A method for planning and designing a highway network across provincial administrative regions

Abstract

The application discloses a kind of interprovincial administrative region highway network planning design method, it is related to traffic highway network planning design field.By regarding interprovincial administrative boundary as virtual fracture surface, the traffic flow characteristic vector data of section is collected in real time at each interprovincial passage;The characteristics of all sections in the same time window are organized as an unordered set, and they are directly mapped into interprovincial OD matrix through reverse mapping network model;In training, introduce flow conservation, sparsity and time sequence smoothing physical constraints;In short-term prediction, the future section characteristics of each section are independently predicted using time sequence convolution network, and then the future OD matrix is obtained through reverse mapping.The application breaks through the dependence of traditional method on interprovincial OD sample, has low data barrier, fast real-time response and light calculation, and can be widely used in interprovincial passage planning, dynamic guidance and emergency control scenes.

Description

Technical Field

[0001] This invention relates to the field of highway network planning and design, and in particular to a method for planning and designing highway networks that span provincial administrative regions. Background Technology

[0002] Traffic demand forecasting for inter-provincial highway networks is a core foundation for regional transportation planning, inter-provincial corridor construction, and inter-provincial traffic management. Accurate forecasting of inter-provincial traffic demand is of great value for optimizing inter-provincial highway layout, coordinating regional transportation policies, and alleviating congestion on inter-provincial corridors.

[0003] Currently, traffic demand forecasting for inter-provincial highway networks mainly employs a four-stage method: traffic generation, traffic distribution, mode classification, and traffic assignment. This method begins with the occurrence and attraction of trips in each traffic zone, obtains the origin-destination (OD) matrix through a gravity model or intervention opportunity model, and then allocates it to the road network. This method has significant drawbacks: First, it requires a large amount of resident travel survey data, resulting in high costs and long update cycles; second, obtaining inter-provincial data is difficult, as OD survey standards differ between provinces, making it difficult to piece together a complete inter-provincial travel profile; third, the chain-like structure of the four-stage method leads to amplified errors at each stage; and fourth, the gravity model is inherently static and cannot respond to sudden changes in inter-provincial policies (such as temporary traffic restrictions and environmental controls).

[0004] In recent years, methods have emerged that utilize Graph Neural Networks (GNNs) to jointly model road network topology and origin-destination (OD) distribution. These methods can capture spatial dependencies, but their training requires a large number of high-quality historical OD samples as labels, while cross-provincial OD samples are precisely the scarcest. Furthermore, the computational complexity of GNN models increases quadratically with the number of nodes, making it difficult to scale to large-scale road networks covering multiple provinces. In addition, dynamic methods based on traffic assignment allocate travel demand to the road network and iteratively update it using dynamic traffic assignment models. These methods are considered the most theoretically complete, but they require accurate road network supply models and traveler behavior models, parameter calibration is extremely complex, and they have high requirements for real-time data, resulting in a high barrier to engineering application.

[0005] In summary, there is an urgent need for a cross-provincial traffic demand forecasting method that has low data barriers, fast real-time response, can be trained with weak supervision, and is computationally lightweight. To this end, a cross-provincial highway network planning and design method is proposed. Summary of the Invention

[0006] The main objective of this invention is to provide a method for planning and designing highway networks across provincial administrative regions, which can effectively solve the problems in the background art.

[0007] To achieve the above objectives, this invention provides a method for planning and designing a highway network spanning provincial administrative regions, comprising the following steps: Step S1: Deploy virtual fault surfaces across provincial administrative boundaries Data collection sections are set up at each inter-provincial passage, such as provincial border toll stations on highways and provincial border checkpoints on national and provincial roads. Existing traffic flow detection equipment, such as ETC gantries, traffic control stations, loop detectors, and video detectors, is used to collect traffic flow feature vectors at each section in real time.

[0008] The cross-sectional feature vector includes at least the following six dimensions: average speed, speed standard deviation, average flow rate, proportion of large vehicles, coefficient of variation of headway, and event marker. The feature vector is represented as: .

