Traffic prediction method based on multi-view fusion and diffusion diagram convolution

Abstract

The invention discloses a traffic flow prediction method based on multi-view fusion and diffusion diagram convolution. The traffic flow prediction method is suitable for dynamic modeling and space-time dependence extraction in a complex traffic scene. The method comprises the following steps: firstly, extracting short-term fluctuation and long-term periodic characteristics through double-path time slice convolution, and modeling multi-scale time dependence; in the aspect of spatial modeling, three types of graph structures including a static physical graph, a historical semantic graph and a current feature graph are fused, a sparse dynamic graph is generated through a Top-K mechanism, multi-order diffusion graph convolution is executed in combination with a static graph, and local and global spatial features are extracted. Furthermore, a space-time fusion strategy of bidirectional cross gating is provided, the information circulation direction between short-term and long-term features is dynamically adjusted, and coordinated fusion of multi-scale features is realized. And finally, the traffic state of multiple time steps in the future is output through a gating structure and a non-regression prediction module. According to the method, the modeling capability of the model for the complex traffic dependency relationship is effectively enhanced, and the prediction precision and robustness are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of smart transportation and artificial intelligence technology, and specifically to a traffic flow prediction method based on multi-view fusion and diffusion graph convolution, which belongs to the research scope of urban traffic state modeling and prediction. Background Art

[0002] With the continuous advancement of urbanization and the rapid development of intelligent transportation systems, traffic flow forecasting, as a core technology for urban traffic management and scheduling, plays a vital role in alleviating traffic congestion, optimizing travel routes, and improving traffic efficiency. Traditional traffic forecasting methods, such as historical average (HA), autoregressive integrated moving average (ARIMA), and classic machine learning methods like support vector regression and random forests, have achieved some success in early scenarios. However, their modeling capabilities are significantly limited when faced with the strong nonlinearity, multi-source heterogeneity, and complex spatiotemporal dependencies inherent in traffic flow data, making it difficult to meet the real-time and high-precision forecasting requirements of modern transportation systems.

[0003] To this end, the recent rise of deep learning methods, particularly graph neural network (GNN) technology, has garnered widespread attention in the field of traffic flow prediction due to its expertise in modeling non-Euclidean spatial structures. Related research generally enhances the model's ability to represent traffic conditions by constructing static graph structures based on road topology or geographic distances and combining them with time series modeling techniques. However, most of these methods rely on a single spatial perspective and fail to effectively capture the coupling relationships between multidimensional information. Furthermore, most existing models employ fixed graph structures, which are unable to reflect the dynamic evolution of traffic conditions over different time periods, limiting their ability to model emergencies and time-varying patterns.

[0004] Furthermore, traffic flow data often exhibits significant multi-scale temporal dependencies, such as short-term fluctuations and long-term cyclical trends. Using only single-scale convolution or uniform time-domain processing makes it difficult to fully extract the underlying relationships between these temporal features. Some research has attempted to improve the generalization capabilities of models by introducing dynamic graph modeling and multi-scale temporal convolutional structures. However, existing methods still primarily focus on local feature extraction and lack global multi-graph fusion and interaction mechanisms, resulting in insufficient adaptability and robustness to complex traffic scenarios.

[0005] Therefore, there is an urgent need to propose a unified traffic prediction framework that supports multi-perspective spatial fusion, dynamic graph perception, and multi-scale temporal modeling, which can comprehensively capture the dynamic structure and evolution pattern of traffic flow, thereby achieving traffic flow prediction with higher accuracy and stronger generalization ability. Summary of the Invention

[0006] This paper aims to overcome the shortcomings of existing traffic flow prediction methods, such as their limited spatial modeling perspective, insufficient ability to express temporal dependencies, and a lack of dynamic modeling. It provides a traffic flow prediction method based on multi-perspective fusion and diffuse graph convolution. By incorporating a multi-source graph structure modeling mechanism, a temporal multi-scale perception strategy, and cross-gating fusion technology, this method significantly improves the model's generalization and prediction accuracy in complex traffic scenarios.

