Whole-domain dynamic perception space-time traffic flow prediction method based on graph packet representation learning

Abstract

The invention discloses a global dynamic perception space-time traffic flow prediction method based on graph packet representation learning, and belongs to the technical field of traffic flow prediction, and the method comprises the following steps: S1, traffic data input, S2, traffic graph packet construction, S3, graph packet initial feature extraction, S4, time sequence feature extraction, S5, spatial feature extraction, and S6, traffic flow prediction and output. Through a space-time modeling technology of graph packet representation learning and global dynamic perception, space-time characteristic elements of a traffic road network can be comprehensively covered, traditional traffic indexes such as flow and speed are concerned, elements such as road network topological association and cross-regional multi-hop association are also included, a dynamic dependency relationship between a time sequence and a spatial dimension is deeply mined, and a real-time dynamic perception effect is achieved. Therefore, the prediction result can reflect the real evolution law of the traffic flow more accurately, and a more scientific basis is provided for traffic management and decision making.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic flow prediction, more particularly to a global dynamic perception spatio-temporal traffic flow prediction method based on graph package representation learning. BACKGROUND

[0002] With the continuous deepening of urbanization process and the rapid development of intelligent transportation system construction, traffic management is facing the severe challenges brought by the rapid increase of road network scale and the diversification of travel demand. The traditional static management and passive response mode has been difficult to cope with the increasingly serious traffic congestion, safety accidents and environmental pollution problems. Therefore, it is particularly important to conduct dynamic traffic flow prediction with high precision and global coverage.

[0003] The patent document with the authorization announcement number CN118397837A discloses a traffic flow prediction method based on an interactive dynamic diffusion graph convolution network. The interactive dynamic diffusion graph convolution network includes a space-time interactive dynamic graph generator, a stacked space-time block, a jump connection and a multi-layer perceptron. Traffic data is input into the space-time interactive dynamic graph generator to capture the heterogeneous space-time correlation of the traffic data and generate a space-time dynamic graph. The traffic data and the space-time dynamic graph are input into the stacked space-time block to capture the time and space features in the traffic network. Finally, the time and space features captured by each space-time block are aggregated through the jump connection, and then passed through the multi-layer perceptron to obtain the prediction result. The patent document with the authorization announcement number CN118968784A discloses a deep learning traffic flow prediction method based on time-aware hypergraph. First, a hybrid hypergraph is constructed using traffic flow data to describe the spatial structure information of the traffic network, including the position information of the nodes and the high-order relationship between the nodes, and a time-aware hypergraph is constructed to represent the dynamic periodic characteristics of the traffic flow data. Second, combining the spatial structure information and the dynamic periodic characteristics of the traffic network, a time module based on a time encoder and a time attention network is used to extract and mine the time dependence relationship in the traffic flow data, a spatial module based on a spectral domain space hypergraph encoder and a hypergraph attention network is used to deeply mine the spatial dependence relationship in the traffic flow data, and a multi-layer perceptron feature fusioner is used to fuse various time and space features into space-time dependence relationship. Finally, a predictor based on a full connection layer is used to map the space-time dependence relationship into the future traffic flow prediction value.

[0004] While existing technologies can capture the spatiotemporal correlation of traffic data to some extent and pay attention to the periodicity, dynamics, or high-order spatial relationships of nodes in the spread signals, thus improving the accuracy of traffic flow prediction, they do not structurally organize traffic data in continuous time slices. This leads to the loss of temporal correlations and limits the efficiency of temporal feature extraction. Furthermore, the use of simple linear transformations to extract initial features fails to fully integrate graph structure information with temporal correlations, retaining only numerical information, which increases the difficulty of subsequent spatial modeling. They also rely on static adjacency matrices, which cannot adapt to the time-varying characteristics of road networks. Dynamic graph techniques also have local biases and ignore cross-regional multi-hop correlations, leading to the accumulation of long-term prediction errors. In addition, the constructed time-aware structure relies on historical periodic data and is less adaptable to non-periodic sudden traffic changes. They have not formed a collaborative closed loop of "data organization-feature extraction-spatiotemporal modeling," and there is still room for improvement in overall prediction accuracy and scenario adaptability. Summary of the Invention