[0009] The calculation methods for each component are as follows: average speed : Take the arithmetic mean of the speeds of all vehicles passing through the cross section within the time window; speed standard deviation This reflects the dispersion of traffic flow; the standard deviation decreases during congestion and increases during free flow. Average flow : The number of vehicles passing through within a unit of time (e.g., 5 minutes); Large vehicle proportion The proportion of large trucks and buses to the total number of vehicles passing through is used to distinguish between freight-dominated and passenger-dominated directions. Coefficient of variation of headway The ratio of the standard deviation of the headway to the mean. A high value indicates that the vehicles arrive randomly (typical long-distance travel), while a low value indicates that the vehicles are closely following each other (typical commuting or congestion). Event marker features The event marker feature is a binary variable used to indicate whether abnormal events such as traffic accidents, temporary traffic control, or severe weather have occurred in the section within the time window. The data source is the event release system of the road network center of each province.

[0010] Step S2: Construct the cross-sectional feature set Organize the feature vectors of all provincial boundary sections within the same time window t into an unordered set. , is represented as: , where M is the total number of provincial boundary sections within the region participating in the prediction. The main characteristic of this set is its disorder; there is no natural order relationship between the sections, therefore the subsequent model must be insensitive to the input order.

[0011] Step S3: Construct the inverse mapping network model The model adopts a set-to-vector mapping architecture, taking the cross-sectional feature set as input and outputting a vectorized representation of the cross-provincial OD matrix.

[0012] The network structure consists of three parts: The first part is the cross-sectional embedding layer: the feature vector of each cross section. Independent input first multilayer perceptron Each cross-sectional feature is mapped to an embedding vector. : ; The second part is the aggregation layer: a symmetric function is used to aggregate all cross-sectional embedding vectors, and the aggregated global features are output. Preferably, an arithmetic mean aggregator is used, expressed as: ; The third part is the decoding layer: it processes the aggregated global features. Input to the second multilayer perceptron Output the vectorized representation of the cross-provincial OD matrix: Where K represents the number of provincial-level administrative regions involved. For a K 2 A dimensional vector, where each element corresponds to the travel demand of a province pair (including the departure province and the arrival province).

[0013] vector The reshaped OD matrix is ​​represented as a K×K matrix: ; Force diagonal elements to be set to zero: To eliminate the need for travel within the province; Perform total normalization: ,in For assignment operation, The measured total flow at all cross-sections is used to ensure that the total output OD matrix is ​​consistent with the observed total inter-provincial travel volume.

[0014] Step S4: Apply physical constraints To ensure that the OD matrix output by the model is consistent with objective physical laws, a flow conservation constraint is introduced.

[0015] The projection matrix is ​​defined as follows: If section s connects a directed channel from province a to province b, then ,otherwise For a bidirectional cross section, it is split into two directed virtual cross sections.

[0016] Cross-sectional flow The OD matrix should satisfy a linear relationship: ; In theory, the projected flow rate should be equal to the measured cross-sectional flow rate. Since the projection error is consistent, it is treated as the physical constraint loss, i.e., the projection error term. , is represented as: ,in This is the s-th row of the projection matrix; Furthermore, the inter-provincial OD matrix is ​​inherently sparse, meaning that most provincial pairs do not have direct inter-provincial travel demand. Therefore, an L1 sparse prior is applied, with sparse prior terms... Defined as: ,in It is an L1 norm; To encourage temporal smoothness of the OD matrix, a temporal smoothing regularization term is applied: Temporal Smoothing Term Defined as: ; By analyzing the deviation between the actual OD samples and the predicted values, the deviation is minimized, and a supervision error term is defined. for: ,in It is the Frobenius norm. This represents the number of training samples.

[0017] The total loss function can then be expressed as: ,in , , , All of these are hyperparameters. The following condition must be met: when a real OD sample exists, When no real OD sample exists, .

