Traffic flow prediction model for multilayer space-time structure correlation perception

Abstract

The invention relates to the technical field of traffic prediction, in particular to a multi-layer space-time structure correlation perception traffic flow prediction model, which comprises a space-time structure decomposition layer for decomposing original traffic flow data into a road hierarchical structure feature matrix, a dynamic time period feature matrix and a hidden space topology feature matrix; the dynamic adjacency matrix generation layer is used for constructing a self-adaptive adjacency weight matrix based on the hidden space topological feature matrix; the cross-level correlation perception layer performs bidirectional feature modulation on the road hierarchical structure feature matrix and the dynamic time period feature matrix to generate a space-time coupling feature tensor; and the prediction layer inputs the space-time coupling feature tensor into a space-time diagram convolution prediction network and outputs a traffic flow prediction value in a future time period. According to the method, the expression ability of the model on potential heterogeneous association between the nodes is improved, and the generalization ability across regions and time periods is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic prediction, and in particular to a multi-layer spatiotemporal structure correlation-aware traffic flow prediction model. Background Art

[0002] With the development of intelligent transportation systems, data-driven traffic flow forecasting technology has gradually become a core support for urban traffic scheduling and management. Traditional forecasting methods mainly include statistical methods based on time series modeling and end-to-end regression methods based on deep learning. These methods can capture the temporal trends of traffic flow to a certain extent, but they generally rely on modeling in a single time dimension, ignoring the extensive structural heterogeneity of road networks. This makes them difficult to adapt to the modeling needs of multi-level spatial dependencies in large-scale urban transportation systems.

[0003] To overcome the shortcomings of spatial structure modeling, graph neural networks (GNNs) have been widely used in traffic flow prediction tasks in recent years. The basic idea is to represent the traffic network as a graph structure, with nodes representing road sections or intersections and edges representing traffic flow relationships. Information propagation between nodes is modeled through graph convolution operations. However, existing GNN methods often rely on predefined static adjacency matrices, fail to fully account for the hierarchical differences between main roads and branches in the road network structure, and fail to effectively incorporate the dynamic characteristics of periodic fluctuations and non-stationary disturbances in the temporal dimension. This results in limited prediction accuracy in scenarios with complex topological changes or significant periodic disturbances.

[0004] In addition, the spatiotemporal feature fusion process generally adopts a simple splicing or weighting strategy in existing models, which lacks the characterization of the two-way modulation mechanism between the hierarchical structure and time period in the transportation network, making it difficult to capture the deep dependency relationship at multiple levels and scales. Summary of the Invention

[0005] The present invention provides a multi-layer spatiotemporal structure correlation-aware traffic flow prediction model.

[0006] A multi-layer spatiotemporal structure correlation-aware traffic flow prediction model, including:

[0007] The spatiotemporal structure decomposition layer decomposes the original traffic flow data into a road hierarchical structure feature matrix, a dynamic time period feature matrix, and an implicit spatial topology feature matrix;

[0008] The dynamic adjacency matrix generation layer constructs an adaptive adjacency weight matrix based on the implicit spatial topology feature matrix, where the weights between nodes are obtained by dynamically calculating the cross-layer feature similarity;

[0009] The cross-level correlation perception layer performs bidirectional feature modulation on the road hierarchical structure feature matrix and the dynamic time period feature matrix to generate a spatiotemporal coupling feature tensor;

[0010] The prediction layer inputs the spatiotemporal coupling feature tensor into the spatiotemporal graph convolution prediction network and outputs the traffic flow prediction value for the future period.

[0011] Optionally, the spatiotemporal structure decomposition layer specifically includes:

[0012] Multi-scale road-level feature extraction: A hierarchical convolutional network is used to extract multi-scale features from raw traffic flow data. Large-scale dilated convolution kernels are used at main road nodes to capture long-range dependencies, while small-scale conventional convolution kernels are used at branch nodes to extract local features. The output is a road-level structural feature matrix.

[0013] Dynamic time decomposition module: This module uses the variational mode decomposition algorithm to decompose the non-stationary signal of traffic flow time series data, combines it with a periodic significance detector to screen the intrinsic mode components with traffic regularity, and constructs a dynamic time period feature matrix;

[0014] Implicit spatial topology mining: The road hierarchical structure feature matrix and the dynamic time period feature matrix are tensor-concatenated, and the implicit spatial correlation patterns between nodes are mined through the graph attention network to generate the implicit spatial topology feature matrix.

