiers, congestion level distributions, and trend direction summaries; and uploading the compressed gradients after establishing a round index with the corresponding boundary node representation prototypes. Specifically, the gradients obtained by backpropagation of the boundary shared encoder are grouped according to the cross-border channel identifier, and upstream boundary nodes, downstream boundary nodes, and co-directional auxiliary nodes in the same cross-border channel are grouped into... Within the same gradient group, the directional component is extracted from the vector dimension related to the trend direction summary in the gradient, the level component is extracted from the output dimension related to the congestion level distribution, and the time series component is extracted from the dimension related to the time slice encoding and the candidate value of the transmission delay. The field deletion process is completed before generating the federated update package. The fields of vehicle identification, original trajectory points, original checkpoint records, and fields that can be used to infer the single vehicle passage path do not enter the compressed gradient. Prototype merging can use the mean within the group, the median vector, or a weighted vector with uncertainty weights. The round index is used to mark the training round, cross-boundary channel, and time slice to which the compressed gradient belongs. In this embodiment, node-level desensitization compression maintains the aggregable form of boundary transmission training information and avoids the original traffic details from participating in cross-regional transmission.

[0030] In one embodiment, reference Figure 4The construction of the boundary consistency constraints includes: for boundary connection nodes pointing to the same cross-border passage in adjacent road management areas, extracting the congestion level distribution, speed decrease trend vector, queue growth trend vector, and transmission delay candidate values ​​obtained locally in each area; mapping the congestion level distribution to a unified level space, and mapping the speed decrease trend vector and the queue growth trend vector to a unified direction space; forming paired boundary samples according to the upstream and downstream directions of the cross-border passage, and applying level consistency constraints, trend direction consistency constraints, and time delay reachability constraints to the paired boundary samples. Specifically, adjacent areas save the local state of their own boundary connection nodes for the same cross-border passage. The federated aggregation end or entrusted aggregation node does not read the original checkpoint data, but reads the boundary node representation prototype and statistical summary uploaded by each area. The congestion level distribution can be divided into low risk, medium risk, and high risk. The probability vector representation of discrete states such as high risk is used. The velocity decline trend vector can be composed of velocity difference and direction encoding between adjacent time slices. The queue growth trend vector can be obtained by combining the flow aggregation and velocity decline of upstream candidate nodes of the boundary node. The transmission time delay candidate value is calculated by the time offset of the state change on both sides of the cross-border channel. The unified level space is used to eliminate the difference in the scale of congestion level classification between regions. The unified direction space is used to eliminate the difference in naming or coordinate direction between upstream and downstream. The level consistency constraint requires that the congestion level distribution on both sides of the same cross-border channel be close within the interpretable range. The trend direction consistency constraint requires that the velocity decline trend and the queue growth trend match along the passable direction. The time delay reachability constraint requires that the transmission time offset be consistent with the cross-border channel passage logic. This embodiment enables adjacent regions to form comparable boundary situation representations without sharing the original data through boundary consistency constraints.

[0031] In this embodiment, the consistency constraint of paired boundary samples can be calculated using the following formula: ; in, This represents the consistency deviation between boundary node a and boundary node b on both sides of the cross-boundary channel. This represents the divergence operation between two congestion level distributions. and This represents the time-aligned distribution of congestion levels. and This represents the trend direction vector after time alignment. This represents the candidate value of the conduction delay calculated from the current observations. This indicates the reference time lag corresponding to the historical passage time distribution of cross-border channels. and This represents the constraint weights, for example, when the congestion level distributions of two boundary nodes are similar, the angle between the trend direction vectors is small, and... near hour, Lower values ​​correspond to higher participation weights in federated aggregation for model updates; when the level distributions differ significantly or the trends are opposite... As the value increases, the corresponding update amount is given a lower weight in the aggregation.

