Traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution

Abstract

The present application relates to a traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution. The method includes processing traffic flow data, time labels and adaptive spatiotemporal features using a spatiotemporal feature embedding module, extracting temporal features and spatial features from high-dimensional embedded features using an encoding module; the encoding module includes a multi-channel spatiotemporal encoder; fusing spatiotemporal features with temporal features and spatial features using a bidirectional spatiotemporal interaction fusion module; extracting features from spatiotemporal fusion features, traffic flow data, time labels and adjacency matrices using a fine-grained spatiotemporal graph convolution module, and then predicting using a prediction module to obtain traffic flow prediction results. The method uses a multi-channel spatiotemporal encoder to extract multi-scale temporal and spatial features, constructs global spatiotemporal dependencies through bidirectional spatiotemporal interaction fusion attention, and uses a graph convolution network that combines dynamic and static graphs to capture fine-grained dynamic spatiotemporal relationships, thereby improving the accuracy of traffic flow prediction.

Description

Technical Field

[0001] The present application relates to the field of intelligent transportation technology, and in particular to a traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution. Background Art

[0002] With the continuous advancement of urbanization and the continued rise in motor vehicle ownership, urban road congestion is becoming increasingly serious. Building an intelligent transportation system is becoming increasingly important, and traffic flow forecasting is a core component of such systems. Accurate traffic flow forecasting not only enables the foresight of traffic trends and proactive intervention in congestion hotspots, but also provides data support for optimized traffic light control, significantly improving road network resource utilization and reducing traffic accident rates.

[0003] Traffic flow forecasting is a technology that predicts future traffic flow at specific locations and time periods by analyzing spatiotemporal patterns in historical data collected by distributed sensors. Early traffic flow methods were primarily based on statistical methods, such as historical averages (HAs), vector autoregression (VARs), autoregressive smoothing average models (ARMA), and autoregressive integrated moving averages (ARIMA). These methods are only suitable for data with stable trends and are less effective when dealing with traffic flow data with complex nonlinear relationships. Later, machine learning-based methods, such as support vector machines (SVRs), K-nearest neighbors (KNNs), and Kalman filters (KFs), were applied to traffic flow forecasting. While these methods can model nonlinear relationships in traffic data to a certain extent, they only consider the temporal characteristics of the data and fail to account for spatial correlation, failing to capture the underlying spatiotemporal dependencies in traffic data. With the rapid development of deep learning, many deep learning methods have been introduced to traffic flow forecasting, achieving promising results, such as long short-term memory networks (LSTMs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). However, these methods still face various challenges. For example, while LSTM and RNN models complex nonlinear temporal dependencies, they fail to account for the spatial dependencies of nodes. CNNs can capture the spatial relationships between nodes by converting road network data into a grid, but they only work in Euclidean space and have limited effectiveness for irregular road networks or non-Euclidean spatial data.

[0004] To effectively capture the spatial structural characteristics and spatiotemporal dependencies of road networks, graph convolutional neural networks (GCNs) have been introduced to the field of traffic flow prediction. Yu proposed a spatiotemporal graph convolutional neural network model (STGCN). By integrating a fully convolutional architecture with graph convolutional layers and gated temporal convolutional layers, STGCN achieved, for the first time, efficient joint modeling of complex spatiotemporal dependencies in traffic forecasting. Compared to traditional methods, STGCN significantly improved medium- and long-term prediction accuracy and reduced training costs. Building on this, Li et al. proposed a diffused convolutional recurrent neural network (DCRNN). This model uses bidirectional random walks to model the diffusion process of traffic flow and incorporates dynamics into spatiotemporal sequence feature extraction. This model, for the first time, achieves joint dynamic modeling of spatiotemporal dependencies within a directed graph structure, effectively capturing spatiotemporal correlations using diffused convolutions. To address the GCN method's high reliance on prior knowledge, Wu et al. proposed Graph Wavenet. This model dynamically constructs a graph structure using learnable parameters, directly mining hidden spatial dependencies from data without relying on prior knowledge such as the road network adjacency matrix. Fang et al. also proposed an efficient spectral graph attention network (ESGAT) based on discrete wavelet transform to separate high- and low-frequency components of traffic flow. Through wavelet graph position encoding and query sampling strategy, the prediction accuracy is significantly improved with low computational complexity.

[0005] However, existing research still faces the following challenges:

[0006] (1) The spatiotemporal characteristics are highly dynamic

[0007] Many features in traffic data exhibit cyclical variations, such as traffic flow fluctuations during the morning rush hour and weekly cyclical fluctuations. These daily and weekly medium- to long-term cyclical variations are easy to capture. However, when faced with emergencies, weather changes, and other factors, the data detected by sensors at various locations can change rapidly and dynamically over time. These hourly or even minute-by-minute traffic flow fluctuations lack distinct cyclical characteristics and are often difficult to capture. This is because short-term fluctuations are easily masked by long-term cyclical features, and models lack the ability to effectively extract fine-grained features across different time scales.

