Gated multi-graph convolution perception modeling method for traffic flow prediction

Abstract

The invention relates to a gated multi-graph convolution perception modeling method for traffic flow prediction. The method integrates multi-graph structure construction, gating graph convolution and time feature extraction, and aims to solve the problems of strong time fluctuation and heterogeneous spatial relationship in traffic data. The method comprises the following steps of: firstly, respectively constructing a geographic map and a semantic map according to the maximum mutual information measurement between the spatial distribution information of a sensor and historical traffic data; and then, designing a dual-adaptive gating graph convolution module, and dynamically adjusting an information propagation path of a multi-graph structure by introducing an attention mechanism and a gating factor, thereby improving the modeling performance of the model on spatial isomerism dependence. On the time dimension, a time sequence interactive sensing module is constructed in combination with multi-scale causal convolution and an attention mechanism, time dependence characteristics of a short period and a long period are captured, and fusion and expression of time characteristics are completed. According to the method, the modeling precision and stability of the traffic prediction model in a complex traffic scene can be effectively enhanced, and the method has relatively high practical application value.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and intelligent transportation technology, and specifically to a gated multi-graph convolutional perception modeling method for traffic flow prediction, which belongs to the category of spatiotemporal data modeling and traffic flow prediction technology based on graph neural networks. Background Art

[0002] With the accelerating pace of global urbanization, urban population density and transportation demand continue to grow, leading to increasingly serious problems such as traffic congestion, environmental pollution, and inefficient commuting. Against this backdrop, Intelligent Transportation Systems (ITS) have rapidly developed as a key technological tool to alleviate these issues. Traffic flow forecasting, a core ITS technology, analyzes historical traffic data and real-time traffic conditions to predict future traffic trends, providing critical support for autonomous driving decision-making and traffic management and scheduling.

[0003] Modeling complex spatiotemporal dependencies is a key challenge in traffic flow forecasting. On the one hand, traffic flow data, as a multivariate time series, exhibits significant short-term volatility and long-term periodicity. On the other hand, traffic networks are spatially dynamic and heterogeneous. Adjacent nodes may exhibit distinct traffic patterns due to differences in road structure or function, and spatial relationships can evolve dynamically over time, events, and other external factors. Therefore, accurately modeling the coexistence of static and dynamic spatiotemporal structures is crucial to improving forecasting accuracy.

[0004] To address these issues, research has progressed from statistical models to deep learning models. Traditional methods such as historical averages, autoregressive differential moving averages, and support vector machines have some ability to model linear or partially nonlinear patterns, but are limited in handling dynamic and complex relationships. In recent years, recurrent neural networks, long-short-term memory networks, convolutional neural networks, and the Transformer architecture have been widely used in time series modeling, enhancing the models' ability to express volatility and long-term dependencies. However, these methods primarily focus on temporal modeling, and their spatial modeling capabilities are relatively limited.

[0005] Graph Neural Networks (GNNs) have become an important tool for spatial dependency modeling due to their ability to process non-Euclidean data structures. Researchers have combined graph convolutional networks (GCNs) with temporal networks to form spatiotemporal graph neural networks (STGNNs), improving predictive performance. However, these methods often rely on static adjacency matrices, making it difficult to capture the complex dependencies between distant, indirect nodes. Furthermore, they struggle to adapt to the rapid evolution of network structures in emergencies or dynamic environments.

[0006] To alleviate these issues, some research has introduced dynamic similarity metrics (such as dynamic time warping and the Pearson correlation coefficient) to construct new adjacency matrices, or employed trainable adaptive adjacency structures to dynamically adjust node dependencies. Other research has attempted to model time-varying relationships between nodes using dynamic graph convolutional structures. While these improvements have improved model performance, existing methods still suffer from the following shortcomings:

[0007] (1) Relying on static graph structures or fixed similarity metrics, it is difficult to comprehensively model fine-grained dynamic dependencies in heterogeneous multi-source data;

[0008] (2) Temporal and spatial dependency modeling are often performed independently, making efficient fusion impossible;

[0009] (3) There is a lack of adaptive aggregation mechanism under multi-graph structure, and the model is still not stable and adaptable enough when dealing with sudden changes in traffic.