[0005] This invention mainly provides a global dynamic perception spatiotemporal traffic flow prediction method based on map package representation learning, which can solve the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a global dynamic perception spatiotemporal traffic flow prediction method based on map packet representation learning, comprising: S1. Input raw traffic data for multiple consecutive time slices, including the vertex set of road network monitoring stations, the edge set of road network topology associations, and the flow and speed data of nodes; S2. Aggregate the traffic maps of consecutive time slices into a single traffic map package, which contains multiple traffic maps. Each traffic map has a vertex set, an edge set, and a node feature matrix at the corresponding time, and is associated with the traffic state label of the future time to be predicted. S3. Extract the structural features of each traffic map in the map package, calculate the relevance weight of each traffic map to the prediction task, and then aggregate the features of each traffic map according to the weight to obtain the initial features at the map package level. S4. Perform convolution operations with increasing dilation factors on the initial features, filter effective temporal features through gated activation units, and then output multi-scale temporal dependency features. S5. Based on temporal features, a time-varying adjacency matrix is ​​dynamically generated. Physical propagation association and cross-regional semantic association are captured through dual paths respectively. Then, the outputs of the dual paths are fused to obtain global-local spatial features. S6. Output traffic flow data for a preset time interval in the future, with the data corresponding to the predicted values ​​of each monitoring station on the road network.

[0007] Furthermore, in S1, the node features of the original traffic data include three types of features: traffic flow, speed, and occupancy rate. The original traffic data comes from historical data collected by road network monitoring stations.

[0008] Furthermore, in S2, the number of consecutive time slices is a preset historical time window length, the traffic map package is composed of traffic maps of consecutive time slices, and the tags associated with the traffic map package correspond to the traffic status at a preset time interval in the future.

[0009] Furthermore, in S3, the intra-graph structural feature extraction adopts the trend similarity calculation method, and the correlation weight calculation and feature aggregation adopt the attention multi-graph pooling technology, which includes two steps: weight calculation and feature aggregation.

[0010] Furthermore, in S4, the convolution operation with increasing dilation factor is a dilated causal convolution, and the gated activation unit adopts an operation method combining two activation functions.

[0011] Furthermore, in S5, the time-varying adjacency matrix is ​​generated by normalizing trend similarity using the softmax function, and the dual pathways include a diffusion convolutional pathway and a multi-head attention pathway.

[0012] Furthermore, in S6, the preset time intervals include multiple levels such as short-term, medium-term, and long-term, and the output traffic flow data covers all monitoring stations within the road network.

[0013] Furthermore, the in-graph structural features are a fixed-dimensional node feature matrix, and the weight parameters and bias parameters in attention multi-graph pooling are optimized iteratively through model training.

[0014] The beneficial effects of this invention's spatiotemporal traffic flow prediction method based on map packet representation learning are as follows: By combining map packet representation learning with spatiotemporal modeling techniques based on full-domain dynamic perception, it can comprehensively cover the spatiotemporal characteristics of the traffic network. It not only focuses on traditional traffic indicators such as flow and speed but also incorporates factors such as network topology correlation and cross-regional multi-hop correlation, and deeply mines the dynamic dependencies between temporal and spatial dimensions, making the prediction results more accurately reflect the true evolution of traffic flow and providing a more scientific basis for traffic management and decision-making. Furthermore, through the collaborative technology of the map packet representation learning module with modules such as dilated causal convolution and route perceptron, it realizes the transformation of traffic data from "raw collection" to "precise spatiotemporal feature prediction." The efficient transformation of "extraction" breaks the limitations of fragmented data organization and single feature extraction, enabling deep collaboration among modules to capture multi-scale spatiotemporal correlations, improving the accuracy and efficiency of traffic flow prediction, and enhancing the model's adaptability to complex traffic scenarios. At the same time, through multi-path spatial feature extraction and real-time decision support technology, it provides richer and more accurate visualization methods for the spatiotemporal evolution of traffic flow, clearly showing the spatial distribution and dynamic changes of traffic flow. Furthermore, the decision support mechanism based on real-time prediction data and dynamic correlation analysis can quickly provide scientific and effective control suggestions in scenarios such as traffic congestion and emergencies, greatly improving the decision-making level of traffic management and the ability to cope with complex traffic conditions. Attached Figure Description

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0016] Figure 1 This is a schematic diagram of the method flow for the global dynamic perception spatiotemporal traffic flow prediction method based on map package representation learning according to the present invention. Detailed Implementation

[0017] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 like Figure 1 As shown, a technical solution is provided: a global dynamic perception spatiotemporal traffic flow prediction method based on map package representation learning, including: Step 1: Input Traffic Data Input raw traffic data from multiple consecutive time slices, including the vertex set of road network monitoring stations, the edge set of road network topology associations, and the flow and speed data of nodes; Specifically, the node characteristics of raw traffic data include three types of characteristics: traffic flow, speed, and occupancy. Raw traffic data comes from historical data collected by road network monitoring stations.