[0007] To achieve the above object, the present invention proposes the following technical solutions:

[0008] 1. Input Embedding Processing: Raw traffic sequence data (such as flow, speed, occupancy, etc.) is input into the linear transformation module, and a high-level representation of uniform dimensions is obtained through a trainable fully connected mapping. Furthermore, to enhance the model's ability to recognize periodic traffic fluctuations, week-cycle embedding and time-of-day embedding mechanisms are designed to capture weekday / holiday differences and intraday peak / valley patterns, respectively. Ultimately, by concatenating the raw features with the periodic time features, a fused spatiotemporal input representation is obtained, enabling subsequent modules to capture periodic patterns and non-stationary fluctuations in traffic signals.

[0009] 2. Temporal Multi-Perspective Modeling and Processing: This module utilizes a sliding window strategy to divide historical sequences into multiple time slices. It then introduces single-path (shallow temporal features) and dual-path (deep temporal features) causal convolutional networks to extract temporal dependency information within different receptive fields. The former is suitable for modeling short-term fluctuations and local anomalies, while the latter emphasizes capturing cyclical patterns and long-term trends. This module specifically introduces normalized residual connections and an attention-guided mechanism to interactively fuse time slice features at different scales, improving long-term information retention and alleviating the vanishing gradient problem.

[0010] 3. Spatial multi-perspective modeling and processing: The spatial structure consists of three types of adjacency graphs. (1) Static physical graph: This is constructed based on the geographic distance matrix of sensor nodes, using Gaussian kernel function or inverse distance function to generate weights, reflecting the basic road topology; (2) Historical semantic graph: This is constructed based on the feature similarity between nodes and the traffic pattern library, dynamically updating the adjacency weights to reflect the historical evolution trend; (3) Current feature graph: This is constructed based on the real-time feature similarity between nodes to reflect the instantaneous state coupling relationship.

[0011] After fusing the three graph structures, a Top-K sparse filtering mechanism is used to generate a sparse dynamic graph. Multi-order diffusion graph convolution is then performed on both the static and dynamic graphs to extract spatially dependent features at different scales. The convolution kernel weights are node-sensitive, supporting heterogeneous control of information propagation paths, effectively enhancing robustness and local generalization capabilities in complex road networks.

[0012] 4. Multi-perspective spatiotemporal fusion processing: This module proposes a bidirectional cross-gating mechanism (Bidirectional GatingFusion), which separately calculates the gating weights of the short-term branch for long-term features and the regulatory response of the long-term branch to short-term features. This mechanism introduces learnable gating functions in both the channel and node dimensions, enabling the model to dynamically adjust the flow paths of various information and the strength of feature fusion. Compared with traditional direct concatenation or weighted averaging methods, the cross-gating mechanism has stronger expressive power and dynamic adaptability, making it suitable for feature regulation in scenarios with varying traffic densities, emergencies, and holidays.

[0013] 5. Prediction Output Processing: Based on the fused features, gated linear units (GLUs) are used for channel-level selective enhancement, and traffic conditions are predicted for multiple future time steps through a series of fully connected networks. This method employs a non-autoregressive prediction strategy, generating all predictions at once in a global manner. This significantly reduces the cascading propagation of prediction errors in traditional autoregressive methods, while also achieving higher inference efficiency and stability.

[0014] The beneficial technical effects of the present invention include:

[0015] (1) Multi-scale time-dependence perception: Combining sliding window time slice decomposition with multi-path convolution extraction mechanism, it significantly enhances the joint modeling capability of short-term disturbances and long-term trends.

[0016] (2) Dynamic graph structure modeling: Based on the historical evolution and real-time changes of traffic, a multi-graph structure is constructed to effectively perceive the dynamic non-stationary characteristics of traffic status.

[0017] (3) Multi-perspective fusion mechanism: Introducing a cross-gating strategy to dynamically adjust the importance of features at each time scale and spatial path, and improving the model's adaptability to spatiotemporal coupling structures.

[0018] (4) Robust and stable prediction: GLU is used to enhance feature selection capabilities and combined with a non-autoregressive decoder to improve robustness and reasoning efficiency in abnormal traffic fluctuation scenarios.

[0019] (5) Strong generalization and scalability: The modular design can be easily transferred to other types of cities, road levels, or perception systems, and has extensive practical deployment and research and promotion value.

[0020] The purpose of the present invention is to overcome the problems of existing traffic flow prediction methods, such as a single spatial modeling perspective, insufficient time scale expression, and limited dynamic modeling capabilities. A traffic prediction method based on multi-perspective fusion and diffusion graph convolution is provided. By introducing multi-perspective fusion, dynamic graph modeling, and multi-scale time perception mechanisms, the model's ability to model complex spatiotemporal dependencies is enhanced, thereby improving the accuracy and adaptability of traffic flow prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] 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.