[0010] Therefore, it is urgent to propose a unified framework that supports adaptive modeling of multi-graph structures, has the ability to recognize fine-grained spatial relationships, and integrates multi-scale temporal modeling mechanisms to improve the accuracy, stability, and generalization ability of traffic flow prediction. Summary of the Invention

[0011] The present invention aims to address the problems of insufficient accuracy, rigid structure, and lack of fine-grained control in existing traffic flow prediction methods in modeling complex spatial correlations and dynamic spatiotemporal dependencies. It provides a gated multi-graph convolutional perception modeling method for traffic flow prediction. This method can fully utilize geographic location and semantic relationships to construct a multi-graph structure, introduce a dynamic attention mechanism and a gating adjustment strategy, and enhance the model's perception of multi-scale heterogeneous spatiotemporal features, thereby significantly improving the accuracy and generalization ability of traffic flow prediction.

[0012] To achieve the above objectives, the present invention provides a gated multi-graph convolutional perceptual modeling method for traffic flow prediction, the technical solution of which includes the following steps:

[0013] 1. Data embedding layer: Embed the raw traffic flow data, extract high-order features through a fully connected network, and fuse periodic time embedding with trainable adaptive embedding representation to form a unified spatiotemporal embedding representation, providing input support for subsequent modeling.

[0014] 2. Spatiotemporal convolution layer: (1) Constructs a geographic map and a semantic map to reflect the spatial adjacency and semantic similarity between traffic nodes, respectively; (2) Proposes a temporal attention interactive convolution module (TAIC) to fuse the attention mechanism with multi-scale convolution in local time blocks to capture short-term fluctuations and long-term trends; (3) Designs a dual adaptive gated graph convolution module (DAGGC) to introduce a multi-head attention mechanism to spatially aggregate multiple graph structures, and controls the information propagation intensity through a gating mechanism to adaptively learn dynamic spatial dependency structures.

[0015] 3. Spatiotemporal fusion layer: The outputs of the multi-scale spatiotemporal convolution modules are concatenated and fused, and the final spatiotemporal representation is generated through additional convolutional layers and nonlinear mapping.

[0016] 4. Prediction output layer: The fused representation is mapped into traffic flow prediction values ​​for multiple future time steps through a fully connected network and gated linear units (GLUs). A non-autoregressive approach is used to improve stability and computational efficiency.

[0017] 5. Loss function design: Huber Loss is used as the regression loss to enhance the model's robustness to outliers and improve stability in actual deployment.

[0018] Compared with the prior art, the present invention has the following technical advantages and innovations:

[0019] (1) For the first time, a multi-graph structure modeling strategy that integrates geographic graphs, semantic graphs, and dynamic graph attention matrices is introduced to more comprehensively characterize spatial heterogeneity;

[0020] (2) A graph convolution method combining multi-head attention mechanism and gated fusion mechanism is proposed to achieve dynamic control of multi-relationship graph information between nodes;

[0021] (3) Design a temporal attention interactive convolutional structure to effectively learn sequence dependencies at different time granularities;

[0022] In summary, by introducing adaptive mapping, gating mechanism and multi-graph fusion strategy, the present invention improves the prediction accuracy while enhancing the model's ability to model dynamic spatiotemporal dependencies, which has good practical application value and engineering deployment prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments or description of the prior art.

[0024] Figure 1The present invention provides a gated multi-graph convolutional perception modeling method for traffic flow prediction;

[0025] Figure 2 The data embedding layer of the present invention;

[0026] Figure 3 The temporal attention interactive convolution module described in the present invention;

[0027] Figure 4 The dual adaptive gated graph convolution module of the present invention;

[0028] Figure 5 The multi-step prediction results of the method described in the present invention are compared with the true values ​​on the METR-LR dataset;

[0029] Figure 6 The multi-step prediction results of the method described in the present invention are compared with the true values ​​on the PEMS-BAY dataset. DETAILED DESCRIPTION

[0030] To further clarify the objectives, technical solutions, and advantages of the present invention, the following describes specific embodiments of the present invention with reference to the accompanying drawings and examples. It should be understood that these embodiments are intended solely to illustrate the present invention and do not limit its scope. To more clearly describe the present invention, the basic concepts involved in traffic network modeling are first defined.

[0031] In the present invention, the traffic network can be represented as a directed weighted graph G = (V, E, A), where V = {v1, v2, ..., v N} represents the set of sensor nodes, E={e1,e2,...,e M} represents an edge set, which indicates the connection relationship between nodes. is the adjacency matrix, which is used to represent the connection weights between nodes. Traffic observation data is represented as traffic signal sequence Where N represents the number of nodes, and F represents the feature dimension of each node, such as traffic flow, vehicle speed, etc. Given the traffic sequence [X t-P+1 ,...,X t ], predict the traffic status of the next Q time steps [X t +1 ,...,X t+Q ]The objective function can be abstracted as:

[0032]

[0033] To effectively model the multi-level relationships between nodes in a transportation network, the present invention constructs two types of graph structures: geographic graph and semantic graph, to model the correlations in physical space and semantic space respectively.

