Method for predicting local road network flow change under influence of traffic events

Abstract

The invention relates to the technical field of intelligent traffic, in particular to a local road network flow change prediction method under the influence of traffic events. Comprising the steps of generating a high-quality mixed data set, constructing a road network dynamic graph structure fused with traffic event characteristics, constructing and training an end-to-end space-time prediction model based on self-adaptive dynamic partitioning and real-time data closed-loop and online rolling prediction. Through an innovative data generation strategy of'historical data calibration + high-fidelity traffic event simulation ', a large-scale, high-density, accurate-label and consistent-physical-significance'traffic event-traffic flow' mixed space-time synchronization data set is obtained, and the data bottleneck problem of road network flow prediction under traffic events is effectively solved. And a solid data foundation is laid for training and verification of a complex prediction model.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically a method for predicting local road network traffic flow changes under the influence of traffic events. Background Technology

[0002] Traffic flow prediction for highway networks is the core of intelligent traffic management. Accurate traffic flow prediction provides crucial decision support for route guidance, congestion warning, and rapid accident response. Existing traffic flow prediction models are mainly based on historical traffic data and utilize deep learning techniques such as recurrent neural networks, graph convolutional networks, and attention mechanisms to mine the spatiotemporal correlation of traffic flow.

[0003] Currently, the more advanced related technologies mainly fall into two categories: First, "A Traffic Flow Prediction Method Based on Multivariate Spatiotemporal Feature Extraction," which employs a multivariate spatiotemporal feature extraction model. In the temporal dimension, it utilizes a gated convolution-improved Res2Net module to capture traffic flow patterns over time. In the spatial dimension, it captures global, geographical neighborhood, and semantically similar spatial correlations through spatial attention mechanisms and dual-view Geom-GCN. In the spatiotemporal interaction dimension, it captures dynamic interaction relationships across time slices through a dynamic spatiotemporal synchronization aggregation module. This method is suitable for routine scenarios with relatively stable traffic patterns, but it lacks effective modeling capabilities for abrupt and non-stationary changes in traffic flow caused by traffic events, leading to a significant decrease in prediction accuracy. Second, "A Multi-Task Traffic Flow Prediction Method and System for Highway Networks Based on Deep Learning," which uses multi-task learning technology to characterize the common features of inbound and outbound traffic, achieving collaborative prediction of toll station inbound and outbound traffic. However, its spatial modeling units and task definitions are relatively simple, lacking the capture of dynamic propagation effects of global traffic flow on road segments and arterial networks. Furthermore, it treats the road network topology as static, failing to effectively extract the dynamic change patterns of traffic flow on different road segments under traffic events.

[0004] With increasingly complex road traffic networks and frequent sudden traffic incidents, the operational efficiency of road networks has declined sharply, and secondary accidents have even occurred. Accurately predicting the dynamic changes in road network traffic flow under the influence of traffic incidents is crucial for achieving proactive traffic management, dynamic route guidance, and emergency response. However, existing technologies face two fundamental challenges: First, there is the problem of data sparsity. Traditional road network traffic flow data mainly relies on sparsely deployed fixed sensors (such as geomagnetic coils, microwave radar, and ETC gantries) for collection. However, traffic events are random and have spatiotemporal uncertainties, often falling within the "blind spots" between fixed sensors. This makes it difficult to effectively capture sudden changes in traffic flow near the event (such as sudden drops in flow, rapid speed reductions, and queue formation). The performance of machine learning models heavily depends on the scale and quality of training data. Currently, there is a lack of high-density, high-quality traffic data near traffic events, making it difficult for models to learn and characterize the real propagation mechanism of the event's impact, causing prediction accuracy to fail at critical moments.