[0018] Furthermore, when real OD samples When physical constraints and regularization terms are unavailable (cold start scenario), weakly supervised training can be performed using only physical constraints and regularization terms. In this case, the total loss function is: After accumulating a small number of real OD samples, we will switch to the fully supervised mode for fine-tuning.

[0019] Step S5: Conduct short-term traffic demand forecasting Unlike traditional methods that directly predict the future OD matrix, this invention first predicts future cross-sectional features and then obtains the future OD through reverse mapping. This is because cross-sectional features are low-dimensional time series, which are easier to predict than high-dimensional OD matrices. Specifically: For each cross section s, a lightweight temporal prediction model is trained independently. The temporal prediction model uses a temporal convolutional network, represented as follows: ,in For the length of the history window, It is the expansion factor. Learnable convolutional kernels; The cross-sectional feature matrix of the past L time windows As input, output the predicted feature sequence for the next H time windows. ; The predicted features of all cross sections are combined into a set of future cross section features, represented as:

[0020] Then, the cross-sectional feature set of each future time step is input into the pre-trained back-mapping network to obtain the cross-provincial OD matrix for the future time step: Predict the cross-sectional feature set for multiple future time windows, input it into the trained back-mapping network model, and output the cross-provincial OD matrix for future time moments.

[0021] Multi-time resolution prediction output: The multi-temporal resolution prediction output layer simultaneously generates cross-provincial OD prediction results with at least two of the following temporal granularities: Real-time prediction: Based on the current cross-sectional characteristics, output the cross-provincial OD matrix for the current time window with a delay of no more than 5 minutes. ; Short-term forecast: Outputs cross-provincial OD matrix sequences for the next 1 to 6 hours with a time resolution of 15 minutes. ; Event response prediction: When the event labeling features of any cross section When the value changes from 0 to 1, the OD (October Deficit) prediction for the next time window is automatically triggered, and the change in OD compared to the normal scenario is output. .

[0022] Beneficial effects Compared with the prior art, the present invention has the following beneficial effects: This solution only requires each province to share real-time traffic flow characteristics (average speed, flow rate, vehicle type ratio, etc.) at the provincial boundary sections, without having to hand over detailed internal OD data or mobile signaling data.

[0023] This solution directly utilizes real-time cross-sectional features for reverse mapping. When sudden policies (such as temporary traffic restrictions) are issued, the cross-sectional features will change within minutes. The model can immediately capture and update OD predictions, and the response delay can be controlled within minutes. This significantly shortens the time compared to traditional methods that require recalibrating parameters or waiting for new OD surveys, which can take several days to several weeks.

[0024] The reverse mapping network constructed in this scheme consists of only two multilayer perceptrons and one average pooling layer, with the number of parameters typically in the hundreds of thousands. The temporal prediction model for cross-sectional features is an independent TCN, and each cross-sectional model is lightweight and can be computed in parallel. The overall computational load can achieve a response time in seconds on a regular server, which is far lower than the computational complexity of graph neural networks or dynamic traffic assignment methods.

[0025] This solution can output multi-resolution OD predictions ranging from real-time (minutes) to short-term (hours), and has event-triggered prediction capabilities. It can be applied to various scenarios such as dynamic guidance of cross-provincial channels, inter-provincial collaborative management and control, and regional emergency evacuation. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a cross-provincial administrative region highway network planning and design method according to the present invention; Figure 2 This is a schematic diagram of the structure of the reverse mapping network model of the present invention; Figure 3 This is a schematic diagram of the structure of the cross-provincial administrative region highway network traffic demand forecasting system of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] Example 1: Traffic Demand Forecasting for Inter-Provincial Highway Networks This embodiment uses the inter-provincial highway network between Shanghai, Jiangsu, Zhejiang and Anhui (three provinces and one municipality) in the Yangtze River Delta region as an application scenario to explain the specific implementation process of the present invention in detail.