[0034] First, construct the geographic graph G l =(V,E,A dis ) is used to simulate the physical connection relationship between nodes. The adjacency matrix of the geographic graph It is calculated using the Gaussian kernel function, and the weights are defined as follows:

[0035]

[0036] Among them, dist(v i ,v j ) indicates sensor v i With v j σ is the standard deviation of the Gaussian kernel, which is used to adjust the sensitivity of distance to edge weight.

[0037] Secondly, to capture the potential semantic correlation between distant nodes, a semantic graph G is constructed. s =(V,E,A mic ), where the adjacency matrix A mic The nonlinear similarity between node time series is calculated based on the Maximum Mutual Information Coefficient (MIC). The specific calculation method is as follows:

[0038]

[0039] Among them, X i ,X j represents the feature vector of nodes i and j in the historical time series, MIC(,·,) represents the maximum mutual information coefficient, and κ∈(0,1) is the set discrimination threshold. The closer the MIC value is to 1, the stronger the nonlinear correlation between nodes. This method can effectively model complex latent semantic structures in heterogeneous spaces, providing richer adjacency information for subsequent graph convolution.

[0040] In summary, the geographic graph and the semantic graph model the spatial adjacency relationship and dynamic semantic association relationship of nodes in the transportation network from two perspectives, respectively, and will serve as the input basis of the dual adaptive gated graph convolution module of the present invention.

[0041] 1. Data Embedding Layer

[0042] In the implementation process of the present invention, the original traffic data is first embedded to project the traffic observation sequence into a high-dimensional latent space, thereby effectively capturing the periodic characteristics and dynamic spatial differences in the data and providing structural input support for subsequent spatiotemporal modeling.

[0043] Specifically, suppose the traffic signal input sequence is Where T is the time step length, N is the number of traffic sensor nodes, and F is the feature dimension of each node. First, a fully connected transformation is used to map the node features at each moment to the specified hidden dimension D f , and get the original embedding representation:

[0044] E f =FC(x t-T+1 ,x t-T ,...,x t )

[0045] in, Basic feature embeddings representing traffic flows.

[0046] Considering the obvious periodicity in traffic flow data (such as daily fluctuations and weekly patterns), this paper introduces a periodic embedding module to enhance the model's ability to model temporal structures. This module includes two learnable embedding dictionaries: weekly embedding dictionary Japanese embedding dictionary They are used to represent the periodic information of 7 days a week (N=7) and 288 moments a day (the sampling interval is 5 minutes, corresponding to N=288). For each time step, the corresponding week embedding is extracted by indexing operation. and timestamp embedding Concatenate and expand them to the node dimension through broadcasting to obtain the final cycle embedding In order to further enhance the model's ability to model dynamic spatiotemporal differences, spatiotemporal adaptive embedding is introduced This embedding is randomly initialized and then used as a learnable parameter during training, automatically capturing the unique behavioral characteristics of each node at different times through an end-to-end training mechanism. This mechanism enables adaptive learning, where the embedding vector of each node is iteratively optimized during training based on its role and performance in the data, effectively capturing its unique spatiotemporal dynamics.

[0047] In summary, the data embedding layer constructs a unified multi-level spatiotemporal representation X by integrating feature embedding, period embedding, and spatiotemporal adaptive embedding. emb , which provides a concrete basis for subsequent traffic flow prediction. The final expression is as follows:

[0048] Z emb =E f ||E P ||E a

[0049] The hidden dimension D h Equal to 3D f +D a ,|| represents the concatenation operator. For convenience, in the following text, Serves as the input basis for subsequent spatiotemporal graph modeling modules.

[0050] 2. Adaptive Gated Spatiotemporal Convolutional Layer

[0051] In the spatiotemporal modeling module of the present invention, a spatiotemporal convolutional layer is introduced to extract multi-scale temporal dependencies and spatial associations between nodes in traffic flow data. This spatiotemporal convolutional layer is composed of multiple spatiotemporal blocks, each of which contains a TAIC module and a DAGGC module, respectively, for modeling information interactions in the temporal and spatial dimensions.

[0052] To mitigate information leakage during feature stacking, a layer normalization mechanism and a residual connection strategy are introduced in the spatiotemporal block to ensure effective information transfer and fusion between different layers. This design effectively models both short-term and long-term temporal dependencies and, combined with the graph structure, enables the modeling of dynamic spatial relationships between nodes.