[0005] Second, there is the issue of model adaptability. Current mainstream traffic flow prediction models (such as time-series models based on RNN, LSTM, and GRU, and spatiotemporal models based on GCN and GAT) are primarily designed for normalized traffic flows with strong periodicity and stability. They excel at capturing regular patterns such as morning and evening rush hours, but struggle to effectively model abrupt, non-periodic, and drastically fluctuating traffic flow triggered by sudden events. These models typically assume that the spatiotemporal relationships within the road network are fixed or change slowly, failing to dynamically capture new patterns of rapid and uneven propagation of traffic impacts within the road network under event disturbances.

[0006] To address the problem of data sparsity, existing research has attempted to generate data through methods such as data interpolation and Generative Adversarial Networks (GANs). However, the generated data often suffers from defects in physical consistency. Regarding model adaptability, while some studies have introduced attention mechanisms or dynamic graph structures, these are mostly limited to fine-tuning the relationships between nodes and have failed to provide an adaptive mechanism for dynamically reconstructing the spatiotemporal modeling units of the road network based on the impact of traffic events. Therefore, there is an urgent need for a technical method that can overcome the data bottleneck and specifically address the challenge of predicting sudden changes in traffic flow driven by traffic events. Summary of the Invention

[0007] To predict traffic flow changes in road networks under traffic incidents, two main challenges exist: firstly, traffic flow data is scarce, sparse, and difficult to obtain; secondly, existing prediction models suffer from poor adaptability and low accuracy for sudden traffic flow events and non-stationary spatiotemporal patterns, failing to provide data support for traffic management. Therefore, this invention provides a method for predicting local road network traffic flow changes under the influence of traffic incidents.

[0008] To achieve the above objectives, the technical solution provided by this invention is as follows: A method for predicting local road network traffic flow changes under the influence of traffic events includes the following steps: S1: Generate a high-quality hybrid dataset: The hybrid dataset is a spatiotemporally dense traffic flow dataset that combines real and simulated traffic flow data with complete traffic event labels; Specifically, this includes: constructing and calibrating a micro-level traffic network simulation model G based on actual road network GIS data. sim In G sim A dense set of virtual sensors is deployed in the middle to form a node set N. dense Simulated traffic flow data D is generated through multi-scenario traffic event simulation. sim_event D sim_event Compared with historical real traffic flow data D real Align and fuse the datasets in the spatiotemporal dimensions to obtain the hybrid dataset D. mix ={(X Nt i t )|t belongs to T hist}, where X Nt Let i be the traffic flow data for all dense nodes at time t. t The event status label corresponding to time t; S2: Constructing a dynamic graph structure for the road network that integrates traffic event features: Abstracting the target road network into a dynamic graph sequence {G} at discrete time steps. t = (V, E, A) t X t )|t=1,2,...,T}, and for each node v j Construct the event feature embedding vector i at time t. j This forms a node feature matrix X containing embedded event features. t =[X t ;I t ], where V is a traffic node, E is a directed connection edge, and A t Let be the weighted adjacency matrix at time t; S3: Construct and train an end-to-end spatiotemporal prediction model based on adaptive dynamic block partitioning: An allocation matrix is ​​generated through a dynamic allocation network, and an ADP module is constructed using an attention mechanism. Nodes are dynamically clustered into several blocks. The spatiotemporal evolution pattern within each block is learned through a shared spatiotemporal core network. Traffic prediction results are generated through feature reconstruction and node-specific predictors, utilizing a hybrid dataset D. mix The training model uses a loss function L that includes the mean squared error between the predicted and actual values, as well as a regularization term for the assignment matrix. S4: Real-time Data Closed-Loop and Online Rolling Prediction: When an actual traffic event is detected, real-time observation data from physical sensors and supplementary data from simulation models are integrated to form complete road network status data. The system embeds real-time event features to construct a dynamic graph sequence, which is then input into the trained model to output a traffic prediction result for the next time period T'. The data and model input parameters are updated at a fixed frequency until the impact of the event ends.