[0008] (2) The deep relationship between spatiotemporal characteristics is complex

[0009] In traffic systems, various hidden relationships exist between different spatial nodes. For example, distant nodes may exhibit highly similar traffic fluctuation trends, while closer nodes may exhibit significant differences. This is because traffic flow exhibits multiple levels of dynamic spatiotemporal dependencies. For example, in emergencies (such as traffic accidents), traffic fluctuations can rapidly impact surrounding road network nodes, with localized fluctuations having a greater impact. In regular scenarios (such as learning about surrounding areas), however, nodes with similar functions but dispersed locations can exhibit synchronized traffic patterns. In these cases, it is important to pay attention to the potential connections between distant nodes.

[0010] Existing methods are difficult to effectively capture dynamic spatiotemporal characteristics and complex spatiotemporal interactions. Summary of the Invention

[0011] Based on this, it is necessary to provide a traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution to address the above technical problems.

[0012] A traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution, the method comprising:

[0013] The acquired traffic flow data, time labels and adaptive spatiotemporal features are input into the spatiotemporal feature embedding module and mapped through the fully connected layer to obtain high-dimensional embedded features.

[0014] The high-dimensional embedded features are input into the encoding module to obtain temporal features and spatial features; the encoding module includes several multi-channel spatiotemporal encoders; the multi-channel spatiotemporal encoder is used to extract temporal features through multi-channel parallel feature extraction, using the causal convolutional attention mechanism and the temporal self-attention mechanism to jointly extract temporal features, and using the spatial self-attention mechanism to extract spatial features.

[0015] The temporal features and spatial features are input into the bidirectional spatiotemporal interaction fusion module to obtain the spatiotemporal fusion features; the bidirectional spatiotemporal interaction fusion module is used to realize the dual-path attention mapping of time-guided space and space-guided time by establishing a directional attention path between dimensions.

[0016] The spatiotemporal fusion features, traffic flow data, time labels and adjacency matrix are input into the fine-grained spatiotemporal graph convolution module to obtain fine-grained dynamic spatiotemporal features; the fine-grained spatiotemporal graph convolution module is used to combine static adjacency features with dynamic graphs to perform graph convolution on daily features, weekly features and traffic flow data respectively, in order to extract fine-grained dynamic spatiotemporal features at multiple scales.

[0017] The fine-grained dynamic spatiotemporal features are input into the prediction module to obtain the traffic flow prediction results.

[0018] The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution includes the following steps: inputting acquired traffic flow data, time labels, and adaptive spatiotemporal features into a spatiotemporal feature embedding module, mapping them through a fully connected layer to obtain high-dimensional embedded features; inputting the high-dimensional embedded features into an encoding module to obtain temporal and spatial features; the encoding module includes several multi-channel spatiotemporal encoders; inputting the temporal and spatial features into a bidirectional spatiotemporal interaction fusion module to obtain spatiotemporal fusion features; inputting the spatiotemporal fusion features, traffic flow data, time labels, and an adjacency matrix into a fine-grained spatiotemporal graph convolution module to obtain fine-grained dynamic spatiotemporal features; and inputting the fine-grained dynamic spatiotemporal features into a prediction module to obtain traffic flow results. This method utilizes a multi-channel spatiotemporal encoder to extract multi-scale temporal and spatial features, constructs global spatiotemporal dependencies through bidirectional spatiotemporal interaction fusion attention, and captures fine-grained dynamic spatiotemporal relationships using a graph convolutional network that combines dynamic and static graphs. This method improves the accuracy of traffic flow prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 1 is a flow chart of a traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution in one embodiment;

[0020] Figure 2 The figure is an overall block diagram of a traffic flow prediction model based on multi-channel spatiotemporal interaction and fine-grained graph convolution in one embodiment;

[0021] Figure 3 Figure 2 is a structural diagram of a causal convolutional attention module in one embodiment. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0023] Definition 1 (Traffic Network) defines a transportation network as an undirected graph ,in yes A collection of nodes represents sensors located at the corresponding roads in the traffic network. Each sensor records traffic flow information. is the set of edges between adjacent nodes, corresponds to The adjacency matrix of .

[0024] Definition 2 (traffic flow sensor) Indicates the traffic network Nodes at time traffic flow, Indicates the traffic flow characteristic number. Represents the traffic flow sensors of all nodes in all time slices.

[0025] Traffic flow prediction problem definition: The traffic flow prediction problem aims to Traffic flow data for each time frame and transportation network information Training has parameters Model , and use the model to infer that all nodes will Traffic flow data for time steps , the above process is expressed as:

[0026] ;

[0027] in , , is the number of past time frames, is the number of future time frames, is the number of nodes, Input the number of features for the data. In this paper , which only indicates traffic flow.