[0029] Step 1, Scene Parameter Settings Provincial-level administrative regions involved: K=4 (Shanghai, Jiangsu, Zhejiang, Anhui); Total number of data collection sections: M=36 (covering 6 channels between Shanghai and Jiangsu, 4 channels between Shanghai and Zhejiang, 8 channels between Jiangsu and Zhejiang, 10 channels between Jiangsu and Anhui, and 8 channels between Zhejiang and Anhui, counted bidirectionally). Time window granularity: Δt = 5 minutes; Forecast time range: H = 12 time windows in the future (i.e., the next hour). Historical window length: L = 24 time windows (i.e., the past 2 hours).

[0030] Step 2, Prediction Process Step S21: Cross-sectional feature acquisition At 36 provincial border sections, the following data is collected every 5 minutes through the existing ETC gantry system and traffic control stations: ; Event tagging data is accessed in real time from the event publishing system of the road network centers in each province.

[0031] Step S22: Construct a set of cross-sectional features The feature vectors of the 36 cross-sections are organized into an unordered set F(t). Since the set is unordered, the cross-sections do not need to be sorted before being input into the network.

[0032] Step S23: Construct the reverse mapping network Network hyperparameter settings: Input dimension: d=6 (six feature components); Embedding dimension: d emb =32; MLP φ Two layers, with a structure of 6 → 16 → 32, and the activation function is ReLU; MLP ψ It has three layers, with a structure of 32 → 64 → 32 → 16, where the last layer outputs dimension K. 2 =16 (4×4 OD matrix flattened), the activation function is ReLU except for the last layer, and Softmax is used in the last layer to ensure that the output is non-negative and the sum is 1 (before the total amount is normalized). Example of network forward computation (time window t): The 36 cross-sectional features were analyzed by MLP. φ This yields 36 32-dimensional embedding vectors z1, z2, ..., z 36 .

[0033] Aggregation layer: This results in a 32-dimensional global vector.

[0034] Decoding layer: This results in a 16-dimensional vector.

[0035] Reshape into a 4×4 matrix, set the diagonals to zero, and multiply by Normalization is complete.

[0036] Step S24: Physical Constraints and Model Training Construct the projection matrix A∈R 36×16 For example, section S01 connects Shanghai to Jiangsu Province, corresponding to OD index (0,1) (assuming index 0 = Shanghai, 1 = Jiangsu), and A is set... 01,(0,1) =1.

[0037] Training data: The cross-sectional feature data for the entire year of 2023 was used (one sample every 5 minutes, for a total of approximately 105,120 time windows). The actual origin-destination (OD) samples are derived from: ① provincial highway toll data (10% sampled to reconstruct OD); ② quarterly inter-provincial travel questionnaires (approximately 5,000 responses per province). These two types of data are then merged as a monitoring signal.

[0038] Hyperparameter settings: λ2=0.5, λ3=0.01, λ4=0.1, learning rate 0.001, optimizer Adam, training epochs 50.

[0039] Step S25: Short-term forecasting Train a TCN model independently for each cross-section. TCN structure: Number of input channels: 6 (feature dimension) Number of hidden channels: 32 Number of layers: 4, with expansion factors of 1, 2, 4, and 8 respectively. Kernel size: 3 Output: Predict a 6-dimensional feature sequence for the next 12 time steps (1 hour).

[0040] Taking section S01 as an example, the input is the feature sequence (6×24 matrix) of the past 24 time steps (2 hours), and the output is the feature sequence (6×12 matrix) of the future 12 time steps.

[0041] After obtaining the future feature predictions of 36 cross sections, they are aggregated into sets of 36 cross sections according to time steps t+1, t+2, ..., t+12, and then input into the back mapping network step by step to obtain the OD matrix prediction every 5 minutes in the next hour.

[0042] Step 3, Prediction Results In this embodiment, the method of the present invention (FSB) is compared and evaluated with the traditional four-stage method and the improved method based on the gravity model. The evaluation metrics are the mean absolute percentage error (MAPE) of the cross-provincial OD matrix and the response delay under the event scenario.

[0043] ; The results show that the method of the present invention is significantly superior to existing methods in terms of prediction accuracy, event response speed, and cold start capability.

