Dynamic road network information fused reconstruction and extension road section congestion prediction system and method

Abstract

The invention belongs to the technical field of reorganization and extension road section congestion prediction systems, and particularly relates to a dynamic road network information fused reorganization and extension road section congestion prediction system and method. Traffic flow data and traffic organization scheme data of a reorganization and extension road section are obtained, and the traffic flow data at least comprise time information and average driving speed of the road section; the traffic organization scheme data at least comprises the lane number, the speed limit value and the road section passing permission state in the construction stage, a time delay mechanism is adopted to divide the preprocessed data set, the time delay value is set to be M, the data set with the original time window being [p, q] is divided into M + 1 data subsets in a sliding mode, and M is a positive integer; enabling the single-row data of each data subset to reflect the evolution process of the road section state at continuous M + 1 moments, and forming model input features.

Description

Technical Field

[0001] This invention belongs to the technical field of congestion prediction systems for road reconstruction and expansion sections, and particularly relates to a congestion prediction system and method for road reconstruction and expansion sections that integrates dynamic road network information. Background Technology

[0002] During road reconstruction and expansion, road alignment indicators and traffic organization plans exhibit significant dynamism. For example, the number of lanes often decreases from four lanes to two lanes, and speed limits are gradually adjusted to 80 km / h, 60 km / h, or even 40 km / h as construction progresses. Furthermore, road permit status changes frequently due to construction closures or diversions. These factors directly lead to the continuous evolution of traffic flow dependencies between road network nodes. Traditional congestion prediction methods are mostly based on fixed road network topology models, such as graph convolutional networks using static adjacency matrices. These methods fail to explicitly integrate the dynamic adjustments of traffic organization into spatial correlation modeling, making it difficult to accurately depict the real-time impact relationships between road segments under construction scenarios. This results in significant discrepancies between prediction results and actual traffic conditions.

[0003] Existing predictive models still suffer from incomplete spatiotemporal modeling or insufficient interpretability. For instance, time-series models such as LSTM focus solely on the temporal evolution of traffic flow, completely ignoring the spatial relationships between road segments. While traditional dynamic Bayesian networks (DBNs) can capture some dependencies, they lack prior knowledge of traffic management in construction scenarios, such as blacklists of closed road segments and whitelists of permitted road segments, limiting model robustness. Furthermore, most models fail to identify congestion source nodes and propagation paths, hindering traffic management departments' ability to trace and control congestion and failing to meet the practical needs of refined traffic scheduling for road reconstruction and expansion projects. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned technical problems by providing a system and method for predicting congestion on reconstructed and expanded road sections that integrates dynamic road network information.

[0005] A congestion prediction system for reconstructed and expanded road sections that integrates dynamic road network information includes a data acquisition module, a data preprocessing module, a dynamic adjacency matrix generation module, a spatiotemporal feature extraction module, and a congestion prediction output module. The data acquisition module is used to acquire traffic flow data and traffic organization plan data of the road section to be reconstructed and expanded. The traffic flow data includes at least the time information and average driving speed of the road section, and the traffic organization plan data includes at least the number of lanes, speed limit and road section traffic permit status during the construction phase. The data preprocessing module is used to divide the dataset using a time delay mechanism. The time delay value is set to M, and the dataset with the original time window of [p,q] is divided into M+1 data subsets. This allows the single row of data in each data subset to reflect the evolution of the road segment status over M+1 consecutive time periods, forming the model input features. The dynamic adjacency matrix generation module is used to construct a blacklist and whitelist based on the road segment permit status in the traffic organization scheme as a prior constraint, and generate a dynamic adjacency matrix by combining the statistical dependency relationship between road alignment indicators and traffic flow data. The spatiotemporal feature extraction module is used to take the dynamic adjacency matrix as graph structure input to construct a stacked spatiotemporal graph convolutional network structure. The number of layers in the STGCN is the same as the number of layers in the stacked state propagation network. The STGCN extracts spatiotemporal features from the input features and outputs the features of each layer. ; The congestion prediction output module is used to process features from each layer. The summation is mapped to the sample space through a fully connected layer to obtain the traffic state characteristics at the predicted time. At the same time, the congestion index is calculated based on the ratio of the average driving speed of the road segment to the maximum speed limit. Combined with a preset threshold, the congestion status is determined, and finally the congestion prediction result of the road segment to be reconstructed and expanded is output.