[0028] In one embodiment, Figure 1 As shown in FIG, a traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution is provided, which includes the following steps:

[0029] Step 100: Input the acquired traffic flow data, time tags, and adaptive spatiotemporal features into the spatiotemporal feature embedding module and map them through the fully connected layer to obtain high-dimensional embedded features.

[0030] Specifically, the spatiotemporal feature embedding module primarily maps raw data, time tags, and the adaptive spatiotemporal feature matrix to a higher dimension through a fully connected layer to effectively extract spatiotemporal features. The adaptive spatiotemporal feature matrix is ​​a parameter matrix used to capture spatiotemporal features during the feature embedding phase. Initially, there is no data and the matrix is ​​dynamically adjusted during training.

[0031] Step 102: Input the high-dimensional embedded features into the encoding module to obtain temporal features and spatial features; the encoding module includes several multi-channel spatiotemporal encoders; the multi-channel spatiotemporal encoder is used to extract temporal features through a multi-channel parallel feature extraction method, using a causal convolutional attention mechanism and a temporal self-attention mechanism to jointly extract temporal features, and using a spatial self-attention mechanism to extract spatial features.

[0032] Specifically, in order to enhance the multi-scale feature extraction capability of the model, a multi-channel spatiotemporal encoder is proposed. The encoder uses causal convolutional attention and temporal self-attention to jointly extract temporal features, capturing multi-scale long-term and short-term temporal features while ensuring temporal causality, and uses the spatial self-attention module to capture the spatial topological relationship between different sensor nodes.

[0033] The multi-channel encoder consists of causal convolutional attention, temporal self-attention mechanism and spatial self-attention mechanism to extract deep multi-scale temporal and spatial features.

[0034] Abandoning traditional spatiotemporal feature extraction and fusion methods, this approach employs a multi-channel spatiotemporal encoder to extract multi-scale temporal and spatial features. This method employs a multi-channel parallel feature extraction approach and incorporates causal convolutional attention to capture multi-scale long- and short-term temporal features while preserving temporal causality. A temporal self-attention mechanism is employed to adaptively focus on key time nodes and long-term temporal dependencies. A spatial self-attention module is employed to capture spatial topological relationships between different sensor nodes. This parallel feature extraction approach not only avoids coupling interference between different features during the feature extraction phase but also effectively enhances the model's multi-scale feature learning capabilities.

[0035] Step 104: Input the temporal features and spatial features into the bidirectional spatiotemporal interaction fusion module to obtain spatiotemporal fusion features; the bidirectional spatiotemporal interaction fusion module is used to achieve dual-path attention mapping of time-guided space and space-guided time by establishing a directional attention path between dimensions.

[0036] Specifically, in order to achieve the effective fusion of temporal and spatial relationships, a bidirectional spatiotemporal interaction fusion module is proposed. By establishing a directional attention path between dimensions, a dual-path attention mapping of time-guided space and space-guided time is realized.

[0037] The bidirectional spatiotemporal interaction fusion module establishes a directional attention path between dimensions, achieving dual-path attention mapping: time guides space and space guides time. This method can overcome the limitations of separate spatiotemporal feature processing and achieve an effective fusion of temporal and spatial features.

[0038] Step 106: Input the spatiotemporal fusion features, traffic flow data, time tags, and adjacency matrix into the fine-grained spatiotemporal graph convolution module to obtain fine-grained dynamic spatiotemporal features; the fine-grained spatiotemporal graph convolution module is used to combine static adjacency features with dynamic graphs to perform graph convolution on daily features, weekly features, and traffic flow data, respectively, to extract fine-grained dynamic spatiotemporal features at multiple scales.

[0039] Specifically, in order to enhance the model's ability to capture dynamic changes in spatiotemporal features, a fine-grained spatiotemporal graph convolutional feature extraction module is proposed. This module combines static adjacency features with dynamic graphs to perform graph convolutional feature diffusion on weekly features (daily level), daily features (hourly level), and real-time traffic (minute level) features, respectively, to extract fine-grained dynamic spatiotemporal features at multiple scales.

[0040] The fine-grained spatiotemporal graph convolution feature extraction module combines static adjacency features with dynamic graphs to perform graph convolution operations on weekly features (daily level), daily features (hourly level), and real-time traffic (minute level) features, effectively capturing the complex spatiotemporal relationships between nodes to extract fine-grained dynamic spatiotemporal features at multiple scales.

[0041] Step 108: Input the fine-grained dynamic spatiotemporal features into the prediction module to obtain traffic flow prediction results.

[0042] like Figure 2 As shown, the overall model of the traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution proposed in this application consists of a multi-feature embedding layer, a multi-channel spatiotemporal encoder, a bidirectional spatiotemporal interaction fusion module, and a fine-grained spatiotemporal graph convolution feature extraction module.