[0032] Further, the determination of the candidate values ​​for conduction delay includes: performing sliding matching on the velocity decrease trend vectors of the boundary connection nodes on both sides of the same cross-border channel to obtain multiple candidate time offsets; performing directional consistency screening on the queue growth trend vector under each candidate time offset; retaining the candidate time offsets that satisfy the topological reachability relationship of the cross-border channel as candidate values ​​for conduction delay; in the time delay reachability constraint, jointly encoding the candidate values ​​for conduction delay with the historical passage time distribution of the cross-border channel and the boundary signal release window. In specific implementation, the system uses the velocity decrease trend vectors on both sides of the boundary connection node as the matching object, moves the trend sequence on one side within the allowed time slice range, and calculates the vector similarity under each time offset to form a set of candidate time offsets, and then reads the phase... If the queue growth trend vector under the same time offset is consistent with the vehicle flow direction of the cross-border channel, the candidate time offset is retained. If the queue growth direction is opposite to the vehicle flow direction and the topology is unreachable, the candidate time offset is eliminated. The retained candidate time offset also needs to be compared with the historical passage time distribution of the cross-border channel and aligned with the boundary signal release window. If the candidate time offset falls into the actual signal release time period or its adjacent propagation window, a time delay reachability constraint is written. If the candidate time offset is different from the historical passage time distribution and there is no release window in the corresponding time period, it is treated as an abnormal candidate value. This embodiment uses the joint screening of speed trend, queue trend, topology reachability and signal release window to make the propagation time delay reflect the real cross-border channel propagation conditions.

[0033] In one embodiment, the aggregation under the federated learning architecture includes: receiving model updates, boundary node representation prototypes, and uncertainty weights uploaded by each road management region; calculating boundary deviations based on the level consistency constraint, the trend direction consistency constraint, and the time-delay reachability constraint; adjusting the participation weight of each region's model updates in the aggregation based on the boundary deviations; when a region lacks boundary node representation prototypes for consecutive time slices, generating alternative representations using boundary node representation prototypes with consistent cross-boundary directions from adjacent regions, and combining the alternative representations with the current round's model updates to participate in the cross-boundary situation prediction model aggregation. Specifically, the aggregation end establishes a federated update record for each road management region, the record including model update amounts, boundary node representation prototypes, compression gradients, uncertainty weights, cross-boundary channel identifiers, and training round indexes. The aggregation end pairs the boundary node representation prototypes of adjacent regions according to the cross-boundary channel identifier, and calculates the boundary deviation using level consistency constraints, trend direction consistency constraints, and time lag reachability constraints. The update volume of regions with lower boundary deviation and higher uncertainty weight occupies a higher participation proportion in the aggregation, while the update volume of regions with higher boundary deviation or lower uncertainty weight is restricted from participating. When a region lacks a boundary node representation prototype in a continuous time slice, the aggregation end selects a boundary node representation prototype with consistent cross-boundary direction from adjacent regions to generate a substitute representation. The substitute representation and the model update volume of the missing region that can still be uploaded are included in the aggregation. The aggregated model parameters are distributed to each region for the next round of local training and cross-boundary situation prediction. This embodiment reduces the conflict between update volumes of non-uniformly distributed regions and the global model by driving the aggregation weight with boundary consistency.

[0034] In this embodiment, the federated aggregation and alternative characterization can be calculated using the following formula:

[0035]

[0036] ; in, R represents the parameters of the aggregated cross-border situation prediction model, and R represents the set of road management areas participating in the federated aggregation. This indicates the parameter corresponding to the amount of model updates uploaded to region r. This indicates the participation weight of region r in the aggregation. The uncertainty weight of the upload from region r is represented in the summation term of the denominator of the formula. This represents the uncertainty weight uploaded by any region k in the set of road management regions participating in the federal aggregation, and its meaning is similar to... Consistent, used only as a summation subscript in the normalization term of the denominator; This represents the mean deviation of consistency among the boundary samples of region r. This represents the alternative representation of the missing region boundary node a in time slice t. This represents the boundary node characterization of adjacent region boundary node b after alignment according to the propagation time delay. This represents the boundary node representation of the missing region in the previous training round. This represents the fusion coefficient, for example, when region r... lower and At higher levels, As the size increases, the model update amount of the region retains more boundary propagation information in the aggregation; when the previous representation of a missing region is relatively new and the temporally aligned representations of adjacent regions are stable, An intermediate value can be chosen so that the alternative representation can simultaneously preserve the transmission state of adjacent regions and the semantics of the historical boundaries of missing regions.