[0015] Optionally, the dynamic adjacency matrix generation layer specifically includes:

[0016] Cross-layer feature projection: The node feature vectors in the implicit spatial topology feature matrix are concatenated with the road level weight coefficients and the time dimension scaling factors, and a cross-layer feature vector is generated through multi-layer perceptron mapping.

[0017] Calculate the cosine similarity of the cross-layer feature vectors between any two nodes, and superimpose the spatiotemporal attenuation constraint term to generate the original similarity matrix;

[0018] Adjacency matrix generation: The original similarity matrix is sparsified, the first k strongly connected edges are retained, and bidirectional weight normalization is applied to generate an adaptive adjacency weight matrix with dynamic topology awareness.

[0019] Optionally, the spatiotemporal decay constraint term is determined by an exponential decay function of the actual road distance between nodes and a dynamic time offset correlation.

[0020] Optionally, the method for constructing the adaptive adjacency weight matrix includes calculating weight values based on the cross-level connection strength between nodes, dynamic time offset correlation and implicit spatial distance attenuation function, wherein the dynamic time offset correlation is calculated using a sliding time window mutual information entropy method, and the window length is adaptively adjusted according to the traffic flow fluctuation coefficient.

[0021] Optionally, the cross-layer correlation perception layer specifically includes:

[0022] Hierarchical→temporal forward modulation: The road hierarchical structure feature matrix is input into the time dimension modulation gate, and the dynamic time period feature matrix is phase-calibrated by the periodic phase alignment unit to generate a road hierarchical feature vector enhanced in the time dimension.

[0023] Time→Level Inverse Modulation: The dynamic time period feature matrix is input into the spatial dimension modulation gate, and the hierarchical attention mask is used to redistribute the spatial weights of the road hierarchical structure feature matrix to generate a spatial dimension-corrected time period feature vector.

[0024] Spatiotemporal coupling fusion: The enhanced road-level feature vector is subjected to a tensor outer product operation with the corrected time period feature vector, and feature compression is performed through a learnable fusion weight matrix to output a spatiotemporal coupling feature tensor.

[0025] Optionally, the structure of the time dimension modulation gate includes a periodic phase alignment module, a dynamic time period feature enhancer and a residual connection path;

[0026] The structure of the spatial dimension modulation gate includes a hierarchical attention mask generator, a spatial reweighting module and a feature correction unit.

[0027] Optionally, the prediction layer specifically includes:

[0028] Hierarchical-aware convolution processing: The spatiotemporal coupled feature tensor is input into the hierarchical-aware convolution module, which dynamically adjusts the deformation parameters of the deformable convolution kernel based on the road hierarchy to which the node belongs, generating a feature map with enhanced spatial dimension. The receptive field of the convolution kernel for main road nodes is extended to a 3-hop neighborhood, while that for branch road nodes is limited to a 1-hop neighborhood.

[0029] Temporal dilation convolution: A temporal dilation convolution operation is applied to the spatially enhanced feature map. The dilation rate is dynamically set according to the length of the latent period detected, expressed as: Where T dominant is the dominant cycle duration output by the implicit cycle detection module, T base =15 minutes is the base time unit;

[0030] Cross-level feature aggregation: The dilated convolution output features are residually connected with the original spatiotemporal coupling feature tensor, and the weights of the aggregation results are redistributed through the hierarchical attention mechanism to suppress cross-level interference noise;

[0031] Multi-step prediction output: The aggregated features are input into the time axis back propagation unit and expanded by time step to generate traffic flow prediction values for the future period [t+1, t+k], where the prediction step length k is positively correlated with the length of the sliding time window.

[0032] Optionally, the spatiotemporal graph convolution prediction network includes alternately stacked level-aware convolution modules and time-axis expansion convolution modules, wherein the level-aware convolution module automatically adjusts the receptive field range of the convolution kernel according to the road level, and the level-aware convolution module includes a deformable convolution kernel, and the deformation parameters of the deformable convolution kernel are dynamically generated by the topological characteristics of the road level where the node is located.