[0043] The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution includes the following steps: inputting acquired traffic flow data, time labels, and adaptive spatiotemporal features into a spatiotemporal feature embedding module, where they are mapped through a fully connected layer to obtain high-dimensional embedded features; inputting the high-dimensional embedded features into an encoding module to obtain temporal and spatial features; the encoding module includes several multi-channel spatiotemporal encoders; inputting the temporal and spatial features into a bidirectional spatiotemporal interaction fusion module to obtain spatiotemporal fusion features; inputting the spatiotemporal fusion features, traffic flow data, time labels, and an adjacency matrix into a fine-grained spatiotemporal graph convolution module to obtain fine-grained dynamic spatiotemporal features; and inputting the fine-grained dynamic spatiotemporal features into a prediction module to obtain traffic flow results. This method utilizes a multi-channel spatiotemporal encoder to extract multi-scale temporal and spatial features, constructs global spatiotemporal dependencies through bidirectional spatiotemporal interaction fusion attention, and captures fine-grained dynamic spatiotemporal relationships using a graph convolutional network that combines dynamic and static graphs. This method improves the accuracy of traffic flow prediction.

[0044] In one embodiment, the multi-channel spatiotemporal encoder includes: a causal convolutional attention module, a temporal self-attention module, and a spatial self-attention module; step 102 includes: reshaping the high-dimensional embedded features and inputting them into the causal convolutional attention module and the temporal self-attention module of the first multi-channel spatiotemporal encoder respectively to obtain the outputs of the causal convolutional attention module and the temporal self-attention module; fusing the outputs of the causal convolutional attention module and the temporal self-attention module of the first multi-channel spatiotemporal encoder using a gated fusion mechanism to obtain a first temporal feature; inputting the high-dimensional embedded features into the spatial self-attention module of the first multi-channel spatiotemporal encoder to obtain a first spatial feature; using the first temporal feature as the input of the causal convolutional attention module and the temporal self-attention module of the second multi-channel spatiotemporal encoder, and using the first spatial feature as the input of the spatial self-attention module of the second multi-channel spatiotemporal encoder, and so on, until the last multi-channel spatiotemporal encoder to obtain temporal features and spatial features.

[0045] Specifically, in order to better capture the complex spatiotemporal dependencies in traffic flow data, this application proposes a multi-channel spatiotemporal encoder. The encoder adopts a multi-channel parallel feature extraction method and mainly includes three core components. The first is the causal convolution attention module, which is used to capture multi-scale long-term and short-term causal time features. The second is the temporal self-attention module, which captures dynamic and long-term temporal dependencies. The third is the spatial self-attention module, which captures the dynamic spatial relationship between different sensor nodes. This parallel feature extraction method can avoid coupling interference between different features and effectively enhance the multi-granularity feature learning ability of the model.

[0046] First, the high-dimensional embedding output by the embedding layer As the input of the multi-channel spatiotemporal encoder, in order to extract the temporal features, Reshape into: ,by As the input of the causal convolutional attention module and the temporal self-attention module, As the input of the spatial attention module, the three extract features in parallel.

[0047] In one embodiment, Figure 3 As shown, step 102 includes: in the causal convolutional attention module: performing layer normalization processing on the features input into the causal convolutional attention module to obtain normalized features.

[0048] ;

[0049] in, is the normalized feature, The features of the input causal convolutional attention module.

[0050] According to the normalized features, the query, key, and value required for the attention mechanism are calculated respectively; the expressions of query, key, and value are:

[0051] ;

[0052] in, 、 、 are the query, key, and value in the causal convolutional attention mechanism, respectively. 、 and are the three learnable parameter matrices in the causal convolutional attention mechanism.

[0053] Reshape the query, key, and value to obtain the reshaped results of the query, key, and value in the causal convolutional attention mechanism. The expression of the reshaped results of the query, key, and value in the causal convolutional attention mechanism is:

[0054] ;

[0055] in, 、 and They are 、 、 The reshaping result, is the number of attention heads, is the number of channels per attention head, and the total dimension of the data embedding is ; is the number of sensor nodes, is the number of time steps of input data, is the local window size, is the set of real numbers.

[0056] According to the reshaped result of query, key and value, the window attention score matrix under the local window is determined; wherein, the expression of the window attention score matrix under the local window is:

[0057] ;

[0058] in, is the window attention score matrix, is the dimension of the key vector, divided by This can prevent the dot product result from being too large and causing subsequent The gradient of the function vanishes.

[0059] According to the preset The matrix selectively expresses the window attention score matrix and normalizes it through Softmax to obtain the masked window attention score matrix; the expression of the masked window attention score matrix is:

[0060] ;

[0061] ;

[0062] in, is the window attention score matrix after mask processing, is the mask matrix of the window attention score matrix, is negative infinity.

[0063] The masked window attention score matrix is ​​weighted and summed with the values ​​in the causal convolution attention mechanism, and then layer normalization is performed. Then, it is processed through the fully connected layer MLP to obtain the output of the causal convolution time attention module. The output of the causal convolution time attention module is expressed as:

[0064] ;

[0065] in, is the output of the causal convolutional attention module, Indicates the traffic flow characteristic number.