[0006] In view of this, the present invention provides a method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information, comprising the following steps: Step 1: Obtain traffic flow data and traffic organization plan data for the road section to be reconstructed and expanded. The traffic flow data shall include at least the time information and average driving speed of the road section, and the traffic organization plan data shall include at least the number of lanes, speed limit and road section traffic permit status during the construction phase. Step 2: The preprocessed dataset is divided using a time delay mechanism. The time delay value is set to M. The dataset with the original time window of [p,q] is divided into M+1 data subsets. Each data subset's single row of data reflects the evolution of the road segment status over M+1 consecutive time periods, forming the model input features. Step 3: Using the road segment permit status in the traffic organization scheme to construct a blacklist and whitelist as a priori constraints, and combining the statistical dependency relationship between road alignment indicators and traffic flow data, a dynamic adjacency matrix is ​​generated through the following steps; Step 4: The generated dynamic adjacency matrix is ​​used as input to the graph structure to construct a stacked spatiotemporal graph convolutional network (STGCN). The number of layers in the STGCN is the same as the number of layers in the stacked state propagation network. The STGCN extracts spatiotemporal features from the input features and outputs the features of each layer. ; For the features of each layer output in step four The summation is performed, and the summation result is mapped to the sample space through a fully connected layer to obtain the traffic state characteristics at the predicted time. At the same time, the congestion index is calculated based on the ratio of the average driving speed of the road segment to the maximum speed limit. Combined with a preset threshold, the congestion status is determined, and finally the congestion prediction result of the road segment to be reconstructed and expanded is output.

[0007] Preferably, step three specifically includes the following steps: Step S11: Set the adjacency matrix elements of the node's temporal self-connection Let downstream node k point to the adjacency matrix element of upstream node i. ; Step S12: Describe the node pairs using a multivariate Gaussian distribution. The conditional probability distribution, based on the observation dataset. Computational network structure Scoring function The scoring function is the joint product of the conditional probabilities of all nodes; Step S13: Under the constraints of the previous step, the max-minhill-climbing heuristic algorithm is used to traverse the network structure and select the optimal structure with the largest score function. Convert it into a dynamic adjacency matrix ; Preferably, the traffic flow data in step one comes from the map service provider's API interface, the data time granularity is 10 minutes, and the number of covered road segments is divided into N according to the road network topology of the reconstruction and expansion area; Data preprocessing includes missing value imputation, using linear interpolation to handle null values ​​with a missing rate of less than 1%; the dataset is divided into training and test sets in an 8:2 ratio.

[0008] Preferably, in step five, the preset threshold ρ for determining the congestion state is 0.5. When the congestion index... When the time is t, it is determined that road segment i is in a congested state.

[0009] Preferably, the scoring function described in step three The calculation logic is as follows: ; in, For node pairs The conditional probability; and C is the dataset covariance, equal .

[0010] Preferably, the time delay value M in step four is 1, 2 or 3; The top-level feature output of the stacked STGCN ; Where σ is the activation function, Let m be the dynamic adjacency matrix corresponding to the m-th time delay. The input features at time tM are... The weight parameter is the one corresponding to the m-th time delay.

[0011] Preferably, the mapping function of the fully connected layer in step four is: ,in Let ω be a linear weighted summation function, where ω is the weight of the fully connected layer. This outputs the summation results of the features for each layer of STGCN.

[0012] Preferably, the training time unit for the congestion prediction model in step five is 10 minutes. Using step sizes of 12 and 6, predict traffic conditions for the next hour using historical traffic conditions from the past 2 hours; The performance evaluation metrics for the prediction results include mean absolute error, root mean square error, and mean absolute percentage error, and the MAPE < 0.05 for each prediction step size.

[0013] Preferably, it also includes a congestion propagation interpretability analysis step: the source node of congestion is identified by the dynamic adjacency matrix generated in step three, the source node is used as the first-generation parent node, potential child nodes are matched in the network at the current and next time step to form parent-child node pairs, a network-level congestion propagation network is constructed, and the flow path of congestion information is displayed.