[0037] In a preferred embodiment, the generation of the alternative representation includes: selecting candidate prototypes from the boundary node representation prototypes uploaded from adjacent regions that are consistent with the cross-boundary channel direction of the missing region, have adjacent time slices, and have consistent level constraints with a deviation amount less than a preset consistency range; performing time alignment on the candidate prototypes based on the transmission delay candidate value; fusing the time-aligned candidate prototypes with the boundary node representation prototypes of the missing region from the previous training round; writing uncertainty weights into the fused representations, and correcting the uncertainty weights in the next federated training round based on the actual re-uploaded boundary node representation prototypes. Specifically, when a road management area cannot continuously upload boundary node representation prototypes due to communication delays, time slice gaps, or temporary lack of local data, the aggregation end does not directly discard the relevant cross-boundary channels of that area, but instead uses the cross-boundary channel identifier and... The directional marker filters candidate prototypes from the boundary node representation prototypes uploaded from adjacent regions. The candidate prototypes need to meet the following requirements: consistent vehicle flow direction, adjacent time slices with missing time slices, and consistent level constraint deviation within a preset consistency range. The filtered candidate prototypes are time-aligned according to the candidate values ​​of the transmission delay to correspond to the time position when the missing region may receive congestion transmission. The time-aligned candidate prototypes are fused with the boundary node representation prototypes of the missing region in the previous training round. The fused representation is written with an uncertainty weight lower than the measured representation and is compared with the actual retransmitted boundary node representation prototype in the next federated training round. The uncertainty weight is corrected according to the difference in level distribution, trend direction, and time delay between the two. This embodiment maintains the continuity of federated training when the boundary representation is missing for a short time through substitute representation and subsequent correction.

[0038] In this embodiment, the inference process of the cross-boundary situation prediction model is executed using the aggregated model parameters and the latest local regional traffic map samples of each region. The regional platform writes the checkpoint traffic, floating car speed sequence, road network topology connection relationship and signal phase summary in the current time slice into the local regional traffic map sample. The regional private encoder reads the internal node features and outputs the internal traffic state representation of the region. The boundary shared encoder reads the boundary connection nodes, cross-boundary adjacent edge identifiers, boundary time slice attributes and the boundary shared parameters issued in the previous round of federation and outputs the boundary node representation. The cross-boundary situation prediction model performs graph time-series propagation on the internal node representation and the boundary node representation locally. The boundary consistency constraint is transformed in the inference stage. To verify the boundary characterization, the exchangeable grade distribution, trend direction, and time delay information on both sides of the same cross-border channel are compared. If the comparison results are within the same range, the congestion risk level and boundary transmission direction of the corresponding cross-border channel are output. If the comparison results are outside the same range, the uncertainty weight is written into the output result and the corresponding update quantity is reduced in the next federated training round. The output cross-border road network congestion risk level can be used as a risk input for dynamic diversion guidance, and the boundary transmission direction can be used as a model basis for determining the diversion trigger area and the transmission source area. This embodiment extends the consistency calibration of the training stage to the inference stage, so that the prediction output and the boundary transmission characterization maintain the same technical standard.

[0039] In a preferred embodiment, the output of the congestion risk level of the cross-boundary road network can be generated in parallel using node-level and channel-level results. The node-level results are directly obtained from the congestion level distribution of each boundary connection node, while the channel-level results are obtained by fusing the level distribution, trend direction vector, and propagation delay candidate values ​​of the boundary connection nodes on both sides of the same cross-boundary channel. Specifically, the model reads the level component in the boundary node representation and converts it into a risk level distribution; reads the trend direction component and determines whether congestion propagates from within the region to outside the region or from adjacent regions to the current region; and reads the time-series component and determines the time slice location where the risk occurs. If the node-level results on both sides of the channel point to the same propagation direction, the channel-level results are generated according to the boundary node representation prototype with higher confidence. If the node-level results on both sides of the channel are in opposite directions, the representation to be fused is selected based on the boundary consistency deviation and uncertainty weight. The output results retain the cross-border channel identifier, time slice identifier, congestion risk level, boundary propagation direction and uncertainty weight, but do not output vehicle identifiers, original trajectory points or checkpoint original records. This embodiment outputs node-level and channel-level results in parallel, so that the cross-border situation prediction results can maintain a correspondence with the boundary channel state required for subsequent dynamic diversion guidance.