[0053] 2.1 Temporal Attention Interactive Convolutional Module

[0054] Traditional temporal convolutional networks extract time series features based on fixed-size convolution kernels, making it difficult to model dynamic relationships between different temporal locations. To enhance the model's adaptability to temporal changes, this paper proposes a self-attention temporal interactive convolution mechanism, which leverages a multi-head attention mechanism to dynamically weight the relative importance of each time slice in the input sequence.

[0055] Assume the input sequence is First, the input is mapped into query matrix, key matrix and value matrix respectively through linear transformation:

[0056] Q=ZW Q ,K=ZW K ,V=ZW V

[0057] Among them, W Q 、W K 、W V is a trainable weight parameter. The attention weight is calculated by the scaled dot product attention mechanism:

[0058]

[0059] in, is the scaled similarity matrix, d k is the scaling factor. The Softmax function is used to normalize the attention weights.

[0060] To capture both short-term and long-term dependencies, the self-attentional temporal interaction convolution module uses two parallel channels to perform convolutions at different scales: one channel uses a small convolution kernel (e.g., 1×3) to extract local temporal dependencies, while the other uses a large convolution kernel (e.g., 1×7) to extract long-term temporal information. These two channels have different receptive fields in the temporal dimension, allowing them to model local mutations and cyclical trends, respectively.

[0061] H1=Conv2D(Z att ,g 1×3 )

[0062] H2=Conv2D(Z att ,g 1×7 )

[0063] Where Conv2D represents a two-dimensional convolution operation. are the feature maps generated by the small-scale convolution kernel and the large-scale convolution kernel, respectively. Then, the nonlinear activation functions tanh and sigmoid are used to capture the deep interaction between multi-scale features, and the interaction of cross-scale features is achieved through element-level multiplication ⊙; then, the interactive features are connected along the channel dimension to generate a fused representation. In order to gradually fuse features and achieve dimensionality reduction, a 2D convolution kernel g is used. 1×1 The fused features are processed to obtain the final output of the module, as shown below:

[0064]

[0065] Z t =Conv2D(Z t (1) ,Z t (2) ,g 1×1 )

[0066] Through the combination of the above-mentioned self-attention mechanism and multi-scale convolution path, the model can dynamically learn the complex dependency structure between different time steps in the time series, and improve the modeling accuracy of non-stationary traffic time series data.

[0067] 2.2 Dual Adaptive Gated Graph Convolution Module

[0068] To model the complex spatial correlations between different types of nodes in traffic networks, this paper constructs an adaptive gated multi-graph spatial convolution module, which is used to capture multi-scale spatial dependencies under heterogeneous graph structures. It also introduces a gating mechanism during the information propagation process to dynamically adjust the information fusion ratio between adjacent nodes, thereby improving the spatial generalization ability of the model when processing dynamic traffic data.

[0069] In the specific implementation, the present invention first constructs two types of spatial graph structures: geographic graph Gl =(V,E,A dis ) and semantic graph G s =(V,E,A mic The geographic graph is constructed based on the geographic locations of sensor nodes, using a Gaussian kernel function to calculate adjacency weights. The semantic graph, on the other hand, uses the Maximum Mutual Information Coefficient (MIC) to characterize the semantic similarity between nodes based on historical traffic flow data. This method can reflect long-term potential traffic dependencies from a nonlinear perspective.

[0070] In order to further enhance the model's ability to identify spatial dependencies in different graph structures, this paper introduces a multi-head graph attention mechanism to perform graph convolution operations. Suppose the target node v i and its neighbor node v j The input feature is z i and z j The attention score is calculated after mapping through a shared linear transformation matrix W:

[0071] z i =Wh i ,z j =Wh j ,

[0072] The role of this linear transformation is to map the original features into a new feature space for subsequent attention calculation. Then, the projected features of the target node and its neighbor nodes are concatenated, and the non-normalized attention score is calculated through the learnable attention vector a, usually using LeakyReLU as the activation function; i For the target node v j the relative importance of

[0073] e ij =LeakyReLU(aT[z′ i ||z′ j ])

[0074] Then, the scores of all neighbor nodes are normalized by the Softmax function to obtain the final attention coefficient α ij :

[0075]

[0076] In the multi-head attention mechanism, each attention head has an independent parameter vector a k , which can capture different spatial patterns respectively. The outputs of multiple attention heads are concatenated in the feature dimension to form a fused expression:

[0077]

[0078] Where || represents the connection operation; W (k) is the linear transformation matrix of the k-th head.