[0014] Preferably, the application scenario for the road section to be reconstructed and expanded is that during the reconstruction and expansion construction, the road section is changed from the original two-way four-lane traffic to two-way two-lane traffic, and the speed limit in the construction area is adjusted step by step as the construction progresses. The blacklist and whitelist are dynamically updated with the construction stage.

[0015] The beneficial effects of this invention are: The proposed dynamic Bayesian graph convolutional neural network demonstrates outstanding congestion prediction accuracy on road sections undergoing reconstruction and expansion, significantly outperforming baseline models such as LSTM, DBN, and STGCN. Its mean absolute error is as low as 2.4112, and its root mean square error is as low as 3.8017. Furthermore, the mean absolute percentage error at each prediction step is less than 0.05, achieving a prediction accuracy exceeding 95%. Simultaneously, the model incorporates dynamic changes in traffic organization as a priori constraints into the dynamic adjacency matrix generation process through a blacklist / whitelist mechanism. This effectively adapts to the continuous evolution of road network dependencies during reconstruction and expansion. Even in scenarios with sudden changes in traffic conditions, the prediction curve still closely matches the actual situation, achieving the goal of accurately predicting future traffic conditions one hour in advance based on two hours of historical data.

[0016] Furthermore, the model possesses excellent interpretability for congestion propagation. It can accurately identify congestion source nodes through a dynamic adjacency matrix, and using the source node as the initial parent node, it matches potential child nodes in the network at the current and next time steps to form parent-child node pairs, constructing a network-level congestion propagation network that clearly demonstrates the flow path of congestion information. For scenarios where highways are reduced from four lanes to two lanes during reconstruction and expansion, the model can dynamically respond to the need for progressively adjusting speed limits as construction progresses. This provides traffic management departments with support for congestion source tracing and propagation path tracking, meeting the practical application needs of refined traffic scheduling and control in reconstructed and expanded road sections. Attached Figure Description

[0017] Figure 1 This is a diagram of the DBGCN model framework of the present invention; Figure 2 This is an example diagram of the new dataset under the three-time delay of this invention; Figure 3 Example diagram of blacklist / whitelist settings in this invention; Figure 4 This is a diagram showing the average prediction results for all road segments in this invention; Figure 5 This is a map showing the prediction results for a portion of the road sections according to the present invention; Figure 6 This is a schematic diagram illustrating the congestion propagation structure reasoning under external influences according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0019] It should be noted that all directional and positional terms used in this invention, such as "up," "down," "left," "right," "front," "back," "vertical," "horizontal," "inner," "outer," "top," "lower," "lateral," "longitudinal," and "center," are only used to explain the relative positional relationships and connections between components in a specific state (as shown in the accompanying drawings). They are merely for the convenience of describing the invention and do not require the invention to be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on the invention. Furthermore, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0020] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0022] A congestion prediction system for reconstructed and expanded road sections that integrates dynamic road network information includes a data acquisition module, a data preprocessing module, a dynamic adjacency matrix generation module, a spatiotemporal feature extraction module, and a congestion prediction output module. The data acquisition module is used to acquire traffic flow data and traffic organization plan data of the road section to be reconstructed and expanded. The traffic flow data includes at least the time information and average driving speed of the road section, and the traffic organization plan data includes at least the number of lanes, speed limit and road section traffic permit status during the construction phase. The data preprocessing module is used to divide the dataset using a time delay mechanism. The time delay value is set to M, and the dataset with the original time window of [p,q] is divided into M+1 data subsets. This allows the single row of data in each data subset to reflect the evolution of the road segment status over M+1 consecutive time periods, forming the model input features. The dynamic adjacency matrix generation module is used to construct a blacklist and whitelist based on the road segment permit status in the traffic organization scheme as a prior constraint, and generate a dynamic adjacency matrix by combining the statistical dependency relationship between road alignment indicators and traffic flow data. The spatiotemporal feature extraction module is used to take the dynamic adjacency matrix as graph structure input to construct a stacked spatiotemporal graph convolutional network structure. The number of layers in the STGCN is the same as the number of layers in the stacked state propagation network. The STGCN extracts spatiotemporal features from the input features and outputs the features of each layer. ; The congestion prediction output module is used to process features from each layer. The summation is mapped to the sample space through a fully connected layer to obtain the traffic state characteristics at the predicted time. At the same time, the congestion index is calculated based on the ratio of the average driving speed of the road segment to the maximum speed limit. Combined with a preset threshold, the congestion status is determined, and finally the congestion prediction result of the road segment to be reconstructed and expanded is output.