[0040] In this embodiment, a state recording mechanism is used to maintain a closed loop between federated training rounds and inference rounds. The state recording content includes the boundary node representation prototype of each cross-boundary channel, the deviation of the level consistency constraint, the deviation of the trend direction consistency constraint, the candidate value of the propagation delay, the uncertainty weight, the alternative representation marker, and the model version identifier. After the regional platform completes training locally, it uploads the federated update package. After the aggregation terminal completes cross-regional aggregation, it sends out the cross-boundary situation prediction model parameters and boundary prototype calibration information. The regional platform then executes the next round of local training and local inference based on this information. If a cross-boundary channel shows a decreasing velocity trend within a continuous time slice, and The queuing trend is increasing, but the prototype of the boundary node representation returned by the adjacent region is inconsistent with the representation of this region. The system writes the cross-boundary channel into the boundary deviation record and adjusts the participation weight in the subsequent federated aggregation. If the actual supplementary data proves that the previous round of replacement representation is close to the real boundary node representation, the uncertainty weight of the corresponding replacement representation source region is increased. If the actual supplementary data is significantly different from the replacement representation, the weight of the corresponding source region on the same channel is reduced. This embodiment makes model update, boundary calibration and short-term missing compensation within the same data chain through the continuous association of training records, inference records and supplementary correction records.

[0041] In a preferred embodiment, the regional private encoder and the boundary shared encoder can share the input time slice division rule but not all parameters. The input window of the regional private encoder covers the historical time series within the region, and the input window of the boundary shared encoder covers the short-term change series of boundary connection nodes and their adjacent nodes. After the outputs of the two are combined in the cross-border situation prediction model through a gated fusion vector, the gated fusion vector is determined by whether the node belongs to the boundary connection node, the influence value of the boundary node in the dynamic influence matrix, and the uncertainty weight. For internal nodes, the gated fusion vector assigns a higher input weight to the internal traffic state representation of the region. For boundary connection nodes, the gated fusion vector assigns a higher input weight to the boundary node representation. For internal nodes adjacent to the boundary connection nodes, the gated fusion vector adjusts the fusion ratio of the two representations according to the influence value of the dynamic influence matrix. During training, the gated fusion vector is updated together with the local loss and boundary consistency constraints. During inference, the gated fusion vector is calculated with the latest regional traffic map samples. This embodiment achieves clear input boundaries for internal road segment state changes and cross-border transmission state changes in the same prediction model by adaptively fusing the internal and boundary representations by node type.

[0042] In a preferred embodiment, the uncertainty weight can be jointly determined by data integrity, boundary consistency deviation, alternative representation marker, and time slice freshness. Data integrity is calculated based on whether the corresponding boundary connection node has continuous checkpoint flow, floating car speed sequence, and signal phase summary in the current training round. Boundary consistency deviation is calculated based on level consistency constraint, trend direction consistency constraint, and time delay reachability constraint. The alternative representation marker is used to distinguish between the measured boundary node representation prototype and the alternative representation generated by adjacent regions. Time slice freshness is used to distinguish between the current time slice representation and the representation left over from the previous training round. The measured and complete boundary node representation prototype has a higher weight, while the alternative representation or time slice missing representation has a lower weight. If the supplementary data in the next training round is close to the alternative representation, the uncertainty weight of the corresponding source region is restored. If the supplementary data is far from the alternative representation, the weight of the corresponding source region on the same cross-boundary channel is reduced. This embodiment uses uncertainty weight to jointly characterize data integrity and boundary consistency, so that federated aggregation is not dominated by short-term abnormal representation or missing compensation representation.