[0019] The raw traffic data for each time slice needs to be standardized into a structured traffic map format, as shown below: In the formula, This refers to the vertex set consisting of all monitoring stations (such as sensors) in the corresponding road network. For the edge set between stations based on road network connectivity, The node feature matrix; The raw traffic data needs to be divided into time dimensions according to a preset historical time window, for example, let the window include A series of consecutive time slices form a time-series data set: The data source should prioritize real traffic datasets to ensure the authenticity and generalizability of the data, while also adapting to the requirements of subsequent modeling for the length of the input time series. In addition, the raw traffic data needs to be preprocessed to ensure quality, for example, for... For missing values ​​in the traffic map, linear interpolation is used to fill in the missing values ​​(e.g., if speed data for a station is missing in a certain time slot, interpolation is performed based on the speed values ​​of the adjacent time slots). For outliers (e.g., extreme values ​​that exceed the normal traffic flow range), the 3σ criterion is used to identify and replace them with the median of the same historical period for that station, avoiding interference from invalid data in subsequent feature extraction and ensuring the accuracy of traffic maps for each time slot. All possess complete and reliable feature information; Finally, it is necessary to establish the correlation between the raw traffic data and the predicted labels, that is, for each time series set. It is necessary to bind traffic status labels for future preset times (e.g., traffic flow in 15 minutes, 30 minutes, and 60 minutes). The label values ​​are taken from the real observation data of the corresponding monitoring stations, providing effective data pairs to support the model in learning the mapping relationship between "historical time series data → future traffic status".

[0020] Step 2: Construction of the traffic map package Traffic maps from consecutive time slices are aggregated into a single traffic map package, which contains multiple traffic maps. Each traffic map has a vertex set, an edge set, and a node feature matrix for the corresponding time, and is associated with traffic state labels for the future time to be predicted. Specifically, the number of consecutive time slices is the preset length of the historical time window, the traffic map package is composed of traffic maps of consecutive time slices, and the labels associated with the traffic map package correspond to the traffic status at a preset time interval in the future.

[0021] The label Y associated with the traffic map package must be strongly bound to the future prediction task. The label value must be taken from the real traffic status of the road network monitoring station at the corresponding future time (such as the traffic flow value 15 minutes later), forming a data pair of "map package B-label Y", that is, each B uniquely corresponds to one Y, providing structured training data for the subsequent model to learn the mapping relationship "from map package to prediction label" (B→Y). The label dimension must be consistent with the number of nodes in the map package to ensure that the prediction value of each monitoring station can be accurately supervised. At the same time, it is necessary to ensure the consistency and arrangement invariance of each traffic map in the map package: consistency is reflected in the fact that the vertex set and edge set (corresponding to the road network topology) of all traffic maps remain fixed, and only the node features are dynamically updated over time. This can be naturally achieved through the weight aggregation mechanism of attention multi-graph pooling without additional processing. Finally, the completed traffic map package needs to be validated. The validation includes: first, whether the number of traffic maps in the package is consistent with the preset time window length to avoid data loss; second, whether the dimension of the node feature matrix of each traffic map is consistent to ensure that it includes core features such as traffic flow, speed, and occupancy; and third, whether the labels match the time range of the map package (for example, when the historical time window corresponds to 1 hour of data, the label should be the traffic status 15 minutes after that time period). Only data that passes the validation can be used for subsequent map package representation learning to provide high-quality structured input for model training.

[0022] Step 3: Initial Feature Extraction of Image Package The internal structural features of each traffic map in the map package are extracted, the relevance weights of each traffic map to the prediction task are calculated, and then the features of each traffic map are aggregated according to the weights to obtain the initial features at the map package level. Specifically, the extraction of intra-graph structural features adopts the trend similarity calculation method, and the correlation weight calculation and feature aggregation adopt the attention multi-graph pooling technique, which includes two steps: weight calculation and feature aggregation.