[0023] To address the challenges posed by the dynamic changes in road alignment parameters and traffic organization during road reconstruction and expansion projects to congestion prediction, this invention proposes a dynamic Bayesian graph convolutional network to learn the traffic state network, such as... Figure 1 As shown. The core idea is to explicitly integrate the dynamic adjustment of traffic organization into the graph structure construction process to more accurately depict the spatial dependencies between road network nodes under reconstruction and expansion conditions. Specifically, this invention adopts a two-stage modeling paradigm: First, network structure learning is performed, using a dynamic Bayesian network to analyze the statistical dependencies between road alignment indicators and traffic flow data. By setting blacklists and whitelists, key traffic organization changes are used as prior constraints to guide the DBN to learn and generate an adjacency matrix that reflects the current traffic management status. Subsequently, network parameter learning is performed, using the dynamic adjacency matrix output by the DBN that considers traffic organization information as the graph structure input. A graph convolutional neural network is used to extract spatiotemporal features, ultimately achieving the prediction of congestion status on the reconstructed and expanded road sections.

[0024] This invention uses running speed to characterize traffic conditions. To explore how earlier traffic conditions affect current traffic conditions, this invention employs a time lag mechanism during dataset preparation: increasing the time lag by 1 moves the dataset's time window forward by one step. Taking the dataset in the figure below as an example, it represents data samples from time p to time q. If the time lag is set to M, the dataset will be slidingly divided into M+1 subsets.

[0025] If we take M=3 as an example, then the dataset used in the study would be transformed into: Figure 2 The structure shown. The size of the dataset is from the original... Transform into In the new data structure, each row of data reflects the evolution of the road segment's state over four consecutive time points, which facilitates the second module's learning of the dynamic relationships between road segments.

[0026] Adjacency matrices are a digital representation of traffic networks, essentially a directed acyclic graph. However, changes in road alignment and traffic organization during construction phases alter the road condition relationships between road segments. Therefore, adjacency matrices cannot be simply equated with a fixed road network topology. To address this, based on a data-driven approach, this invention uses dynamic Bayesian networks to infer the adjacency relationships between road segments under the influence of changes in road alignment and traffic organization. Figure 1As shown in (b). The following section uses a two-time-slice DAG as an example to illustrate the congestion propagation structure learning process.

[0027] Step 1: Constrain the directed causal relationships between nodes based on empirical knowledge. During congestion propagation, two natural constraints are: (1) Due to changes in road alignment or traffic organization, once congestion occurs, it will only propagate from the source node to its upstream neighbor nodes. (2) The current state of a node is influenced by its previous state, i.e., node self-connection. Therefore, in any two-time-slice DAG, the adjacency matrix is ​​predefined as follows:

[0028] Step 2: Construct a scoring function to find the optimal structure. The essence of the optimal structure is multiple parent-child node pairs. A consortium, in which, For each node pair Due to the parent node Non-uniqueness, right The stochastic process exerting the effect is appropriately described using a multivariate Gaussian distribution. Therefore, given the observed data... Under these conditions, node pairs The conditional probability distribution can be written as:

[0029] in, This represents the covariance of the road segment data; equal .

[0030] Known observation dataset Then a certain network structure The corresponding scoring function should be the joint probability of all node pairs, which can be expressed as:

[0031] Due to each structure The formation of is equally likely, then Since it is a constant, equation (5.9) is equivalent to:

[0032] Step 3: After determining the scoring function, network structure learning becomes a search problem. Under the constraints of Step 1, the max-minhill-climbing heuristic is used to traverse all possible network structures. The structure with the highest score is determined as the optimal structure. This is then converted into the adjacency matrix required for GCN, as follows:

[0033] This section, through the learned graph structure, treats pairs of road segments with influence relationships as parent-child node pairs, introduces a spatiotemporal graph convolutional network, and uses a deep network approach to measure the degree of influence between parent and child nodes.