[0043] In this embodiment, the content read by the cross-boundary dynamic diversion guidance side is limited to the congestion risk level of the cross-boundary road network and the boundary transmission direction. The diversion guidance system can determine whether to trigger boundary channel diversion based on the risk level, determine the risk source area and affected area based on the boundary transmission direction, and determine whether to wait for the next time slice verification based on the uncertainty weight. The method itself does not require cross-regional sharing of original vehicle trajectories, checkpoint details, or complete signal control schemes. The dynamic diversion guidance side receives the channel-level risk status output by the cross-boundary situation prediction model. If the risk level of a cross-boundary channel increases and the transmission direction points from the adjacent area to this area, the diversion guidance system can write the channel identifier into the candidate diversion constraint. If the risk level increases but the uncertainty weight is low, the diversion guidance system can write the channel identifier into the pending verification state and wait for the next time slice model output. If the risk level is low and the transmission direction is stable, the diversion guidance system can maintain the current channel state. This embodiment, by limiting the output fields and the transmission boundary direction, enables the cross-boundary situation prediction results to enter the subsequent guidance and control chain without changing the technical boundary of the original data not leaving the domain.

[0044] In this embodiment, the complete processing chain can consist of local sample construction, separation encoding, local training, desensitization and compression, federated aggregation, boundary consistency calibration, alternative representation compensation, and local inference. After the regional platform completes the construction of regional traffic map samples locally, it uses a regional private encoder to generate internal node representations and a boundary shared encoder to generate boundary node representations. It updates the local model using internal state reconstruction terms, boundary state prediction terms, and private representation isolation terms. Subsequently, it performs node-level desensitization and compression on the gradient of the boundary shared encoder, uploads the boundary node representation prototype, model update amount, and uncertainty weights, and the federated aggregation end calculates the boundary deviation amount based on the boundary consistency constraints and adjusts the model update amount to participate in the weights. It generates alternative representations for boundary node representation prototypes with missing continuous time slices. After aggregating to obtain the cross-border situation prediction model, it distributes it to each region. The regional platform uses the latest local regional traffic map samples to complete the output of the cross-border road network congestion risk level and boundary transmission direction. This embodiment forms a reproducible technical closed loop between data fusion, situation prediction, and dynamic diversion guidance input by constraining each processing link within the same cross-border boundary representation logic.

Claims

1. A method for congestion early warning and dynamic diversion guidance in cross-border road networks, characterized in that, include: Traffic flow at checkpoints, floating car speed sequences, road network topology connections, and signal phase summaries within different road management areas are used to construct regional traffic map samples. In each region, a local private encoder and a boundary shared encoder are used to generate internal node representations and boundary node representations. Upload the boundary node representation prototype, model update amount, and uncertainty weights under the federated learning architecture; Boundary consistency constraints are constructed based on the congestion level, speed decline trend, and transmission delay of adjacent regional boundary nodes; Based on the boundary consistency constraints, a cross-boundary situation prediction model is obtained, which outputs the congestion risk level of the cross-boundary road network and the direction of boundary transmission. The region-private encoder and the boundary-shared encoder are trained using a separate parameter update method; The regional private encoder receives internal road segment nodes, internal topological edges, and local historical time series to generate a regional traffic state representation; the boundary shared encoder receives boundary connection nodes, cross-boundary adjacent edge identifiers, exchangeable statistical summaries on both sides of the boundary, and boundary time slice attributes to generate a boundary node representation; internal state reconstruction term, boundary state prediction term, and private representation isolation term are set in the local loss respectively, so that the internal node representation and the boundary node representation are updated separately in the same training round; The construction of the boundary consistency constraint includes: for the boundary connection nodes in adjacent road management areas that point to the same cross-border passage, extracting the local congestion level distribution, speed decrease trend vector, queue growth trend vector and transmission delay candidate value of each area; The congestion level distribution is mapped to a unified level space, and the speed decrease trend vector and the queue increase trend vector are mapped to a unified direction space. Paired boundary samples are formed according to the upstream and downstream directions of the cross-border channel, and level consistency constraints, trend direction consistency constraints and time lag reachability constraints are applied to the paired boundary samples.

2. The method for cross-border road network congestion early warning and dynamic diversion guidance according to claim 1, characterized in that, The construction of the regional traffic map sample includes: determining road segment nodes, boundary connection nodes, and traversable edges between nodes according to the road management area; Write the traffic flow at the checkpoint into the node traffic sequence according to the road segment node affiliation, write the floating car speed sequence into the node speed sequence according to the trajectory matching result, and write the signal phase summary into the time slice attribute corresponding to the boundary connection node. A static topology matrix is ​​generated based on the node connection relationship, and a dynamic influence matrix is ​​generated based on the velocity change direction and flow migration direction within adjacent time slices; The static topology matrix, the dynamic influence matrix, and the node time series are collectively encapsulated into a local regional traffic map sample.