[0009] Preferably, in step S1, the microscopic traffic network simulation model G simThe construction and calibration were performed using microscopic traffic simulation software to build a model, based on actual physical sensor nodes within a historical period. Nreal Multi-dimensional traffic flow time series data D real Based on this, the travel demand matrix, route selection behavior, and driving behavior parameters are adjusted and iteratively calibrated to ensure that the simulation output D in the event-free state is optimized. sim With D real The statistical distribution consistency measure satisfies S(D) real D sim ) th The microscopic traffic simulation software is VISSIM or SUMO, and the multi-dimensional traffic flow data includes traffic volume, speed, and occupancy attributes. th This is a preset threshold.

[0010] Preferably, in step S1, the set of nodes N of the dense virtual sensor dense =N real ∪N virtual , where N real N represents the actual physical sensor location. virtual The deployment locations are for incident-prone areas, key bottleneck sections, and upstream / downstream junctions, and |N dense |Much greater than|N real |

[0011] Preferably, in step S1, the event scenario set I = {i} for multi-scenario traffic event simulation. k |k=1,2,...,M}, each event scenario ik is represented by a tuple (loc... k , t k , t dur_k lane close_k severity k type k The definitions correspond to the geographical location of the event, the simulation start time, the duration, the number of lanes occupied, the severity, and the event type, which include rear-end collisions, spills, construction, and congestion.

[0012] Preferably, in step S2, the event feature embedding vector i j The construction uses a sine function to adjust the severity parameter value of traffic events, and encodes and fuses event attributes and spatiotemporal relationships between nodes into a low-dimensional dense vector. Event attributes include event location, duration, number of lanes occupied, severity, and event type.

[0013] ​Preferably, the construction of the ADP module in step S3 is specifically as follows: synchronously constructing ST-Encoder as a shallow spatiotemporal feature extractor, generating a node-block soft allocation matrix through a dynamic allocation network, and learning the dynamic similarity of nodes by combining an attention mechanism, so that the block allocation strategy changes in real time with traffic conditions and event impacts.

[0014] Preferably, the shared spatiotemporal core network in step S3 is a stacked spatiotemporal convolutional block or a Transformer block. The block-level feature representation is obtained by weighted aggregation of the features of all nodes within the block. After being learned by the spatiotemporal core network, it is weighted and mapped back to the node level through the transpose of the allocation matrix.

[0015] Preferably, the supplementary data for real-time data fusion in step S4 is based on the current road network status and event attributes as initial conditions, and is obtained through a calibrated microscopic traffic network simulation model G. sim Rapidly generated simulations are used to cover blind spots of physical sensors.

[0016] Preferably, the fixed frequency of rolling updates in step S4 is every 5 minutes. Each update replaces the old data with the latest measured data, and the simulation model input parameters are updated synchronously to regenerate supplementary data. The new data with real labels collected online is used for incremental learning and optimization of the model.