[0034] To accommodate adjacency matrices at different levels, this invention constructs a stacked structure based on STGCN (red box). Figure 1 (c) In the diagram, the number of layers in the STGCN structure is the same as the number of layers in the stacked state propagation network. Within each layer, traffic states are propagated downwards through each STGCN module, constituting information transmission. It is important to note the time delay value. It is an adjustable parameter to facilitate the analysis of the long-term effects of traffic conditions.

[0035] Taking the top layer as an example, the time lag is M, and the model will choose the current time period. Before The data from each time period is used as the feature input, and is passed down time by time through STGCN. The final feature output is:

[0036] Finally, the feature outputs of each layer are summed to obtain... And use a fully connected layer to map it to the sample space, as follows:

[0037] in, The fully connected layer is essentially a linear weighted summation function; Indicates the weight.

[0038] The research area of ​​this invention is located on a highway, which is divided into 164 road segments according to its topology.

[0039] The data used in this invention includes traffic flow data and traffic organization schemes. The traffic flow data comes from a map service provider. Data on time and average speed per minute for 164 topological road segments within the study area, at a granular level of 10 minutes, were obtained via API interface from October 22nd to November 10th, 2024. Verification showed that the data missing rate was less than 1%, and a small number of missing values ​​were filled using linear interpolation, further ensuring the integrity and quality of the data.

[0040] The traffic organization schemes are all derived from the official reconstruction and expansion project design scheme and construction organization plan documents of the highway. These documents record the implementation details of each stage of the project. Based on the traffic organization design in the reconstruction and expansion scheme, this invention manually constructs corresponding blacklists and whitelists for each construction stage, thereby inputting the actual traffic organization into the model. The whitelist contains road sections that allow direct passage in the scheme, while the blacklist contains connections that are prohibited due to closure or diversion. Figure 4 For example, since no construction is underway on the left side of the road, normal connections from upstream to downstream are whitelisted and not blacklisted. However, construction is underway on road segment 2 on the right side, so segments 1 to 2 and 2 to 3 are blacklisted. This ensures that information about the closure of segment 2 is input into the subsequent model. The blacklist and whitelist are dynamically updated as construction progresses and are input during network structure learning.

[0041] Ultimately, this invention divides one month's data into a training set and a test set in an 8:2 ratio. The data format for each time step in the dataset includes the actual speed value, the number of lanes, and the speed limit. The actual speed value is a variable that changes over time and is also the model's prediction target; the number of lanes and the speed limit are adjusted according to the construction progress and remain constant within the same construction phase.

[0042] Table 1. Dataset format;

[0043] This invention selects the average vehicle speed of a road segment to characterize the traffic conditions of that segment. To avoid the influence of road heterogeneity on the model, we normalize the speed as shown in formula (12):

[0044] in Indicates road segment The average driving speed during the t-th time interval, This refers to the maximum speed limit for that section of road.

[0045] in addition, It is a continuous value between 0 and 1, which can linearly represent the degree of congestion. This invention calls it the congestion index. The lower the value, the lower the speed and the more congested the road segment. Conversely, the higher the value, the more uncongested the road segment. According to Saberi et al., congestion status can be determined by the following formula:

[0046] To more comprehensively identify the entire process from the onset to the dissipation of congestion, this invention refers to relevant literature on traffic congestion prediction and determines a threshold. It is 0.5, that is, with The criteria determine that the current road segment is in a congested state.

[0047] Since congestion indices are derived from traffic speeds, congestion propagation inference is achieved through traffic speed prediction under special circumstances. Many current studies involve traffic congestion propagation inference or short-term traffic prediction, using different data structures and requiring different data fields. Therefore, the DBGCN model proposed in this invention cannot be compared with all models proposed to date. Our primary goal is to demonstrate whether the dynamic adjacency matrix learned after considering dynamic changes in traffic organization can enhance the model's inference ability. Secondly, it aims to verify whether early traffic conditions affect the current traffic conditions; this is also the purpose of constructing a stacked model framework in this invention. Therefore, this invention selects three representative models as benchmarks: (1) LSTM: LSTM is a variant of Recurrent Neural Network (RNN) specifically designed to solve long-term dependency problems, making it well-suited for processing time series data. LSTM can retain historical information across multiple time steps, giving it a unique advantage in time series prediction of congestion indices. Since the evolution of traffic flow and congestion often exhibits long-term dependencies, LSTM can capture the impact of past congestion states on the congestion index at present and future times. Therefore, LSTM can effectively model the spatiotemporal dynamics of congestion states when dealing with time series prediction problems of traffic congestion.