[0044] Step 4, Event Response Example During a recent test, Zhejiang Province temporarily closed two provincial border crossings with Anhui Province (G56 Hangzhou-Ruili Expressway and G3 Beijing-Taipei Expressway) due to a major event. The announcement was made at 8:00 AM, and the flags for sections S35 and S36 were displayed at 8:05 AM. event The marker changed from 0 to 1, and the flow rates at both cross-sections dropped to near zero within the next 5 minutes.

[0045] In the next prediction cycle at 8:10, the model of this invention automatically inputs the features of these two cross-sections into the inverse mapping network. The output OD matrix shows that the travel demand from Zhejiang to Anhui decreased from approximately 420 pcu / h during normal periods to 65 pcu / h. Correspondingly, the predicted traffic flow on the alternative routes (via the G50 Shanghai-Chongqing Expressway) for these remaining trips increased by approximately 380 pcu / h. The entire update process was completed at 8:12, at least 3 days earlier than traditional methods.

[0046] Step 5, Applying Cold Start Mode For a newly constructed inter-provincial corridor (such as the newly opened Su-Tai Expressway connecting Jiangsu and Zhejiang), there are no historical OD samples in the initial stage of operation. A weakly supervised mode is adopted: Training was performed using only measured features of the cross-section. After accumulating 500 real OD samples on the 30th day after the system was launched, the system was switched to fully supervised fine-tuning, and the prediction accuracy decreased from the initial MAPE of 18.3% to 9.8%.

[0047] Example 2: Operating System and Device The technical solution of this invention can be deployed as a traffic demand forecasting system for cross-provincial administrative region highway networks, comprising: Data acquisition module: Detection equipment (ETC gantry, traffic control station, video detector) deployed at provincial border sections to collect data such as speed, traffic flow, and vehicle type in real time.

[0048] Event Access Module: Accesses the event publishing system of the road network center in each province via API to obtain event information such as traffic control, accidents, and weather.

[0049] Feature extraction module: Converts the raw detection data into cross-sectional feature vectors every 5 minutes.

[0050] Back-mapping prediction module: Deployed on a cloud server, it runs a pre-trained back-mapping network model.

[0051] Short-term prediction module: Deployed on edge nodes or in the cloud, it runs the TCN time-series prediction model for each cross section.

[0052] Output interface: Outputs real-time OD matrix, OD prediction sequence for the next 1-6 hours, event-triggered early warning, and OD change in standard data formats (such as JSON, Protobuf).

[0053] The system can be connected to provincial or national transportation operation coordination centers (TOCC) to provide decision support for cross-provincial collaborative management, dynamic guidance, and emergency dispatch.

[0054] Industrial applicability The traffic demand forecasting method for cross-provincial administrative region highway networks proposed in this invention has clear industrial applicability: High feasibility: The required data all come from existing traffic detection facilities (ETC gantries, traffic control stations, video detectors, etc.), requiring no additional hardware investment; the model structure is simple and can be deployed and run on existing traffic data center servers.

[0055] It has a wide range of applications: it can be applied to multiple fields such as the feasibility study of cross-provincial expressway construction, traffic flow analysis after the cancellation of provincial border toll stations, cross-provincial traffic control during major events, and emergency evacuation of regional road networks.

[0056] Significant commercial value: This invention can reduce the cost and barriers of cross-provincial traffic data sharing, improve the efficiency of cross-provincial traffic collaboration, and reduce traffic delays caused by provincial border control.

[0057] Highly scalable: This method is not limited to highway networks, but can be extended to other modes of transportation such as inter-provincial railway passenger flow forecasting and inter-provincial civil aviation travel analysis; the definition of cross-sectional features can be adjusted and expanded according to specific application scenarios.