[0022] Figure 1 A traffic prediction method based on multi-view fusion and diffusion graph convolution;

[0023] Figure 2 Temporal Multi-Perspective Example;

[0024] Figure 3 Spatial multi-perspective example;

[0025] Figure 4 Dual-path temporal modeling based on temporal patch convolution;

[0026] Figure 5 Dynamic spatial modeling with adaptive diffuse convolution. DETAILED DESCRIPTION

[0027] To better understand the technical solution of the present invention, the specific implementation of the present invention is described below in conjunction with the application background of the traffic flow prediction task.

[0028] In the task of traffic flow prediction, the traffic network is usually modeled as a graph structure G = (V, E, A), where V represents the node set, which represents the observation points in the traffic system (such as sensors, road sections or intersections); E represents the edge set, which represents the connection relationship between nodes; A is the adjacency matrix, which represents the connection strength or weight between nodes and quantifies the mutual correlation between nodes in the spatial structure.

[0029] Assume that the traffic observation data is the traffic signal matrix represents the state data of N nodes in the traffic system within a certain time period, where F represents the characteristic dimension of each node (such as traffic flow, speed or density, etc.). The present invention aims to use historical traffic observation data to learn the spatiotemporal pattern of traffic flow in order to predict the traffic state of each node in multiple time steps in the future. The prediction process is based on the modeling of spatiotemporal sequences, that is, the traffic sequence X before a given historical moment t is t-T:t-1 Based on this, the traffic status of the next T′ steps is predicted It can be expressed as:

[0030]

[0031] Here, f(·) represents the traffic flow prediction model, G1 and G2 represent two types of spatial graph structures: G1 is a topological structure graph constructed based on static physical distance, and G2 is a time-varying graph structure dynamically generated according to historical traffic characteristics.

[0032] In order to fully capture short-term fluctuations and long-term cyclical trends, the present invention divides historical input into short-term sequences and long-term series Where T s and T l are the short-term and long-term time window lengths, T s <<T l Correspondingly, the model extracts short-term and long-term spatiotemporal features respectively, and combines the information of the two graph structures G1 and G2 to achieve multi-view and multi-scale modeling.

[0033] In the present invention, by constructing a multi-graph structure and combining time embedding, adaptive diffusion convolution and cross-gating mechanism, the joint modeling of dynamic spatial relationships and periodic temporal changes is achieved, thereby improving the accuracy, robustness and generalization ability of traffic flow prediction.

[0034] 1. Input embedding processing

[0035] In the present invention, in order to comprehensively extract the spatiotemporal features of traffic flow data, an input embedding module is used to map the original traffic data into a high-dimensional space to obtain a more expressive feature representation.

[0036] First, traffic data contains a variety of time-dependent characteristics, such as short-term fluctuations and long-term cyclical trends. To meet the modeling requirements of the time dimension, we first perform feature transformation on the original traffic data through a full connection operation to extract preliminary high-order features. The specific calculation formula is as follows:

[0037] H f =σ(W f X+b f )

[0038] Among them, X represents the original traffic signal data, W f is a trainable linear transformation matrix, b f is the bias term, σ(·) represents the activation function, which is used to enhance the nonlinear expression ability of features, and H f Represents the high-order feature representation after mapping.

[0039] On this basis, in order to model the periodic time information in traffic data, a learnable time embedding mechanism is introduced, using two embedding dictionaries to represent the week embedding and time point embedding respectively. Time Embedding Dictionary where N w =7 represents the 7 days of a week, N d =288 means 288 sampling time points in a day (sampling interval is 5 minutes). By extracting the week index T in the traffic data week and timestamp index T day , and using the above two embedding dictionaries, the corresponding periodic time embedding features can be obtained:

[0040] H week =E week [T week ]; H day =E day [T day ]

[0041] The above high-order features H f 、Week embedded H week and time embedding H day Perform splicing and fusion to obtain the final embedding representation H emb , whose dimension is TN 3d f , expressed as:

[0042] H emb =Concact(H week ,H day ,H f )

[0043] By fusing the aforementioned linear feature transformation with periodic temporal embedding, the model's ability to represent spatiotemporal dependencies can be significantly improved. To simplify the subsequent description, this paper will refer to this embedding as X for the subsequent modeling process.