[0033] Beneficial effects of the present invention:

[0034] This paper combines a "hierarchical-aware convolutional network" with a "variational mode decomposition + periodic saliency detection" mechanism to simultaneously model the spatial structure of traffic flow (main road / branch road differences) and non-stationary temporal cycle characteristics (such as morning and evening rush hour and weekend variations). This architecture enables the model to dynamically adjust its receptive field and time window, breaking through the limitations of traditional static adjacency graph and fixed time scale modeling methods, improving its adaptability to complex traffic pattern changes and predictive accuracy.

[0035] This invention innovatively designs a bidirectional modulation mechanism of "level→time forward modulation" and "time→level reverse modulation", and combines it with the tensor outer product fusion method of learnable weights to realize deep interactive modeling of structure-time bidirectional features, breaking the limitations of traditional feature splicing or simple weighted fusion, and improving the model's ability to express potential heterogeneous associations between nodes (such as the influence between low-frequency branches and high-frequency main roads), effectively improving the generalization ability across regions and time periods.

[0036] The present invention generates a dynamic adjacency matrix through cross-layer feature similarity calculation and spatiotemporal attenuation constraints, and combines it with the time axis dilation convolution operation driven by implicit cycles to give the model the ability to adaptively construct connections in a dynamic topological environment and the ability to perceive history under periodic changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0038] Figure 1 Schematic diagram of the prediction model structure of an embodiment of the present invention;

[0039] Figure 2 Schematic diagram of the prediction layer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0041] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0042] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0043] like Figure 1-Figure 2 As shown in Figure 1, a multi-layer spatiotemporal structure correlation-aware traffic flow prediction model includes:

[0044] The spatiotemporal structure decomposition layer decomposes the original traffic flow data into a road hierarchical structure feature matrix, a dynamic time period feature matrix, and an implicit spatial topology feature matrix;

[0045] The dynamic adjacency matrix generation layer constructs an adaptive adjacency weight matrix based on the implicit spatial topology feature matrix, where the weights between nodes are obtained by dynamically calculating the cross-layer feature similarity;

[0046] The cross-level correlation perception layer performs bidirectional feature modulation on the road hierarchical structure feature matrix and the dynamic time period feature matrix to generate a spatiotemporal coupling feature tensor;

[0047] The prediction layer inputs the spatiotemporal coupling feature tensor into the spatiotemporal graph convolution prediction network and outputs the traffic flow prediction value for the future period.

[0048] The spatiotemporal structure decomposition layer specifically includes:

[0049] aMulti-scale road level feature extraction:

[0050] The original traffic flow data X is extracted from multi-scale features through a hierarchical perceptual convolutional network, where a large-scale hole convolution kernel Conv is used for the main road nodes. dilate To capture long-range dependencies, branch nodes use small-scale conventional convolution kernels Conv local Extract local features; output road hierarchical structure feature matrix: F road =HPCN(X), where X represents the original traffic flow input data (number of nodes × time steps × number of channels), HPCN(·) represents the hierarchical perceptual convolutional network, and F road Represents the road hierarchical structure feature matrix;

[0051] b Dynamic time decomposition module:

[0052] The variational mode decomposition (VMD) algorithm is used to analyze the original time series traffic data X t Decompose the non-stationary signal and obtain several intrinsic mode function (IMF) sequences: Combined with period saliency detector Screen the modal components with periodicity and construct the time period characteristic matrix: Among them, X t represents the time series of node traffic flow, VMD(·) represents the variational mode decomposition function, and IMF i represents the i-th eigenmode function, represents the periodic significance detection function, θ represents the periodic significance threshold, F time Represents the dynamic time period characteristic matrix;

[0053] c Implicit spatial topology mining:

[0054] The output F of the previous step road With F time Perform channel dimension splicing to construct the fusion tensor F joint :F joint =concat(F road ,F time ); input graph attention network (GAT), mine the implicit spatial topological relationship between nodes, and output the implicit spatial topological feature matrix: F topo =GAT(F joint ,A init ); where concat(·) represents the feature concatenation operation (by feature dimension), GAT(·) represents the graph attention network module, and A init represents the initial adjacency matrix (which can be a degree-normalized adjacency matrix or an identity matrix), F topo Represents the implicit spatial topological feature matrix.