[0048] (2) Dynamic Bayesian Network. A dynamic Bayesian network is a directed probabilistic graphical model that uses probability theory to describe the transmission of different traffic states between adjacent road segments. For continuous input features, this transmission process uses a Gaussian mixture model as the model, in which child nodes can infer the state of the child node by aggregating the information of its parent node.

[0049] (3) STGCN: STGCN is a deep learning model that integrates graph convolution and temporal convolution, capable of simultaneously modeling the spatial relationships and temporal evolution of traffic data. In this invention, STGCN utilizes the topology of the road network, extracts the spatial dependencies between road segments through graph convolution, and combines temporal convolution to capture the temporal variation characteristics of congestion, thereby achieving accurate prediction of future traffic conditions. This model is suitable for traffic scenarios with complex spatiotemporal dependencies and has strong expressive power and predictive performance.

[0050] In the data input layer, as the time lag value M increases, traffic state information from earlier time slices will be aggregated into the current traffic state. Considering that the duration of traffic congestion is finite, the value of M should not be set too large. This invention sets the value of M to 1, 2, and 3.

[0051] In model training, this invention uses the mean squared error between the true and inferred values ​​as the loss function, employs the Adam optimization method to update the parameters in the graph convolutional network, and sets the learning rate to 10⁻³, which accelerates the convergence of the training process. Furthermore, early stopping is used to prevent overfitting. The model training process stops if the mean squared error (MSE) does not significantly improve within 10 consecutive iterations.

[0052] In this invention, historical data from the past month was used for model training, with a time unit of 10 minutes. Input data included road topology and traffic conditions. The model's input and output step sizes were 12 and 6, respectively, meaning it used two hours of historical traffic conditions to predict the traffic conditions for the next hour. MAE, RMSE, and MAPE were selected as model performance evaluation metrics, and the results are shown in Table 2. As the prediction step size increased, the prediction accuracy decreased slightly, but the MAPE remained below 0.05, indicating a prediction accuracy above 95%.

[0053] Table 2. Model performance under different prediction step sizes;

[0054] To verify the predictive performance of the proposed DBGCN model under scenarios involving dynamic changes in road alignment indicators and traffic organization, comparative experiments were conducted on the aforementioned real-world highway dataset. Mean absolute error and root mean square error were used to evaluate the performance of the congestion propagation inference model in this study.

[0055] Table 3 Performance metrics for different models;

[0056] Table 3 shows the comparison results of the DBGCN model with several baseline models in terms of mean absolute error (MAE) and root mean square error (RMSE). As can be seen from the table, the DBGCN model proposed in this invention significantly outperforms all baseline models in both MAE and RMSE, achieving the lowest MAE and RMSE values ​​respectively, indicating a clear advantage in overall prediction accuracy. Furthermore, it is worth noting that models such as DBN and STGCN, which consider spatial relationships, actually perform worse than the traditional time-series prediction model LSTM. This indicates that changes in traffic organization in the reconstruction and expansion area do indeed pose a challenge to the traditional model's learning of spatial relationships, while the model proposed in this invention effectively solves this problem and improves prediction accuracy.

[0057] To further analyze the model's predictive performance, Figure 5 and Figure 6The average prediction results of the DBGCN model across all 164 road segments and its time-period prediction performance on some road segments are presented. It can be seen that the DBGCN prediction curve closely matches the actual traffic condition trends, especially maintaining high fitting accuracy during abrupt changes in traffic conditions, indicating that the model effectively captures the spatiotemporal evolution of macroscopic traffic flow.

[0058] Another advantage of the DBGCN proposed in this invention is its good interpretability. Specifically, the model can identify the source nodes causing congestion based on changes in congestion patterns. These source nodes may be construction zones or zones experiencing sudden incidents. Once the source node is identified, the model uses it as the initial parent node and then searches for potential child nodes. These child nodes exist in the network for the current and next time periods, forming parent-child node pairs with the parent node. In this way, the model can gradually construct a network-level congestion propagation network, demonstrating the flow of congestion information within the propagation network.