[0044] 2. Temporal multi-perspective modeling and processing

[0045] In this paper, in order to enhance the modeling ability of time series dependency, a time patch convolution (TPConv) structure based on time patch and dual-path causal convolution is proposed to extract multi-scale time series features from historical traffic data.

[0046] This method divides the original time series into multiple time slices using a sliding window and uses different channel structures to extract short-term and long-term dynamic change features. To meet causal requirements, causal convolution is used on each time slice to avoid information leakage.

[0047] Assume that the original time series is X, the window length is L, and the sliding step is S. The number of time slices generated by the sliding window is:

[0048]

[0049] Where T represents the total length of the time series. The i-th time slice is defined as:

[0050]

[0051] Each time slice As local input, it is input into the single-path or dual-path causal convolution structure. To ensure that the convolution calculation only relies on historical information, TPConv adopts the following causal convolution form:

[0052]

[0053] Where θ is a one-dimensional convolution kernel, and the convolution operation only uses the features of the first k time steps.

[0054] To enhance the modeling capability of multi-scale temporal dependencies, TPConv adopts a dual-channel structure, performing one layer of convolution (single path) and two layers of stacked convolution (dual path) operations respectively. The output is defined as follows:

[0055]

[0056] Among them, θ1, θ2, θ3 are one-dimensional convolution kernels of different sizes, GELU represents the Gaussian error linear unit activation function, and F norm Represents the normalization function.

[0057] To integrate the outputs of the above two channels, the following fusion method is used:

[0058]

[0059] Among them, F con represents the channel dimension splicing function, F res represents the residual connection structure, θ4 is a 1×1 convolution kernel used for dimensionality reduction fusion.

[0060] After splicing the outputs of all time slices in sequence, we can get the global time feature representation H patch :

[0061]

[0062] Finally, in order to further integrate multi-segment information, a fully connected network is used to patch Perform feature compression to obtain the time modeling output X T :

[0063] X T =FC(H patch )

[0064] Through the above-mentioned TPConv module, the present invention can realize the efficient extraction of multi-scale time-dependent features, providing a robust time feature foundation for subsequent spatial modeling and spatiotemporal fusion.

[0065] 3. Spatial multi-view modeling and processing

[0066] In this paper, in order to effectively capture the dynamic spatial relationships in transportation networks, a spatial modeling method based on multi-graph structure fusion is designed, which integrates static topological information and data-driven dynamic graph structure to improve the accuracy and adaptability of spatial dependency representation.

[0067] First, in terms of spatial graph structure construction, the physical distance information between traffic nodes is used to construct a static adjacency matrix A. s , which is calculated as follows:

[0068]

[0069] Among them, d ij represents the geographical distance between node i and node j, σ is a hyperparameter that controls the range of distance decay, and ε is the set sparsification threshold used to filter weak connections.

[0070] To enhance the model's adaptability to dynamic changes, two data-driven methods are introduced to construct a dynamic adjacency matrix:

[0071] The first is based on historical traffic characteristics and a learnable traffic pattern library The similarity calculation between is defined as follows:

[0072]

[0073] The second is based on the autocorrelation between the current node features and is defined as:

[0074]

[0075] in, Represents the input features of each node at the current time step, d represents the feature dimension, ReLU represents the rectified linear unit activation function, and softmax is used for row normalization.

[0076] In order to fuse the above two dynamic graph structures, a fully connected fusion operation is used to generate a fusion graph:

[0077] A D =FC(A dyn,1 ‖A dyn,2 )

[0078] Where ‖ represents the tensor concatenation operation and FC represents the fully connected layer.

[0079] Since the fused adjacency matrix is in a fully connected form, it may introduce redundant connections and noise information. Therefore, an adaptive sparse matrix M is introduced to retain the most relevant τ neighbor connections of each node, and a Top-K screening mechanism is used to construct a sparse mask matrix:

[0080]

[0081] Finally, the sparse dynamic graph adjacency matrix is obtained:

[0082] A D =A D ⊙ M

[0083] where ⊙ represents the Hadamard (element-wise) product.

[0084] In the graph convolution propagation stage, the adaptive diffusion graph convolution network (ADGCN) is used to convolution the static adjacency matrix A. s and sparse dynamic graphs Multi-order diffusion propagation is performed simultaneously to enhance spatial modeling capabilities. The calculation formula is as follows:

[0085]

[0086] Among them, H (l) is the node feature of the lth layer, and are the normalized adjacency matrices of the static graph and the dynamic graph at the kth order, W s With W d are the learnable convolution kernel weights.