[0055] The dynamic adjacency matrix generation layer specifically includes:

[0056] a Cross-layer feature projection:

[0057] The implicit space topological feature matrix F topo Each node feature vector in is input into the multi-layer perceptron (MLP) and mapped into a cross-layer feature vector H of uniform dimension:

[0058] Among them, F topo Represents the implicit spatial topology feature matrix, H represents the cross-layer feature vector matrix, W1, W2 represent the mapping weight matrix of MLP, and the mapping weight structure includes λ road Road level adjustment coefficient and λ time Time dimension scaling coefficient, b1, b2 represent bias terms, σ(·) represents the activation function;

[0059] b Dynamic similarity calculation:

[0060] For any two nodes i, j, the cross-layer feature vector H i ,H j Perform cosine similarity calculation and superimpose the spatiotemporal attenuation constraint term to construct the original similarity matrix S: Among them, γ ij is the space-time attenuation constraint term, and γ ij =exp(-α·D ij )·ρ(ΔT ij ), S ij represents the original similarity value between nodes i and j, H i ,H j represents the cross-layer feature vector of node i, j, ||·|| represents the L2 norm of the vector, D ij represents the road space distance between nodes i and j, α represents the distance attenuation factor, ΔT ij represents the time offset of traffic flow between nodes, ρ(·) represents the time offset correlation function, which reflects the similarity of traffic delay;

[0061] c adjacency matrix generation:

[0062] Perform sparse processing on the similarity matrix S, retain the node connection edges corresponding to the first k maximum values in each row, and construct a sparse similarity matrix And perform bidirectional weight normalization to generate the adaptive adjacency matrix A:

[0063] A=Norm(TopK(S,k)), the normalization method is usually symmetric normalization:

[0064] Among them, TopK(S,k) means retaining the top k connections with the largest similarity by row to form a sparse similarity matrix D represents the sparse similarity matrix , Norm(·) represents the symmetric normalization operation, and A represents the final generated dynamic adaptive adjacency weight matrix.

[0065] The cross-layer feature vector H is used for similarity calculation; the generated adjacency matrix A is used as the input graph structure of the subsequent spatiotemporal graph convolution prediction network to participate in traffic flow prediction modeling.

[0066] The cross-level correlation perception layer specifically includes:

[0067] a level → time forward modulation:

[0068] The road hierarchy feature matrix F road Input cycle phase alignment module, dynamic time period feature matrix F time To perform a phase calibration: represents the periodic phase alignment unit, Represents the time period feature matrix after the time dimension is enhanced;

[0069] b Time→level reverse modulation:

[0070] The time period feature matrix F time Input spatial dimension modulation module, based on the hierarchical attention mask M road Perform attention weighting on the road structure feature matrix: in represents the spatial modulation gate, M road represents the hierarchical attention mask matrix, ⊙ represents element-wise multiplication, Represents the road structure feature matrix after spatial dimension correction;

[0071] c. Space-time coupling fusion:

[0072] Perform the tensor outer product operation on the two enhanced eigenvectors to construct the three-dimensional coupling tensor T spatio : Then use the learnable fusion weight matrix W fuse Compress and fuse tensors: F fuse =Fuse(T spatio ,W fuse ), Represents vector outer product operation (generating a two-dimensional matrix), T spatio represents the spatiotemporal coupling tensor at the node level, W fuse Represents the fusion weight tensor, F fuse Represents the final spatiotemporal coupling feature tensor, which is used as the input of the traffic flow prediction network.

[0073] The prediction layer specifically includes:

[0074] a-level perceptual convolution processing:

[0075] The spatiotemporal coupling feature tensor is input into the hierarchical perception convolution module, which dynamically adjusts the deformation parameters of the deformable convolution kernel according to the road level to which the node belongs, and generates a feature map with enhanced spatial dimension. The receptive field of the convolution kernel of the main road node is extended to the 3-hop neighborhood, while the receptive field of the branch node is limited to the 1-hop neighborhood, which can be expressed as follows:

[0076] Among them, H i represents the input feature vector of the i-th node; represents the spatial enhanced feature vector of the i-th node; represents the hop count neighborhood defined by the road hierarchy l∈{main road, branch road}; Represents the weight of the convolution kernel on the node i, j pair, which is adjusted with the level; represents a nonlinear transformation function;