[0066] Specifically, in order to better learn temporal correlation, this method proposes to use the causal convolutional attention mechanism CTMSA to extract temporal features and construct a causal convolutional attention module. First, in order to ensure the causal relationship of time, that is, to ensure that the state of each time step is closely related to the previous state and cannot peek into the future state, the constraint The time step characteristics can only be obtained through Secondly, in order to improve the model's ability to extract multi-scale time dependencies, window attention is used for calculation. Each time, only the time correlation of the time step within the window is extracted. In the first layer of attention, the window size is set to 3, in the second layer, the window size is set to 6, and in the third layer, the window size is set to 12. This not only extracts multi-scale time dependencies, but also effectively improves computational efficiency. The structure of the causal convolutional attention module is as follows: Figure 3 shown.

[0067] First, the input of the causal convolutional attention module is ,in is the number of sensor nodes, is the number of time steps of input data, is the total dimension of data embedding. In order to alleviate gradient explosion and gradient disappearance and stabilize the training process, we first use Perform normalization processing.

[0068] Assume the local window size is , using the multi-head attention mechanism idea, with window size The scales are grouped into Group, will Reshape into: Then use the query, key and value expressions to find the query required by the attention mechanism (Query), key (Key), Value (Value), then , , The matrix is ​​reshaped to obtain the reshaped results shown in the expressions of the reshaped results of the query, key, and value in the above causal convolutional attention mechanism.

[0069] Then the expression of the window attention score matrix under the local window is used to calculate the self-attention score matrix under the local window . Self-attention score matrix under local window It can capture the temporal relationship of different time steps. At the same time, in order to ensure the causality of the time series, a Matrix pair self-attention score matrix under local window For selective expression, assuming the window size is hour, The matrix is:

[0070] ;

[0071] in, is negative infinity.

[0072] A matrix is ​​a The lower triangular matrix of , which means that each time step is only connected to the previous time step. The upper triangular value is set (negative infinity), the purpose of this setting is to pass When normalizing, the values ​​of the upper triangle are made close to 0, so as not to affect the calculation of the lower triangle elements. Matrix pair self-attention score matrix under local window selectively expressed and Normalization is performed to obtain the result shown in the expression of the window attention score matrix after the above mask processing.

[0073] Finally, the masked window attention score matrix is ​​used right A weighted sum is performed, and then normalized through LayerNorm. The data is then restored to the same shape and dimension as the input through the fully connected layer MLP to obtain the output of the causal convolutional time attention module. The output of the causal convolutional time attention module is shown in the expression of the output of the causal convolutional time attention module.

[0074] In one embodiment, step 102 includes: in the temporal self-attention module: mapping the input features to the attention space through the learnable parameter matrix to obtain the query, key, and value required by the self-attention mechanism; the query, key, and value required by the self-attention mechanism are respectively:

[0075] ;

[0076] ;

[0077] ;

[0078] in, , , are the query, key, and value in the temporal attention mechanism, respectively. , , They are three independently learnable parameter matrices in the temporal self-attention mechanism, is the input feature.

[0079] Based on the query, key, and value required by the self-attention mechanism, the output of the temporal self-attention module is determined to be:

[0080] ;

[0081] in, is the output of the temporal self-attention module, is the dimension of the key vector, is the transpose of the key in the temporal attention mechanism.

[0082] In one embodiment, step 102 includes: in the spatial self-attention module: mapping the input features to the attention space through the learnable parameter matrix to obtain the query, key, and value required by the spatial self-attention mechanism; the query, key, and value required by the spatial self-attention mechanism are:

[0083] ;

[0084] ;

[0085] ;

[0086] in, , , are the query, key, and value in the spatial self-attention mechanism, respectively. is the input of the spatial attention mechanism, is the set of real numbers, Embed the total dimensions for the data; is the number of sensor nodes, is the number of time steps of input data, 、 and are the three learnable parameter matrices in the spatial self-attention mechanism.

[0087] Based on the query, key, and value required by the spatial self-attention mechanism, the output of the attention head is determined to be:

[0088] ;

[0089] in, is the self-attention mechanism, is the dimension of the key vector.

[0090] According to the output of the spatial self-attention mechanism and the multi-head attention mechanism, the output of the spatial self-attention module is:

[0091] ;

[0092] ;

[0093] in, is the output of the spatial self-attention module, , , Respectively i The query, key, and value of each attention head, For the i An attention head, It is a multi-head attention mechanism. is the number of attention heads.

[0094] Specifically, unlike the temporal attention module, the input of the spatial attention mechanism ,This module focuses on modeling the spatial dependencies between nodes, such as road topology association, ,regional functional similarity, etc. Its calculation method is similar to but not exactly the same as ,the temporal sub-attention.