[0023] The process begins with preliminary feature mining within the graph. For each time slice of the traffic map package, feature transformation is performed by combining the road network topology and node features. Graph convolution operations are then used to deeply fuse the topological associations between node features and edges, avoiding the limitation of traditional linear transformations that only retain numerical information. This ensures that the extracted features simultaneously include both node-edge structural features and traffic state numerical features, as detailed below: In the formula, and They are nodes ,node In the historical time window Temporal feature vectors within, The weight matrix is ​​a learnable matrix; Simultaneously, the attention weights of each traffic map feature within the map package are calculated to quantify the contribution of features from different time slots to the prediction task. The weight calculation is achieved through linear transformation combined with activation functions, and the output weight values ​​can accurately capture the feature importance of key time slots (such as the pre-peak period) and weaken the interference of noisy time slots. In addition, based on attention weights, all in-graph features are aggregated, and the "structure-numerical" fusion features of multiple time slices are aggregated into initial features of a unified dimension by weighted summation. This not only preserves key time series information, but also achieves the normalization of feature dimensions, and solves the problem of loss of association caused by the fragmentation of continuous time slice data. Finally, the initial features obtained from aggregation are validated for effectiveness. The validation includes: whether the feature dimensions match the input requirements of the subsequent modeling module, whether the feature value distribution meets the preset range (normalization is used to ensure that the values ​​are within a reasonable range), and whether the features of different image packages are distinguishable (redundant features are eliminated by calculating the similarity between features). Only features that pass the validation can be input into the subsequent time series feature extraction stage, laying the foundation for improving the overall prediction accuracy of the model.

[0024] Step 4: Temporal Feature Extraction The initial features are subjected to convolution operations with increasing dilation factors. Effective temporal features are then filtered through gating activation units to output multi-scale temporal dependency features. Specifically, the convolution operation with increasing dilation factor is called dilated causal convolution, and the gated activation unit adopts the operation method of combining two activation functions.

[0025] The initial input features are information carriers that have been structured and aggregated from traffic map packages. These features have integrated traffic maps from multiple consecutive time slices into a unified unit, implying the flow propagation trend and node topological relationships between time slices. This changes the input format of traditional fragmented time data and provides a structured foundation for time series modeling, as detailed below: In the formula, As the expansion factor, The kernel length is 1. This is the result of the convolution; Dilated causal convolution expands the receptive field through progressively increasing dilation factors, enabling precise coverage of the time span from recent to long-term within the graph package. It also prioritizes capturing the core trends carried by key time slices aggregated by weights, avoiding breaks in temporal correlations caused by fragmentation of the original data. The formula is as follows: In the formula, and The calculation results for different convolution kernels correspond to the "memory gate" and the "forget gate," respectively. For element-wise multiplication, It is the sigmoid function; Meanwhile, the gating activation mechanism dynamically filters the features generated by convolution, retaining effective information that is strongly correlated with future traffic status and eliminating weak noise introduced during the image packet aggregation process, thereby further improving the relevance and purity of the time-series features. The final generated temporal features will fully preserve the structured temporal patterns and multi-scale dependency information of the traffic map package, providing accurate temporal references for the construction of dynamic topology in the subsequent spatial feature extraction stage, realizing the synergistic linkage of temporal and spatial information, and fundamentally optimizing the dynamic perception effect of the entire domain.

[0026] Step 5: Spatial Feature Extraction A time-varying adjacency matrix is ​​dynamically generated based on temporal features. Physical propagation associations and cross-regional semantic associations are captured through dual paths, and then the outputs of the two paths are fused to obtain global-local spatial features. Specifically, the time-varying adjacency matrix is ​​generated by normalizing trend similarity using the softmax function, and the dual-pathway includes a diffusion convolutional path and a multi-head attention path.

[0027] The progressive graph construction process relies on the structured information carried by the initial features aggregated by the traffic map package. This feature integrates the node trend similarity and road network topology association of multiple consecutive time slices, which can provide a cross-time dimension reference for the generation of dynamic adjacency matrix, avoid the misjudgment of topological relationship caused by relying only on data at a single moment, and make the constructed graph structure more in line with the real dynamic law of traffic flow evolution over time. Furthermore, in the dual-path feature fusion mechanism, the diffusing convolutional path can accurately model the actual propagation path of traffic flow in the road network by leveraging the physical road network association information contained in the map bag, while the multi-head attention path can adaptively identify cross-regional multi-hop semantic associations based on the weight distribution of key time slices within the map bag, effectively compensating for the deficiency of insufficient global association capture caused by local aggregation in traditional methods. The weights of the multi-head attention path are as follows: In the formula, , , For learnable feature vectors, For feature dimension, For the number of attention heads, This is the attention weight matrix. For attention output features; The final generated spatial features will fully integrate the structured topological information and dynamic temporal patterns of the map package, forming a deep synergy with the aforementioned temporal features. This enables the model to synchronously perceive the temporal evolution trend and spatial propagation characteristics of traffic flow, fundamentally improving the accuracy of dynamic perception across the entire domain, especially in scenarios with drastic changes in spatial relationships, such as during peak hours.