[0079] To support dynamic graph modeling, the present invention further adopts a gating mechanism to perform weighted adjustment on the attention output information flow. t As a basis, let it perform element-by-element multiplication operation with the attention coefficient matrix M (A t ⊙M), and obtain the dynamic attention graph structure under the fusion adjacency relationship.

[0080] On this basis, the node embedding matrix E is introduced s =tanh(αW1),E t =tanh(αW2), respectively as the representation of the source node and the target node, and then construct the metric between the node pairs by embedding similarity:

[0081]

[0082] The above formula uses ReLU to sparse the matrix and uses softmax to normalize the matrix. Like other adaptive graph convolution models, we can directly obtain the adjacency matrix or Laplacian matrix. The output of the adaptive graph convolution can be expressed as:

[0083]

[0084] Then a gating mechanism is used to control how much new layer information each node fuses and how much of the previous layer feature representation is retained:

[0085]

[0086] This gating mechanism can dynamically determine the proportion of spatial information propagation between different nodes, thereby effectively solving the over-smoothing problem caused by the fixed adjacency structure in traditional graph convolution, and improving the model's performance in capturing non-stationary traffic scenarios such as long-distance dependencies and abnormal event disturbances.

[0087] The final output of the dual adaptive gated graph convolution is expressed as:

[0088]

[0089] Where g(·) is a nonlinear activation function, such as ReLU and tanh. Finally, we perform weighted fusion on the output of the distance weight map and the MIC map processed by the multi-layer graph neural network to obtain the final output Z of the AGMGCN module. s :

[0090]

[0091] Where β and λ are the weight coefficients of the two network outputs, respectively, used to balance the impact of geographic distance and nonlinear correlation on the final spatial features. These two weights can be given by training or design, reflecting the importance of different features.

[0092] 3. Space-time fusion layer

[0093] In this implementation, the spatiotemporal fusion layer is used to integrate the outputs of the aforementioned multiple spatiotemporal convolutional modules to obtain the final spatiotemporal representation. This fusion process first stacks and fuses the output features of multiple spatiotemporal convolutional blocks to produce a fused feature tensor, Z. Subsequently, a temporal convolutional module is added at the end of the model to further model the local temporal dependencies of the target time period. Finally, two fully connected layers map and compress the fused features to generate the target prediction result.

[0094] Specifically, the output layer receives the fused hidden representation Z of the last layer and maps it to the prediction space layer by layer through activation functions and linear transformations to complete the final traffic flow prediction. The calculation process is as follows:

[0095] Y=FC(Relu(FC(Z)))

[0096] in, It represents the traffic flow prediction value output by the model, N is the number of sensor nodes, and T is the prediction time step.

[0097] 4. Loss Function

[0098] During model training, to improve prediction robustness and reduce the impact of outliers on training results, this paper uses the Huber loss function as the objective optimization function. The Huber loss function combines the advantages of the mean squared error (MSE) and the mean absolute error (MAE), using a quadratic loss for small error segments and a linear penalty for large error segments. This improves robustness to outliers while maintaining strong fitting capabilities.

[0099] The Huber loss function is defined as follows:

[0100]

[0101] Among them, Y represents the actual traffic flow, represents the traffic flow predicted by the model, and δ is a threshold constant used to regulate the segmentation boundary of the error.

[0102] By introducing this loss function, the present invention can reduce the training instability caused by abnormal fluctuations in traffic data while ensuring prediction accuracy, thereby improving the convergence efficiency and generalization ability of the model.

[0103] Based on the above-mentioned inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the computer program contains instructions for executing the traffic flow prediction method described in the present invention. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a programmable logic device (such as an FPGA), or other computing components suitable for implementing the functions of the method of the present invention. When the processor runs the program instructions, it implements the various steps of the present invention, including input embedding, spatiotemporal modeling, graph convolution operations, spatiotemporal fusion, and prediction output.

[0104] Based on the same inventive concept, the present invention further provides a computer-readable storage medium storing a computer program. When executed by a processor in a computer device, the computer program is used to implement the traffic flow prediction method described in the present invention. The computer-readable storage medium can be a non-transitory computer-readable medium such as a magnetic disk, optical disk, USB flash drive, hard disk, flash memory, ROM, RAM, CD-ROM, DVD, EPROM, EEPROM, or any combination of these media, and all fall within the scope of protection of the present invention.