[0095] In one embodiment, the outputs of the causal convolutional attention module and the temporal self-attention module are fused using a gated fusion mechanism to obtain the first temporal feature:

[0096] ;

[0097] ;

[0098] in, is the time series feature, is the output of the causal convolutional attention module, is the output of the temporal self-attention module, is the weight, is the output of the spatial self-attention module, for The weight of for The weight of .

[0099] Specifically, the input of the temporal attention mechanism is ,in Indicates the number of traffic nodes, is the time step, is the feature dimension. Temporal self-attention can be used to extract temporal context information and dynamic dependencies at different distances in the time dimension, such as periodic traffic fluctuations, temporal correlations of event triggers, etc. First, the input features are mapped to the attention space through a learnable parameter matrix to obtain the query required by the self-attention mechanism. ,key Sum .

[0100] The correlation weight matrix between time steps is then calculated by scaling the dot product, and the weight matrix is ​​used to calculate the value matrix Weighted summation is performed to extract temporal context features and obtain the output of the temporal self-attention module. .

[0101] The gated fusion mechanism is used to fuse the temporal features obtained by the causal convolutional attention and temporal attention. First, the output of the causal convolutional attention module is , the output of the temporal self-attention module ,pass Generate weight representation Z: ;in, are two learnable parameters, is the generated weight representation, and then the two parts are weighted and summed by the weight to obtain the first time series feature as shown in the above first time series feature expression .

[0102] In one embodiment, step 104 includes: establishing a dual-path mapping of time-guided space and space-guided time; wherein the attention mapping path expression of time-guided space is:

[0103] ;

[0104] ;

[0105] ;

[0106] in, is the query projected by the time feature, is the spatial feature, is the time feature, and are the keys and values ​​projected by spatial features, is the time query projection matrix, is the spatial key projection matrix, is the spatial value projection matrix.

[0107] The expression of the spatially guided temporal attention mapping pathway is:

[0108] ;

[0109] ;

[0110] ;

[0111] in, is the query projected by spatial features, and The keys and values ​​projected by the time feature, is the time query projection matrix, is the time key projection matrix, is the time value projection matrix.

[0112] According to the query projected by the time feature, the key and value projected by the spatial feature, the query projected by the spatial feature, the key and value projected by the time feature, dual-path parallel computing is used to make the time and space dimensions interact, and residual connection is introduced to fuse the obtained time-guided spatial path features and time-guided spatial path features to obtain the spatiotemporal fusion feature; the expression of the spatiotemporal fusion feature is:

[0113] ;

[0114] ;

[0115] ;

[0116] in, is the spatiotemporal fusion feature, is the time-guided spatial pathway feature, is the spatial guidance time path feature, is the self-attention mechanism, Normalizes operations for a layer.

[0117] Specifically, the dynamic changes in traffic flow characteristics are essentially the synergistic interaction of temporal and spatial features. Existing methods primarily fuse spatiotemporal features by serially stacking temporal and spatial modules. This approach results in temporal features focusing primarily on historical trends, while spatial features primarily model topological associations, making it difficult to capture dynamic interactions across dimensions. To address this issue, this paper proposes a bidirectional spatiotemporal interactive attention mechanism. This design overcomes the limitations of separate spatiotemporal feature processing and effectively integrates temporal and spatial relationships by establishing a directed attention pathway between dimensions.

[0118] First, we establish the dual-path attention mapping of time-guided space and space-guided time respectively. Assume that the attention mapping of time-guided space is , the spatially guided temporal attention map is The temporal feature tensor obtained in the previous module is , the spatial feature tensor is , respectively, perform multi-head projection, and the time-guided spatial attention mapping pathway is , as shown in the expression of the time-guided spatial attention mapping pathway. According to the time feature tensor Projected, and Then the spatial feature tensor get. For the temporal query projection matrix, we learn the mapping of temporal patterns to the attention space, which can decompose the features into multiple groups of subspace query vectors, each group focusing on different temporal evolution patterns, such as overall trends, periodic changes, and sudden events. The spatial key projection matrix is ​​used to establish the response relationship between spatial nodes and query space, converting spatial features into key vectors that can be retrieved by temporal queries, and representing which spatial nodes should be activated when the temporal features focus on a certain pattern. It is a spatial value projection matrix that encodes the information of spatial nodes, stores detailed information of spatial nodes, and provides specific feature values ​​after the attention weight is determined.

[0119] The mapping pathway of spatially guided temporal attention is , as shown in the mapping pathway expression of spatially guided temporal attention. Different from the temporally guided spatial pathway, the query By spatial feature tensor Projected, and The time eigenvector Projected. It is the spatial query projection matrix, which converts the spatial node state into multiple sets of query intents to establish a mapping relationship from spatial features to temporal patterns. The time key projection matrix encodes the retrievable features of historical time steps and is converted into a key vector that can be matched by spatial queries, representing which historical moments contain relevant information when the spatial node pays attention to a certain change state. It is a time-value projection matrix that stores the complete time-step context information and provides fine-grained history information when the attention is activated.