[0028] Step 6: Traffic Flow Forecasting and Output Output traffic flow data for a preset time interval in the future, with the data corresponding to the predicted values ​​of each monitoring station on the road network; Specifically, the preset time intervals will include short-term, medium-term, and long-term levels, and the output traffic flow data will cover all monitoring stations within the road network.

[0029] When performing feature fusion, the integrated temporal and spatial features deeply carry the structured information of the traffic map package. This information integrates the traffic evolution trend, node trend similarity and dynamic topological association of multiple consecutive time slices, providing a more comprehensive global spatiotemporal reference benchmark for feature mapping. It effectively avoids the mapping distortion problem caused by the deviation of traditional single-moment data, allowing the prediction model to more accurately transform the fused features into the traffic state at the time to be predicted. Meanwhile, the output prediction results can accurately correspond to the core patterns of key time slices captured by the traffic map package. Relying on the temporal evolution logic and spatial propagation characteristics contained in the map package, it can significantly suppress the accumulation of errors in long-term prediction scenarios. Especially during periods of drastic changes in traffic flow dynamics, such as morning and evening rush hours, the prediction values ​​can better match the real state of the traffic flow of each node in the road network as it changes with time and space, fully demonstrating the prediction advantages of full-domain dynamic perception.

[0030] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A global dynamic perception spatio-temporal traffic flow prediction method based on graph package representation learning, characterized in that: S1, input continuous multiple time slice original traffic data, including vertex set of road network monitoring station, edge set of road network topological correlation, node traffic, speed data; S2, aggregate the traffic graphs of continuous time slices into a single traffic graph package, which includes multiple traffic graphs, each traffic graph having a vertex set, an edge set, and a node feature matrix corresponding to the time, and is associated with a traffic state label of a future time to be predicted; S3, extract intra-graph structure features for each traffic graph in the graph package, calculate the relevance weight of each traffic graph to the prediction task, and then aggregate the traffic graph features according to the weight to obtain the initial features of the graph package level; S4, perform convolution operation with increasing dilation factor on the initial features, filter effective time sequence features through the gating activation unit, and then output multi-scale time sequence dependent features; S5, dynamically generate a time-varying adjacency matrix based on the time sequence features, capture physical propagation correlation and cross-region semantic correlation through a double channel respectively, and then fuse the global-local spatial features output by the double channel; S6, output traffic flow data of a future preset time interval, which corresponds to the prediction value of each monitoring station in the road network.

2. The global dynamic perception spatio-temporal traffic flow prediction method based on graph-based representation learning according to claim 1, characterized in that: In S1, the node features of the original traffic data include traffic flow, speed, and occupancy rate, and the original traffic data is derived from historical data collected by road network monitoring stations.

3. The global dynamic perception spatio-temporal traffic flow prediction method based on graph-based representation learning according to claim 1, characterized in that: In S2, the number of continuous time slices is a preset historical time window length, and the traffic graph package is composed of traffic graphs of continuous time slices, and the label associated with the traffic graph package corresponds to the traffic state of a future preset time interval.

4. The global dynamic perception spatio-temporal traffic flow prediction method based on graph-based representation learning according to claim 1, characterized in that: In S3, the intra-graph structure feature extraction adopts a trend similarity calculation method, the relevance weight calculation and feature aggregation adopt an attention multi-graph pooling technology, including two steps of weight calculation and feature aggregation.

5. The global dynamic perception spatio-temporal traffic flow prediction method based on graph-based representation learning according to claim 4, characterized in that: In S4, the convolution operation with increasing dilation factor is a dilated causal convolution, and the gating activation unit adopts a combination operation mode of double activation functions.

6. The global dynamic perception spatio-temporal traffic flow prediction method based on graph-based representation learning according to claim 1, characterized in that: In S5, the time-varying adjacency matrix is generated by normalizing the trend similarity through a softmax function, and the double channel includes a diffusion convolution channel and a multi-head attention channel.

7. The global dynamic perception spatio-temporal traffic flow prediction method based on graph-based representation learning according to claim 1, characterized in that: In S6, the future preset time interval includes multiple grades such as short-term, medium-term, and long-term, and the output traffic flow data covers all monitoring stations in the road network.

8. The global dynamic perception spatio-temporal traffic flow prediction method based on graph-based representation learning according to claim 4, characterized in that: The intra-graph structure features are fixed-dimensional node feature matrices, and the weight parameters and bias parameters in the attention multi-graph pooling are optimized through model training iteration.

Images