[0087] Through the joint modeling mechanism of static and dynamic graphs, the present invention can fully capture the multi-perspective and dynamically changing spatial dependencies in the transportation network, effectively improving the comprehensiveness and stability of spatial information extraction.

[0088] 4. Spatiotemporal fusion processing

[0089] To effectively integrate short-term and long-term spatiotemporal features, this paper proposes a multi-perspective spatiotemporal fusion method based on a cross-gated weighted fusion mechanism. This method dynamically integrates feature information at different time scales, enhancing the model's ability to model spatiotemporal dependency structures.

[0090] First, the spatiotemporal features from the short-term and long-term branches are represented as H short and H long Input to the interactive modeling module, combined with the static adjacency matrix A s and the dynamic graph adjacency matrix A d , perform spatiotemporal interaction operations:

[0091] H short ,H long =ST(H short ,H long ,A S ,A D )

[0092] Here, ST(·) represents the multi-perspective spatiotemporal interaction modeling function, which is used to achieve spatiotemporal information fusion and enhancement at different scales.

[0093] In the feature fusion stage, a cross-gating strategy is introduced to achieve bidirectional regulation between short-term and long-term features. By calculating the gating weight, short-term features are guided to influence long-term features, while also allowing long-term features to have a reverse regulatory effect on short-term features. The calculation formula for the gating weight is as follows:

[0094] Among them, [,] represents the feature concatenation operation, X T is the output feature of the temporal modeling module, and is a learnable linear transformation parameter, and σ(·) is an activation function (such as sigmoid) used to map the gating weights to the [0,1] interval.

[0095]

[0096] Subsequently, based on the calculated gating weights g1 and g2, a bidirectional weighted fusion operation is performed to obtain the final fusion feature representation H fusion , the specific calculation is as follows:

[0097] H fusion =g2·(H short W short )+g1·(H long W long )

[0098] Among them, W short With W long is a trainable weight matrix for projecting short-term and long-term features.

[0099] Through the above-mentioned gating mechanism, the model can adaptively fuse information from different time scales, achieve reasonable information transmission and integration, and thus improve the representation ability of spatiotemporal features and the final prediction effect.

[0100] 5. Prediction output processing

[0101] In this paper, the extracted fused spatiotemporal features are input into the output prediction module to generate the final traffic flow forecast. To improve feature selection and expression capabilities, a gated linear unit mechanism is introduced. This gated structure dynamically adjusts feature channels, achieving selective fusion and suppression of effective information.

[0102] The prediction stage consists of two sub-stages: (1) the gated feature processing stage, which uses the GLU structure to process the fusion feature H fusion Perform nonlinear transformation and screening; (2) In the regression mapping stage, use a fully connected network to project the above features and generate prediction output.

[0103] The calculation formula of the prediction signal is as follows:

[0104]

[0105] Where FC(·) represents a fully connected layer; ⊙ represents the Hadamard product (element-wise multiplication); σ(·) represents the Sigmoid activation function; ReLU(·) represents the rectified linear unit activation function; Represents the traffic flow prediction results output by the model.

[0106] The above structure enhances the expressibility of features through a nonlinear gating mechanism and outputs traffic prediction values for multiple time steps through a single-step regression method, avoiding the cumulative propagation of errors in the time dimension and improving the stability and inference efficiency of the overall prediction.

[0107] The present invention provides a traffic flow prediction method based on multi-perspective spatiotemporal dynamic graph convolution, which can adaptively construct a multi-graph structure that integrates static and dynamic relationships from the original traffic data, and deeply explore the potential spatial dependency features. The traffic sequence is modeled at multiple scales through the time slice convolution structure, effectively extracting short-term fluctuations and long-term trend features at different time granularities. At the same time, a spatiotemporal fusion module based on a cross-gating mechanism is designed to achieve information interaction and dynamic weighting between multi-perspective features, thereby improving the model's ability to express complex spatiotemporal patterns. In addition, the output module adopts a non-autoregressive method to further improve the prediction stability and reasoning efficiency. The present invention can update the traffic graph structure in real time and optimize the original topological expression. It has good accuracy, robustness and generalization capabilities, can provide a reliable data basis for prediction scheduling and decision support in intelligent transportation systems, and has broad practical application value.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present disclosure, rather than to limit them. Although the present disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present disclosure.