[0120] The dual-path parallel computing is used to make the time and space dimensions interact, and the residual connection is introduced to alleviate the gradient disappearance. Finally, the time guides the spatial path features. and space-guided time pathways Fusion results in the spatiotemporal fusion feature as shown in the expression of spatiotemporal fusion feature .

[0121] This method proposes a bidirectional spatiotemporal interaction fusion module. By establishing a directional attention path between dimensions, it achieves dual-path attention mapping, where time guides space and space guides time. This method can overcome the limitations of separate spatiotemporal feature processing and effectively integrate temporal and spatial relationships.

[0122] In one embodiment, the time tag includes: time-of-day features and day-of-week features; step 106 includes: high-dimensional embedding of traffic flow data to obtain a traffic flow embedding matrix; high-dimensional embedding and expansion of the time-of-day features and day-of-week features to obtain a time-of-day feature embedding matrix and a day-of-week feature embedding matrix; the traffic flow embedding matrix, the time-of-day feature embedding matrix and the day-of-week feature embedding matrix are processed using a self-attention mechanism to obtain traffic flow attention features, time-of-day attention features and day-of-week attention features; a static graph of the traffic network is extracted according to the adjacency matrix, and a dynamic graph is constructed, and the static graph and the dynamic graph are fused and processed through a fully connected layer to obtain a dynamic-static fusion graph; graph convolution is performed on the traffic flow attention features, time-of-day attention features and day-of-week attention features according to the dynamic-static fusion graph, and the obtained graph convolution results are fused with the spatiotemporal fusion features to obtain a fused dynamic graph convolution feature; the fused dynamic graph convolution feature is processed through a self-attention mechanism to obtain a fine-grained dynamic spatiotemporal feature.

[0123] Among them, the time feature refers to the hour of the day, and the week feature refers to the day of the week.

[0124] In one embodiment, the process of constructing a dynamic graph includes: defining a node embedding matrix for dynamically encoding semantic features of nodes, defining an edge embedding matrix for dynamically modeling spatial dependencies between nodes; generating a dynamic graph using the node embedding matrix and the edge embedding matrix; the expression of the dynamic graph is:

[0125] ;

[0126] in, For dynamic graph, is the node embedding matrix, is the edge embedding matrix, is the transpose of the node embedding matrix.

[0127] Specifically, traffic flow prediction in traffic systems faces the problem of difficulty in capturing dynamic changes in spatiotemporal relationships. Traditional methods often rely on a combination of predefined static graph structures and time series features, which makes it difficult to capture the real-time spatiotemporal correlations in road networks; at the same time, traditional methods usually simply splice and fuse time-of-day features, weekly features, and real-time traffic flow when embedding features. The three are heterogeneous features of different scales, and after simple splicing, they share the same projection space, which can easily cause high-frequency real-time traffic fluctuations to mask long-period patterns, leading to information conflicts and loss of key patterns. In order to solve the above problems, this application proposes a fine-grained graph convolution feature extraction module that combines dynamic and static graphs. This module uses a combination of static adjacency features and dynamic graphs to perform graph convolution operations on time-of-day features, weekly features, and real-time traffic flow features, respectively, to extract fine-grained dynamic spatiotemporal features at multiple scales.

[0128] Since the original data has a low dimension and cannot effectively represent the intrinsic characteristics of the data, the data is first embedded into a higher dimension to capture high-dimensional complex features:

[0129] ;

[0130] in 、 、 They are time of day features, weekday features and traffic flow data. is the trainable time feature embedding matrix, is the number of timestamps in a day, which is 288 in this example. is a predefined embedding dimension. is the trainable Zhoutian feature embedding matrix, is the day of the week, in this case 7. ,in is the traffic flow data, time step, is the number of sensor nodes. Then, in order to align the dimensions, and Expand to and .

[0131] Use the self-attention mechanism to extract high-dimensional features of each data:

[0132] ;

[0133] in, For data m high-dimensional features, 、 、 Represents three sets of data, , is the number of feature dimensions of each data.

[0134] Extracting a static graph of the traffic network using the adjacency matrix And build dynamic graph To build a dynamic graph, we first define the node embedding matrix Used to dynamically encode the semantic features of nodes and then define the edge embedding matrix It is used to dynamically model the spatial dependencies between nodes, and finally generate a dynamic graph using the node embedding matrix and the edge embedding matrix:

[0135] ;

[0136] Combine static images with dynamic images, static images Used to encode road network physical connections, dynamic diagrams Used to learn dynamic spatiotemporal patterns, the fusion of the two can combine physical topology and data-driven dynamic relationships to achieve scene-adaptive spatial dependency modeling. right Perform graph convolution feature propagation:

[0137] ;

[0138] in For the The feature matrix of the layer nodes, is the activation function. for The degree matrix satisfies . For the adjacency matrix with self-loops added, (is the identity matrix), which can be retained by adding self-loops The characteristics of each node in the . Indicates the history The node feature matrix of the layer, in this embodiment, hour, Corresponding to the input data , . is a learnable parameter matrix. The above process combines static adjacency features with dynamic graphs to perform graph convolution operations on time-of-day features, weekday features, and real-time traffic features. This effectively captures the complex spatiotemporal relationships between nodes and extracts fine-grained dynamic spatiotemporal features at multiple scales.