[0017] Preferably, the hybrid dataset D mix During the generation process of D sim_event With D real The spatiotemporal alignment and fusion ensures the rationality of data in terms of physical consistency.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs an innovative data generation strategy of "historical data calibration + high-fidelity traffic event simulation" to obtain a large-scale, high-density, accurately labeled, and physically consistent hybrid spatiotemporally synchronized dataset of "traffic events-traffic flow." This effectively solves the data bottleneck problem in road network traffic flow prediction under traffic events, laying a solid data foundation for the training and validation of complex prediction models. Simultaneously, by explicitly introducing traffic event information and processing it into data features understandable to the model, the model can directly perceive the occurrence of traffic events and their differentiated impacts on different nodes of the road network, rather than implicitly inferring from traffic flow data, significantly enhancing the prediction accuracy. The model's interpretability and event perception capabilities are enhanced. Furthermore, the innovative Adaptive Dynamic Blocking (ADP) module dynamically aggregates road network nodes into blocks with similar spatiotemporal evolution patterns based on real-time traffic conditions and the impact of traffic events. This breaks through the limitations of traditional models that rely on fixed graph structures, flexibly capturing the non-stationary traffic impact propagation effects under traffic event disturbances. It significantly improves the modeling ability for sudden, highly volatile traffic flow mutation patterns, achieving accurate and robust prediction of road network traffic changes under the influence of traffic events, and providing reliable data support for proactive traffic management, dynamic route guidance, and emergency response. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the prediction method of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The specific embodiments of the present invention are as follows: A method for predicting local road network traffic flow changes under the influence of traffic events includes the following steps: S1: Generate a high-quality hybrid dataset: The hybrid dataset is a spatiotemporally dense traffic flow dataset that combines real and simulated traffic flow data with complete traffic event labels; Specifically, this includes: constructing and calibrating a micro-level traffic network simulation model Gsim based on actual road network GIS data, and deploying dense virtual sensors in Gsim to form a node set N. dense Simulated traffic flow data D is generated through multi-scenario traffic event simulation. sim_event D sim_event Compared with historical real traffic flow data D real Align and fuse the datasets in the spatiotemporal dimensions to obtain the hybrid dataset D. mix ={(X Nt i t )|t belongs to T hist}, where X Nt Let i be the traffic flow data for all dense nodes at time t. t Let be the event state label at time t, and let D be the mixed dataset. mix During the generation process of D sim_event With D real Spatiotemporal alignment and fusion ensure the reasonableness of data in terms of physical consistency; Step S1 Microscopic traffic network simulation model G sim The construction and calibration were performed using microscopic traffic simulation software to build a model, with the actual physical sensor nodes N within a historical period as the basis. real Multi-dimensional traffic flow time series data D real Based on this, the travel demand matrix, route selection behavior, and driving behavior parameters are adjusted and iteratively calibrated to ensure that the simulation output D in the event-free state is optimized. sim With D real The statistical distribution consistency measure satisfies S(D) real D sim ) th The microscopic traffic simulation software is VISSIM or SUMO, and the multi-dimensional traffic flow data includes traffic volume, speed, and occupancy attributes. th The preset threshold; The set of nodes N of the dense virtual sensor in step S1 dense =N real ∪N virtual , where N real N represents the actual physical sensor location. virtual The deployment locations are for incident-prone areas, key bottleneck sections, and upstream / downstream junctions, and |N dense |Much greater than|N real |

[0023] In step S1, the event scenario set I = {i} for multi-scenario traffic event simulation k |k=1,2,...,M}, each event scenario ik is represented by a tuple (loc... k , t k , t​dur_k lane close_k severity k type k The definitions correspond to the geographical location of the event, the simulation start time, the duration, the number of lanes occupied, the severity, and the event type, which include rear-end collisions, spills, construction, and congestion.

[0024] S2: Constructing a dynamic graph structure for the road network that integrates traffic event features: Abstracting the target road network into a dynamic graph sequence {G} at discrete time steps. t = (V, E, A) t X t )|t=1,2,...,T}, and for each node v j Construct the event feature embedding vector i at time t. j This forms a node feature matrix X containing embedded event features. t =[X t ;I t ], where V is a traffic node, E is a directed connection edge, and A t Let be the weighted adjacency matrix at time t; In step S2, the event feature embedding vector i j The construction uses a sine function to adjust the severity parameter value of traffic events, and encodes and fuses event attributes and spatiotemporal relationships between nodes into a low-dimensional dense vector. Event attributes include event location, duration, number of lanes occupied, severity, and event type. S3: Construct and train an end-to-end spatiotemporal prediction model based on adaptive dynamic block partitioning: An allocation matrix is ​​generated through a dynamic allocation network, and an ADP module is constructed using an attention mechanism. Nodes are dynamically clustered into several blocks. The spatiotemporal evolution pattern within each block is learned through a shared spatiotemporal core network. Traffic prediction results are generated through feature reconstruction and node-specific predictors, utilizing a hybrid dataset D. mix The training model uses a loss function L that includes the mean squared error between the predicted and actual values, as well as a regularization term for the assignment matrix. The construction of the ADP module in step S3 is as follows: ST-Encoder is constructed synchronously as a shallow spatiotemporal feature extractor. A node-block soft allocation matrix is ​​generated through a dynamic allocation network. The dynamic similarity of nodes is learned by combining an attention mechanism, so that the block allocation strategy changes in real time with traffic conditions and event impacts.