[0058] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for planning and designing a highway network spanning provincial administrative regions, characterized in that, Includes the following steps: S1: Set up data collection sections at each inter-provincial passage to collect traffic flow feature vector data of the sections in real time; S2: Construct a cross-sectional feature set, and organize the feature vectors of all provincial boundary cross-sections within the same time window into an unordered set; S3: Construct a reverse mapping network model. The model adopts a set-to-vector mapping architecture, taking the cross-sectional feature set as input and outputting a vectorized representation of the cross-provincial OD matrix. It includes: a first multilayer perceptron, which maps each cross-sectional feature to an embedding vector; and an aggregation layer, which uses a symmetric function to aggregate all cross-sectional embedding vectors. The second multilayer perceptron is used to decode the aggregated global features into OD vectors; S4: Use the projection matrix to establish a linear relationship between the output OD vector and the measured total flow at each cross section, and use the projection error term as part of the loss function; S5: Train a time series prediction model independently for each cross section, predict the cross section feature set for multiple future time windows, and input it into the trained back-mapping network model to output the cross-provincial OD matrix for future time moments.

2. The method for planning and designing a highway network spanning provincial administrative regions according to claim 1, characterized in that, The feature vector At least include average speed Speed ​​standard deviation Average flow Large vehicles account for a certain percentage and the coefficient of variation of the headway It also includes event tagging features. The event marker feature is a binary variable used to indicate whether a traffic accident, temporary traffic control or severe weather event has occurred at the cross section within the corresponding time window; The feature vector is represented as follows: .

3. The method for planning and designing a highway network spanning provincial administrative regions according to claim 1, characterized in that, The reverse mapping network model is represented as follows: ,in For the vectorized representation of the cross-provincial OD matrix, The number of provincial-level administrative regions involved. Let be the eigenvector of section s. For the first multilayer perceptron, For the second multilayer perceptron, The total number of cross-sections; The aggregation layer employs an arithmetic mean operator to ensure that the model remains invariant to changes in the number of cross sections.

4. The method for planning and designing a highway network across provincial administrative regions according to claim 3, characterized in that, Step S3 further includes: reshaping the vectorized representation into an OD matrix. , is represented as: And force the diagonal elements to be set to zero: ; Perform total normalization: ,in For assignment operation, This represents the measured total flow rate across all cross-sections.

5. The method for planning and designing a highway network spanning provincial administrative regions according to claim 1, characterized in that, The projection matrix is ​​represented as: Defined as: if section s connects a directed channel from province a to province b, then ,otherwise ; The projection error term Defined as: ,in The s-th row of the projection matrix, The measured total flow rate at section s is given.

6. The method for planning and designing a highway network spanning provincial administrative regions according to claim 1, characterized in that, The loss function It also includes the supervision error term. sparse prior terms Time series smoothing term Defined as: ,in , , , All of these are hyperparameters. The following condition must be met: when a real OD sample exists, When no real OD sample exists, .

7. The method for planning and designing a highway network across provincial administrative regions according to claim 6, characterized in that, The supervision error term Defined as: ,in It is the Frobenius norm. This represents the number of training samples; The sparse priors Defined as: ,in It is an L1 norm; The time-series smoothing term Defined as: .

8. The method for planning and designing a highway network across provincial administrative regions according to claim 1, characterized in that, The temporal prediction model employs a temporal convolutional network, represented as follows: ,in For the length of the history window, It is the expansion factor. The kernel is a learnable convolutional kernel; the temporal prediction model uses the cross-sectional feature matrix of the past L time windows. As input, output the predicted feature sequence for the next H time windows. .

9. The method for planning and designing a highway network spanning provincial administrative regions according to claim 1, characterized in that, The method further includes: simultaneously generating cross-provincial OD prediction results with at least two time granularities through a multi-time resolution prediction output layer: Real-time prediction: Based on the current cross-sectional characteristics, output the cross-provincial OD matrix for the current time window with a delay of no more than 5 minutes. ; Short-term forecast: Outputs cross-provincial OD matrix sequences for the next 1 to 6 hours with a time resolution of 15 minutes. ; Event response prediction: When the event labeling features of any cross section When the value changes from 0 to 1, the OD (October Deficit) prediction for the next time window is automatically triggered, and the change in OD compared to the normal scenario is output. .

Images