[0077] b Time axis dilation convolution:

[0078] A time-axis dilation convolution operation is applied to the spatially enhanced feature map, and the dilation rate r is determined by the dominant period length T output by the implicit period detection module. p Dynamic setting, the calculation formula is Where r represents the expansion rate, which determines the jump step size between each time step; T p represents the duration of the dominant cycle, and Δt represents the sliding time window step of the time axis; Indicates rounding up operation;

[0079] c Cross-level feature aggregation:

[0080] Perform a residual connection between the dilated convolution output feature and the original spatiotemporal coupling feature tensor to obtain the intermediate fusion feature: H (res) =H (dilated) +H (coupled) ; Then, a hierarchical attention mechanism is introduced to redistribute the weights of the fused features: Among them, H (dilated) represents the time enhancement feature, H (coupled) represents the spatiotemporal coupling characteristic tensor, H (res) represents the feature vector after residual connection, represents the final feature vector after the aggregation of the i-th node, α i Represents the level attention weight of the i-th node, defined as w represents the attention weight vector, represents the transpose of the attention weight vector, W aRepresents the linear transformation matrix; ReLU(·) represents the rectified linear unit activation function; Represents the feature vector after residual connection of the i-th node; represents the residual connection feature vector of the jth node; exp(·) represents the exponential function, which is used to enhance the discrimination; ∑ j Indicates that a normalized sum operation is performed on all nodes j.

[0081] d Multi-step prediction output:

[0082] The aggregated feature H (agg) Input the time axis back propagation unit and expand it by time step to generate traffic flow prediction values for future time periods: in represents the traffic flow prediction value from time step t+1 to t+T; T represents the prediction step size (determined by the sliding window length Δt); represents the aggregated feature vector at time step t; Decoder(·) represents the time series decoder.

[0083] Optionally, the spatiotemporal graph convolution prediction network includes alternately stacked level-aware convolution modules and time-axis expansion convolution modules, wherein the level-aware convolution module automatically adjusts the receptive field range of the convolution kernel according to the road level, and the level-aware convolution module includes a deformable convolution kernel, and the deformation parameters of the deformable convolution kernel are dynamically generated by the topological characteristics of the road level where the node is located.

[0084] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0085] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A multi-layer spatiotemporal structure correlation-aware traffic flow prediction model, characterized by: include: The spatiotemporal structure decomposition layer decomposes the original traffic flow data into a road hierarchical structure feature matrix, a dynamic time period feature matrix, and an implicit spatial topology feature matrix; The dynamic adjacency matrix generation layer constructs an adaptive adjacency weight matrix based on the implicit spatial topology feature matrix, where the weights between nodes are obtained by dynamically calculating the cross-layer feature similarity; The cross-level correlation perception layer performs bidirectional feature modulation on the road hierarchical structure feature matrix and the dynamic time period feature matrix to generate a spatiotemporal coupling feature tensor; The prediction layer inputs the spatiotemporal coupling feature tensor into the spatiotemporal graph convolution prediction network and outputs the traffic flow prediction value for the future period.

2. The multi-layer spatiotemporal structure correlation-aware traffic flow prediction model according to claim 1, characterized in that: The spatiotemporal structure decomposition layer specifically includes: Multi-scale road-level feature extraction: A hierarchical convolutional network is used to extract multi-scale features from raw traffic flow data. Large-scale dilated convolution kernels are used at main road nodes to capture long-range dependencies, while small-scale conventional convolution kernels are used at branch nodes to extract local features. The output is a road-level structural feature matrix. Dynamic time decomposition module: This module uses the variational mode decomposition algorithm to decompose the non-stationary signal of traffic flow time series data, combines it with a periodic significance detector to screen the intrinsic mode components with traffic regularity, and constructs a dynamic time period feature matrix; Implicit spatial topology mining: The road hierarchical structure feature matrix and the dynamic time period feature matrix are tensor-concatenated, and the implicit spatial correlation patterns between nodes are mined through the graph attention network to generate the implicit spatial topology feature matrix.