[0059] Taking a partially merging section of road as an example, Figure 6 This section presents partial inference results from the adjacency matrix construction during a congestion event. Red nodes represent the current congestion source node, while blue nodes represent potential child nodes for the next moment. Arrows indicate the propagation path from parent to child nodes, clearly demonstrating how congestion gradually spreads from the source node to the entire network.

[0060] This invention addresses the difficulty of predicting traffic conditions during road reconstruction and expansion by proposing a dynamic Bayesian graph convolutional neural network. This network aims to deeply integrate road alignment indicators with the dynamic changes in traffic organization to achieve accurate real-time prediction of congestion conditions on road sections undergoing reconstruction and expansion.

[0061] Experimental results of the proposed DBGCN model on a real highway dataset demonstrate that it significantly outperforms several baseline models, including LSTM, DBN, and STGCN, in terms of both mean absolute error (MAE) and root mean square error (RMSE). Specifically, the DBGCN model achieves the lowest MAE and RMSE values, indicating a clear advantage in prediction accuracy.

[0062] Furthermore, this invention incorporates a blacklist / whitelist mechanism to integrate prior knowledge of traffic organization into the graph structure learning process. This allows the proposed DBGCN model to respond to changes in road alignment and traffic organization, generating a dynamic adjacency matrix that conforms to the current construction environment. This feature makes the model more robust and flexible in predicting traffic flow in complex dynamic scenarios.

[0063] In summary, the DBGCN model proposed in this invention provides an effective solution for predicting congestion in road sections undergoing reconstruction and expansion. By deeply integrating road alignment indicators with dynamic traffic organization information, DBGCN not only improves prediction accuracy but also enhances the understanding of complex traffic environments. Future research could explore how to incorporate more external factors into the model to further improve prediction performance. Simultaneously, optimizing the model structure and parameter settings to improve computational efficiency are also directions worthy of in-depth investigation. In conclusion, this invention provides new ideas and technical support for intelligent traffic management and urban planning.

[0064] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A congestion prediction system for reconstructed and expanded road sections integrating dynamic road network information, characterized in that: It includes a data acquisition module, a data preprocessing module, a dynamic adjacency matrix generation module, a spatiotemporal feature extraction module, and a congestion prediction output module; The data acquisition module is used to acquire traffic flow data and traffic organization plan data of the road section to be reconstructed and expanded. The traffic flow data includes at least the time information and average driving speed of the road section, and the traffic organization plan data includes at least the number of lanes, speed limit and road section traffic permit status during the construction phase. The data preprocessing module is used to divide the dataset using a time delay mechanism. The time delay value is set to M, and the dataset with the original time window of [p,q] is slidably divided into M+1 data subsets, so that the single row of data in each data subset reflects the evolution process of the road segment status at consecutive M+1 time points, forming the model input features. The dynamic adjacency matrix generation module is used to construct a blacklist and whitelist based on the road segment permit status in the traffic organization scheme as a prior constraint, and generate a dynamic adjacency matrix by combining the statistical dependency relationship between road alignment indicators and traffic flow data. The spatiotemporal feature extraction module is used to take the dynamic adjacency matrix as graph structure input to construct a stacked spatiotemporal graph convolutional network structure. The number of layers in the STGCN is the same as the number of layers in the stacked state propagation network. The STGCN extracts spatiotemporal features from the input features and outputs the features of each layer. ; The congestion prediction output module is used to process features from each layer. The summation is mapped to the sample space through a fully connected layer to obtain the traffic state characteristics at the predicted time. At the same time, the congestion index is calculated based on the ratio of the average driving speed of the road segment to the maximum speed limit. Combined with a preset threshold, the congestion status is determined, and finally the congestion prediction result of the road segment to be reconstructed and expanded is output.