[0025] The shared spatiotemporal core network in step S3 consists of stacked spatiotemporal convolutional blocks or Transformer blocks. The block-level feature representation is obtained by weighted aggregation of all node features within the block. After learning by the spatiotemporal core network, it is weighted and mapped back to the node level through the transpose of the allocation matrix. S4: Real-time Data Closed-Loop and Online Rolling Prediction: When an actual traffic event is detected, real-time observation data from physical sensors and supplementary data from simulation models are integrated to form complete road network status data. The system embeds real-time event features to construct a dynamic graph sequence, which is then input into the trained model to output the traffic prediction result for the next time period T'. The data and model input parameters are updated at a fixed frequency until the impact of the event ends.

[0026] The supplementary data for real-time data fusion in step S4, using the current road network state and event attributes as initial conditions, is obtained through the calibrated microscopic traffic road network simulation model G. sim Rapidly generated simulations are used to cover blind spots of physical sensors.

[0027] In step S4, the fixed frequency of rolling updates is every 5 minutes. Each time an update is performed, the old data is replaced with the latest measured data, and the input parameters of the simulation model are updated synchronously to regenerate supplementary data. The new data with real labels collected online is used for incremental learning and optimization of the model.

[0028] Embodiment 1 of the present invention is as follows: Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that variations may be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting local road network traffic flow changes under the influence of traffic incidents, characterized in that, Includes the following steps: S1: Generate a high-quality hybrid dataset: The hybrid dataset is a spatiotemporally dense traffic flow dataset that combines real and simulated traffic flow data with complete traffic event labels; Specifically, this includes: constructing and calibrating a micro-level traffic network simulation model G based on actual road network GIS data. sim In G sim A dense set of virtual sensors is deployed in the middle to form a node set N. dense Simulated traffic flow data D is generated through multi-scenario traffic event simulation. sim_event D sim_event Compared with historical real traffic flow data D real Align and fuse the datasets in the spatiotemporal dimensions to obtain the hybrid dataset D. mix ={(X Nt i t )|t belongs to T hist }, where X Nt Let i be the traffic flow data for all dense nodes at time t. t The event status label corresponding to time t; S2: Constructing a dynamic graph structure for the road network that integrates traffic event features: Abstracting the target road network into a dynamic graph sequence {G} at discrete time steps. t = (V, E, A) t X t )|t=1,2,...,T}, and for each node v j Construct the event feature embedding vector i at time t. j This forms a node feature matrix X containing embedded event features. t =[X t ;I t ], where V is a traffic node, E is a directed connection edge, and A t Let be the weighted adjacency matrix at time t; S3: Construct and train an end-to-end spatiotemporal prediction model based on adaptive dynamic block partitioning: An allocation matrix is ​​generated through a dynamic allocation network, and an ADP module is constructed using an attention mechanism. Nodes are dynamically clustered into several blocks. The spatiotemporal evolution pattern within each block is learned through a shared spatiotemporal core network. Traffic prediction results are generated through feature reconstruction and node-specific predictors, utilizing a hybrid dataset D. mix The training model uses a loss function L that includes the mean squared error between the predicted and actual values, as well as a regularization term for the assignment matrix. S4: Real-time Data Closed-Loop and Online Rolling Prediction: When an actual traffic event is detected, real-time observation data from physical sensors and supplementary data from simulation models are integrated to form complete road network status data. The system embeds real-time event features to construct a dynamic graph sequence, which is then input into the trained model to output the traffic prediction result for the next time period T'. The data and model input parameters are updated at a fixed frequency until the impact of the event ends.