[0105] In this specification, terms such as "one embodiment," "example," and "one approach" are used solely to illustrate the technical features of a specific implementation and do not limit the scope of protection of the present invention. The technical features mentioned in this specification may be combined or substituted for other embodiments without departing from the spirit of the present invention.

[0106] This paper discloses a gated multi-graph convolutional perceptual modeling method for traffic flow prediction. Its basic principles are: by fusing physical distance graphs with semantic relevance graphs to construct a dynamic spatiotemporal graph structure, combining a gating mechanism with multi-graph convolution to enhance spatial structure expression and communication capabilities; and combining a multi-scale temporal modeling strategy with an attention mechanism to improve the accuracy and robustness of temporal feature extraction. In specific implementations, the modules of this invention can be deployed independently or integrated into an integrated model, demonstrating excellent flexibility, scalability, and practicality.

[0107] Finally, it should be noted that the embodiments described in this specification are merely preferred embodiments of the present invention, intended to clearly illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Any equivalent substitution, structural modification, parameter fine-tuning, model combination, and other technical means performed by those skilled in the art without departing from the original concept of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A gated multi-graph convolutional perception modeling method for traffic flow prediction, characterized by: The steps include: Step 1: Data embedding: preprocess the raw traffic flow data and use a fully connected layer to map it into a high-order representation of uniform dimension. The periodic time embedding and node adaptive embedding features are then integrated to construct a unified input feature. Step 2: Construct a multi-graph structure. Build a geographic graph based on the physical distance between sensor nodes and a semantic graph based on the maximum mutual information coefficient between historical traffic flows. Generate static and dynamic adjacency matrices respectively. Step 3: Temporal attention interaction modeling divides the historical input sequence into multiple time slices, uses multi-scale convolution kernels and temporal attention mechanisms to extract short-term and long-term dependency features, and fuses them through a gating mechanism to generate a temporal representation; Step 4: Dual adaptive gated graph convolution modeling, performing multi-head graph attention convolution operations on the multi-graph structure constructed in step 2, controlling the information propagation and fusion of different graph structures through the gating mechanism, and generating spatial modeling features; Step 5: Spatiotemporal fusion: concatenate the temporal features and spatial features obtained in steps 3 and 4, and input them into a convolutional network or a gated linear unit to obtain a fused representation. Step 6: Output prediction. Input the fusion representation into the fully connected network and output the traffic flow prediction results for multiple time steps in the future. The prediction process adopts a non-autoregressive method.

2. The method according to claim 1, wherein: The adjacency weight of the geographic graph is obtained by calculating the physical distance between nodes, and the adjacency matrix is ​​constructed using a Gaussian kernel function.

3. The method according to claim 1, wherein: The adjacency weight of the semantic graph is calculated based on the maximum mutual information coefficient between historical traffic flow features, and a threshold mechanism is combined to eliminate weakly associated edges.

4. The method according to claim 1, wherein: The temporal attention interaction module consists of two parallel channels, which use convolution kernels of different scales to extract short-term and long-term dependency features respectively, and fuse them through a gating mechanism.

5. The method according to claim 1, wherein: The multi-graph convolution adopts a multi-head attention mechanism, performs attention calculation on each graph structure independently, and splices and fuses the multi-graph convolution results.

6. The method according to claim 1, wherein: The gating mechanism calculates the inter-node propagation weight based on the sigmoid activation function, which is used to dynamically adjust the spatial feature aggregation strength and enhance the adaptability to heterogeneous graph structures.

7. The method according to claim 1, wherein: The output layer includes a gated linear unit (GLU) and two fully connected layers in series to generate the final prediction result.

8. The method according to claim 1, wherein: The training loss function is Huber loss, which is used to enhance the robustness of the model to abnormal traffic points.

9. The method according to claim 1, wherein: The input is a fixed-length historical traffic signal sequence, and the output is a traffic flow prediction value for multiple consecutive time steps in the future.

10. The method according to claim 1, wherein: The method is applicable to multi-step traffic flow prediction tasks in urban main roads, highways or regional traffic sensing networks.

11. A computing device, characterized in that include: one or more processors; A memory for storing one or more computer programs; wherein, when executed by a processor, the one or more computer programs cause the computing device to perform any one of the following methods: A gated multi-graph convolutional perception modeling method for traffic flow prediction according to any one of claims 1 to 10.

12. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor of a computing device, the computing device is caused to perform any one of the following methods: A gated multi-graph convolutional perception modeling method for traffic flow prediction according to any one of claims 1 to 10.

Images