3. The method for cross-border road network congestion early warning and dynamic diversion guidance according to claim 2, characterized in that, The generation of the dynamic impact matrix includes: within each regional traffic map sample, using the speed difference, flow difference, and node adjacency direction of consecutive time slices as input, calculating the local traffic impact value between node pairs; For node pairs containing boundary connection nodes, add a cross-boundary direction marker and match the cross-boundary direction marker with the release direction in the signal phase summary; When the velocity difference direction and the flow difference direction of a node pair are inconsistent, the influential edge in the reachable direction is retained according to the topological reachability relationship; Write the influence values ​​of internal node pairs and boundary node pairs into the same dynamic influence matrix.

4. The method for cross-border road network congestion early warning and dynamic diversion guidance according to claim 3, characterized in that, The separation parameter update method includes: in each federated training round, first update the parameters of the regional private encoder with the internal traffic status representation of the region, and then freeze the parameters of the regional private encoder related to the internal nodes. The parameters of the boundary shared encoder are then updated using the boundary node representation, boundary time slice attributes, and boundary node representation prototypes returned by adjacent regions. Before uploading the parameters, the gradient of the boundary shared encoder is desensitized and compressed at the node level, and the compressed gradient and the boundary node representation prototype are written into the federated update package.

5. The method for cross-border road network congestion early warning and dynamic diversion guidance according to claim 1, characterized in that, The determination of the candidate values ​​of the conduction time delay includes: performing sliding matching on the velocity decrease trend vector of the boundary connection nodes on both sides of the same cross-border channel to obtain multiple candidate time offsets; Perform directional consistency screening on the queue growth trend vector under each candidate time offset; Candidate time offsets that satisfy the topological reachability of cross-boundary channels are retained as candidate values ​​for conduction delays; In the time-delay reachability constraint, the candidate values ​​of the conduction time delay are jointly encoded with the historical passage time distribution of the cross-border channel and the boundary signal release window.

6. The method for cross-border road network congestion early warning and dynamic diversion guidance according to claim 1, characterized in that, The aggregation under the federated learning architecture includes: receiving model update data, boundary node representation prototypes, and uncertainty weights uploaded by each road management area; The boundary deviation is calculated based on the level consistency constraint, the trend direction consistency constraint, and the time delay reachability constraint. The participation weight of the model update amount in the aggregation is adjusted according to the boundary deviation amount; When a region lacks a boundary node representation prototype for a continuous time slice, an alternative representation is generated using the boundary node representation prototype with the same cross-boundary direction in adjacent regions. The alternative representation and the current model update are then used together in the cross-boundary situation prediction model aggregation.

7. The method for cross-border road network congestion early warning and dynamic diversion guidance according to claim 4, characterized in that, The node-level desensitization compression includes: grouping the gradient of the boundary shared encoder according to the cross-boundary channel to which the boundary connection node belongs; For each set of gradients, extract the directional component, level component, and temporal component, and delete the fields corresponding to vehicle identification, original trajectory points, and checkpoint original records; The gradient components after deleting fields are prototyped and merged to form a compressed gradient that only contains cross-boundary channel identifiers, time slice identifiers, congestion level distributions, and trend direction summaries. After establishing a round index between the compression gradient and the corresponding boundary node representation prototype, the data is uploaded.

8. The method for cross-border road network congestion early warning and dynamic diversion guidance according to claim 6, characterized in that, The generation of the alternative representation includes: selecting candidate prototypes from the boundary node representation prototypes uploaded from adjacent regions that are consistent with the cross-boundary channel direction of the missing region, adjacent in time slice, and have the same level of constraint deviation with a preset consistency range. The candidate prototypes are time-aligned based on the proposed conduction delay values; The time-aligned candidate prototypes are fused with the boundary node representation prototypes of the missing regions from the previous training round. The fused representation is written with uncertainty weights, and the uncertainty weights are corrected in the next federated training round based on the actual retransmitted boundary node representation prototype.

Images