Claims

1. A traffic prediction method based on multi-view fusion and diffusion graph convolution, characterized by: The method comprises the following steps: Step S1: Input embedding processing, input the original traffic flow data into the embedding module, and perform linear transformation on it using a fully connected method to obtain a high-order feature representation H of unified dimension f ; Extract timestamp information and embed it into the dictionary with the week embedding dictionary Obtain periodic time features from ; concatenate the original features with the periodic embedding to generate a fused embedding representation H emb For simplicity, in the subsequent steps, the embedded representation is denoted as X, and step S2 is continued; Step S2: Temporal multi-perspective modeling processing, using a sliding window method to divide the historical sequence into multiple time segments, and using a single-path causal convolutional network and a dual-path stacked causal convolutional network in each time segment to extract short-term and long-term time dependency features Fusion of each time slice feature H patch , and obtain a unified temporal modeling representation X through the fully connected layer T ; Step S3: spatial multi-view modeling processing, constructing a multi-view spatial graph including a static adjacency structure and a dynamic graph structure; the static adjacency matrix A S Constructed based on the geographical distance of nodes; the dynamic graph includes the current node and the pattern library The first dynamic graph adjacency matrix A constructed based on the similarity between dyn,1 The second dynamic graph adjacency matrix A constructed by the current feature similarity between nodes dyn,2 After fusing the two, we perform sparsification through the Top-K mechanism to obtain the sparse dynamic graph adjacency matrix A. D ; Perform multi-order diffusion graph convolution on static graphs and sparse dynamic graphs respectively to extract the corresponding spatial dependency features H (l+1) ; Step S4: Spatiotemporal fusion processing, design a cross-gating mechanism to calculate the gating weight g1 of short-term features to long-term features, and the gating weight g2 of long-term features to short-term features; perform weighted fusion of bidirectional information flows through the gating mechanism to generate the final spatiotemporal feature representation H fusion ; Step S5: Prediction output processing, the fused spatiotemporal features are sent to the gate control linear unit for nonlinear channel adjustment, and then the fully connected layer is used to generate traffic flow prediction results for multiple future time steps The prediction is performed in a non-autoregressive manner to avoid error accumulation.

2. The method according to claim 1, wherein The input embedding process includes: using a fully connected layer to perform linear mapping on the original traffic flow data to extract high-order features; using a learnable daily embedding dictionary and a learnable weekly embedding dictionary to generate periodic time embeddings respectively; and splicing the time embeddings with the original features to form a unified input.

3. The method according to claim 1, wherein The temporal multi-view modeling includes: dividing the input sequence into multiple time slices, constructing single-path and dual-path convolution channels in parallel, and using different receptive fields to extract short-term fluctuation and long-term periodic dependency features respectively; the features of all time slices are concatenated and fully connected to generate a temporal modeling result.

4. The method according to claim 1, wherein The spatial multi-view modeling includes: constructing a first dynamic graph adjacency matrix (based on the similarity between nodes and pattern libraries) and a second dynamic graph adjacency matrix (based on the similarity of current features) respectively; after fusing the two, performing correlation scoring and screening the most relevant adjacency relationships based on the Top-K mechanism to generate a sparse dynamic graph structure.

5. The method according to claim 1, wherein The sparsity mechanism retains the most relevant K adjacent edges of each node in the adjacency matrix through the Top-K strategy, and uses them for dynamic graph convolution operations, reducing computational complexity and enhancing adjacency selectivity.

6. The method according to claim 1, wherein The diffusion graph convolution includes: performing multi-order diffusion convolution operations on a static adjacency matrix and a sparse dynamic graph adjacency matrix respectively; and weightedly combining convolution outputs of different orders through trainable fusion weights to generate a final spatial modeling representation.

7. The method according to claim 1, wherein The cross-gating mechanism includes two gating function modules, which respectively calculate the information flow weight of short-term features on long-term features and the adjustment weight of long-term features on short-term features; and improve the model's adaptability to multi-scale features through bidirectional gating adjustment.

8. The method according to claim 1, wherein The prediction output module includes: a gated linear unit for performing feature channel adjustment, followed by at least one fully connected network structure to output multi-step prediction results; the prediction strategy adopts a non-autoregressive structure to simultaneously generate all target time step outputs, improving inference speed and stability.

Images