3. The multi-layer spatiotemporal structure correlation-aware traffic flow prediction model according to claim 1, characterized in that: The dynamic adjacency matrix generation layer specifically includes: Cross-layer feature projection: The node feature vectors in the implicit spatial topology feature matrix are concatenated with the road level weight coefficients and the time dimension scaling factors, and a cross-layer feature vector is generated through multi-layer perceptron mapping. Calculate the cosine similarity of the cross-layer feature vectors between any two nodes, and superimpose the spatiotemporal attenuation constraint term to generate the original similarity matrix; Adjacency matrix generation: The original similarity matrix is sparsified, the first k strongly connected edges are retained, and bidirectional weight normalization is applied to generate an adaptive adjacency weight matrix with dynamic topology awareness.

4. The multi-layer spatiotemporal structure correlation-aware traffic flow prediction model according to claim 3, characterized in that: The spatiotemporal decay constraint term is determined by the exponential decay function of the actual road distance between nodes and the dynamic time offset correlation.

5. The multi-layer spatiotemporal structure correlation-aware traffic flow prediction model according to claim 3, characterized in that: The method for constructing the adaptive adjacency weight matrix includes calculating weight values based on the cross-level connection strength between nodes, dynamic time offset correlation and implicit spatial distance decay function, wherein the dynamic time offset correlation is calculated using a sliding time window mutual information entropy method, and the window length is adaptively adjusted according to the traffic flow fluctuation coefficient.

6. The multi-layer spatiotemporal structure correlation-aware traffic flow prediction model according to claim 1, characterized in that: The cross-level correlation perception layer specifically includes: Hierarchical→temporal forward modulation: The road hierarchical structure feature matrix is input into the time dimension modulation gate, and the dynamic time period feature matrix is phase-calibrated by the periodic phase alignment unit to generate a road hierarchical feature vector enhanced in the time dimension. Time→Level Inverse Modulation: The dynamic time period feature matrix is input into the spatial dimension modulation gate, and the hierarchical attention mask is used to redistribute the spatial weights of the road hierarchical structure feature matrix to generate a spatial dimension-corrected time period feature vector. Spatiotemporal coupling fusion: The enhanced road-level feature vector is subjected to a tensor outer product operation with the corrected time period feature vector, and feature compression is performed through a learnable fusion weight matrix to output a spatiotemporal coupling feature tensor.

7. The multi-layer spatiotemporal structure correlation-aware traffic flow prediction model according to claim 6, characterized in that: The structure of the time dimension modulation gate includes a periodic phase alignment module, a dynamic time period feature enhancer and a residual connection path; The structure of the spatial dimension modulation gate includes a hierarchical attention mask generator, a spatial reweighting module and a feature correction unit.

8. The multi-layer spatiotemporal structure correlation-aware traffic flow prediction model according to claim 1, characterized in that: The prediction layer specifically includes: Hierarchical-aware convolution processing: The spatiotemporal coupled feature tensor is input into the hierarchical-aware convolution module, which dynamically adjusts the deformation parameters of the deformable convolution kernel based on the road hierarchy to which the node belongs, generating a feature map with enhanced spatial dimension. The receptive field of the convolution kernel for main road nodes is extended to a 3-hop neighborhood, while that for branch road nodes is limited to a 1-hop neighborhood. Temporal dilation convolution: A temporal dilation convolution operation is applied to the spatially enhanced feature map. The dilation rate is dynamically set according to the length of the latent period detected, expressed as: Where T dominant is the dominant cycle duration output by the implicit cycle detection module, T base =15 minutes is the base time unit; Cross-level feature aggregation: The dilated convolution output features are residually connected with the original spatiotemporal coupling feature tensor, and the weights of the aggregation results are redistributed through the hierarchical attention mechanism to suppress cross-level interference noise; Multi-step prediction output: The aggregated features are input into the time axis back propagation unit and expanded by time step to generate traffic flow prediction values for the future period [t+1, t+k], where the prediction step length k is positively correlated with the length of the sliding time window.

9. The multi-layer spatiotemporal structure correlation-aware traffic flow prediction model according to claim 8, characterized in that: The spatiotemporal graph convolution prediction network includes alternately stacked level-aware convolution modules and time-axis dilation convolution modules, wherein the level-aware convolution module automatically adjusts the receptive field range of the convolution kernel according to the road level. The level-aware convolution module includes a deformable convolution kernel, and the deformation parameters of the deformable convolution kernel are dynamically generated by the topological characteristics of the road level where the node is located.

Images