2. A method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information, based on the system described in claim 1, characterized in that: Includes the following steps: Step 1: Obtain traffic flow data and traffic organization plan data for the road section to be reconstructed and expanded. The traffic flow data shall include at least the time information and average driving speed of the road section, and the traffic organization plan data shall include at least the number of lanes, speed limit and road section traffic permit status during the construction phase. Step 2: The preprocessed dataset is divided using a time delay mechanism. Let the time delay value be M. The dataset with the original time window of [p,q] is divided into M+1 data subsets. Each data subset's single row of data reflects the evolution of the road segment status over M+1 consecutive time periods, forming the model input features. Step 3: Using the road segment permit status in the traffic organization scheme to construct a blacklist and whitelist as a priori constraints, and combining the statistical dependency relationship between road alignment indicators and traffic flow data, a dynamic adjacency matrix is ​​generated through the following steps; Step 4: The generated dynamic adjacency matrix is ​​used as input to the graph structure to construct a stacked spatiotemporal graph convolutional network (STGCN). The number of layers in the STGCN is the same as the number of layers in the stacked state propagation network. The STGCN extracts spatiotemporal features from the input features and outputs the features of each layer. ; For the features of each layer output in step four The summation is performed, and the summation result is mapped to the sample space through a fully connected layer to obtain the traffic state characteristics at the predicted time. At the same time, the congestion index is calculated based on the ratio of the average driving speed of the road segment to the maximum speed limit. Combined with a preset threshold, the congestion status is determined, and finally the congestion prediction result of the road segment to be reconstructed and expanded is output.

3. The method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information as described in claim 2, characterized in that: Step three specifically includes the following steps: Step S11: Set the adjacency matrix elements of the node's temporal self-connection Let downstream node k point to the adjacency matrix element of upstream node i. ; Step S12: Describe the node pairs using a multivariate Gaussian distribution. The conditional probability distribution, based on the observation dataset. Computational network structure Scoring function The scoring function is the joint product of the conditional probabilities of all nodes; Step S13: Under the constraints of the previous step, the max-minhill-climbing heuristic algorithm is used to traverse the network structure and select the optimal structure with the largest score function. Convert it into a dynamic adjacency matrix .

4. The method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information as described in claim 3, characterized in that: The traffic flow data mentioned in step one comes from the map service provider's API interface, with a data time granularity of 10 minutes, and the number of covered road segments is divided into N according to the road network topology of the reconstruction and expansion area; Data preprocessing includes missing value imputation, using linear interpolation to handle null values ​​with a missing rate of less than 1%; the dataset is divided into training and test sets in an 8:2 ratio.

5. The method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information as described in claim 4, characterized in that: In step five, the preset threshold ρ for determining the congestion state is 0.

5. When the congestion index... When the time is t, it is determined that road segment i is in a congested state.

6. The method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information as described in claim 5, characterized in that: The scoring function described in step three The calculation logic is as follows: ; in, For node pairs The conditional probability; and C is the dataset covariance, equal .

7. The method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information as described in claim 6, characterized in that: The time delay value M mentioned in step four can be 1, 2, or 3; The top-level feature output of the stacked STGCN ; Where σ is the activation function, Let m be the dynamic adjacency matrix corresponding to the m-th time delay. Let be the input features at time tM. The weight parameter is the one corresponding to the m-th time delay.

8. The method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information as described in claim 7, characterized in that: The mapping function of the fully connected layer in step four is: ,in Let ω be a linear weighted summation function, where ω is the weight of the fully connected layer. This provides the summation results of the output features for each layer of STGCN.

9. The method for predicting congestion on reconstructed and expanded road sections by integrating dynamic road network information as described in claim 8, characterized in that: The training time unit for the congestion prediction model described in step five is 10 minutes. Using step sizes of 12 and 6, predict traffic conditions for the next hour using historical traffic conditions from the past 2 hours; The performance evaluation indicators of the prediction results include mean absolute error, root mean square error and mean absolute percentage error, and MAPE < 0.05 at each prediction step size. It also includes the congestion propagation interpretability analysis step: the congestion source node is identified through the dynamic adjacency matrix generated in step three, the source node is used as the first generation parent node, potential child nodes are matched in the network at the current and next time moment to form parent-child node pairs, and a network-level congestion propagation network is constructed to show the flow path of congestion information. The application scenario for the road section to be reconstructed and expanded is that during the reconstruction and expansion construction period, the road section is changed from the original two-way four-lane traffic to two-way two-lane traffic, and the speed limit in the construction area is adjusted step by step as the construction progresses. The blacklist and whitelist are dynamically updated with the construction stage.

Images