2. The method for predicting local road network traffic flow changes under the influence of traffic incidents according to claim 1, characterized in that: Step S1 Microscopic traffic network simulation model G sim The construction and calibration were performed using microscopic traffic simulation software to build a model, based on actual physical sensor nodes within a historical period. Nreal Multi-dimensional traffic flow time series data D real Based on this, the travel demand matrix, route selection behavior, and driving behavior parameters are adjusted and iteratively calibrated to ensure that the simulation output D in the event-free state is optimized. sim With D real The statistical distribution consistency measure satisfies S(D) real D sim ) th The microscopic traffic simulation software is VISSIM or SUMO, and the multi-dimensional traffic flow data includes traffic volume, speed, and occupancy attributes. th This is a preset threshold.​ 3. The method for predicting local road network traffic flow changes under the influence of traffic incidents according to claim 1, characterized in that: The node set N of the dense virtual sensor in step S1 dense =N real ∪N virtual , where N real N represents the actual physical sensor location. virtual The deployment locations are for incident-prone areas, key bottleneck sections, and upstream / downstream junctions, and |N dense |Much greater than|N real | 4. The method for predicting local road network traffic flow changes under the influence of traffic incidents according to claim 1, characterized in that: In step S1, the event scenario set I={i} for multi-scenario traffic event simulation is... k |k=1,2,...,M}, each event scenario ik is represented by a tuple (loc... k , t k , t dur_k lane close_k severity k type k The definitions correspond to the geographical location of the event, the simulation start time, the duration, the number of lanes occupied, the severity, and the event type, which include rear-end collisions, spills, construction, and congestion.

5. The method for predicting local road network traffic flow changes under the influence of traffic incidents according to claim 1, characterized in that: In step S2, the event feature embedding vector i j The construction uses a sine function to adjust the severity parameter value of traffic events, and encodes and fuses event attributes and spatiotemporal relationships between nodes into a low-dimensional dense vector. Event attributes include event location, duration, number of lanes occupied, severity, and event type.

6. The method for predicting local road network traffic flow changes under the influence of traffic incidents according to claim 1, characterized in that: The construction of the ADP module in step S3 is specifically as follows: ST-Encoder is constructed synchronously as a shallow spatiotemporal feature extractor, a node-block soft allocation matrix is ​​generated through a dynamic allocation network, and the dynamic similarity of nodes is learned by combining an attention mechanism, so that the block allocation strategy changes in real time with traffic conditions and event impacts.

7. The method for predicting local road network traffic flow changes under the influence of traffic events according to claim 1, characterized in that: The shared spatiotemporal core network in step S3 consists of stacked spatiotemporal convolutional blocks or Transformer blocks. The block-level feature representation is obtained by weighted aggregation of the features of all nodes within the block. After being learned by the spatiotemporal core network, it is weighted and mapped back to the node level through the transpose of the allocation matrix.

8. The method for predicting local road network traffic flow changes under the influence of traffic incidents according to claim 1, characterized in that: The supplementary data for real-time data fusion in step S4, using the current road network state and event attributes as initial conditions, is obtained through a calibrated microscopic traffic network simulation model G. sim Rapidly generated simulations are used to cover blind spots of physical sensors.

9. The method for predicting local road network traffic flow changes under the influence of traffic events according to claim 1, characterized in that: In step S4, the fixed frequency of rolling updates is every 5 minutes. Each time an update is performed, the old data is replaced with the latest measured data, and the input parameters of the simulation model are updated synchronously to regenerate supplementary data. The new data with real labels collected online is used for incremental learning and optimization of the model.

10. The method for predicting local road network traffic flow changes under the influence of traffic incidents according to claim 1, characterized in that: The hybrid dataset D mix During the generation process of D sim_event With D real The spatiotemporal alignment and fusion ensures the rationality of data in terms of physical consistency.

Images