[0139] The fine-grained spatiotemporal graph convolution feature extraction module combines static adjacency features with dynamic graphs to perform graph convolution operations on weekly features (daily level), daily features (hourly level), and real-time traffic (minute level) features to extract fine-grained dynamic spatiotemporal features at multiple scales.

[0140] In one embodiment, the prediction module includes a fully connected layer.

[0141] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0142] In a validation example, four real-world public traffic datasets, PEMS03, PEMS04, PEMS05, and PEMS06, collected by a city transportation agency's PeMS (Performance Measurement System), were used. These datasets were collected in real time by sensors deployed on four highways in California, USA. The datasets primarily consist of two parts: traffic data and an adjacency matrix representing the traffic network. Traffic data was collected every five minutes, with each acquisition lasting 30 seconds, resulting in 12 frames per hour, for a total of 288 data points per day. These data points include traffic flow, road occupancy, and traffic speed. This example only used traffic flow information for the experiment. Dataset details are shown in Table 1.

[0143] Table 1 Dataset details

[0144]

[0145] Adjacency matrix in the dataset is a The matrix of is the number of sensor nodes. The adjacency matrix describes the spatial relationship and connectivity relationship of different nodes in the transportation network.

[0146] (1) Experimental setup

[0147] The dataset division for this experiment adopts the common division method in the field of traffic flow prediction, and the four datasets are divided into a training set (60%), a validation set (20%), and a test set (20%) in chronological order. All experiments were run on NVIDIA 4090 24G, using the Ubuntu 22.04 system, and the experimental environment was Pytorch 2.12, Python 1.10, and CUDA 11.8. The experimental task of this study is multi-step prediction, that is, using the traffic flow data of the past 1 hour (12 time steps) to predict the traffic flow data of the next 1 hour (12 time steps). The main hyperparameters of the model are set as follows: the input data step size in_steps is set to 12, the output data step size out_steps is set to 12, and the input embedding dimension d is set to 24. In the attention mechanism, the number of attention heads num_heads is set to 4, the number of attention layers num_layers is set to 3, and the forward feedback dimension feed_forward_dim is set to 24.

[0148] Set to 256. The number of training rounds epoch is set to 100, batchSize is set to 16, the initial learning rate lr is set to 0.001, and the learning rate is adjusted in the 25th, 45th, and 65th rounds respectively. The adjustment strategy is to multiply the learning rate by the variation coefficient each time. The variation coefficient in this experiment is 0.1.

[0149] (2) Evaluation indicators

[0150] The model is evaluated using three standard indicators in traffic flow prediction: (1) Mean Absolute Error (MAE), which is a basic indicator that reflects the actual situation of prediction accuracy. (2) Root Mean Square Error (RMSE), which is more sensitive to outliers. (3) Mean Absolute Percentage Error (MAPE), which can eliminate the influence of different data units to a certain extent. The above three evaluation indicators reflect the degree of deviation between the predicted value and the actual value from different perspectives. The smaller the value, the more accurate the prediction result. The calculation formula of the evaluation indicator is defined as follows:

[0151] ;

[0152] ;

[0153] ;

[0154] in represents the true value, Represents the predicted value. Indicates the number of observation samples. In the experiment, each model is used to conduct experiments in steps 1 to 12, that is, , and take the average value of the prediction results under 12 time steps as the experimental result.

[0155] (3) Experimental results and analysis

[0156] Table 2 Performance comparison of each model on the PEMS03, PEMS04, PEMS07, and PEMS08 datasets

[0157]

[0158] The performance comparison of our method with six mainstream models is shown in the table above. Table 2 shows the average prediction results for time steps 1 to 12, comparing the three performance metrics of the models on four traffic flow prediction datasets. The best results for each metric are indicated in bold, and the suboptimal results are underlined. As can be seen from Table 2, our model outperforms the other models in all metrics on all four datasets.

[0159] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0160] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution, characterized by: The method comprises: The acquired traffic flow data, time tags, and adaptive spatiotemporal features are input into the spatiotemporal feature embedding module and mapped through the fully connected layer to obtain high-dimensional embedded features. Inputting the high-dimensional embedded features into an encoding module to obtain temporal features and spatial features; the encoding module includes a plurality of multi-channel spatiotemporal encoders; the multi-channel spatiotemporal encoders are used to extract temporal features by using a multi-channel parallel feature extraction method, adopting a causal convolutional attention mechanism and a temporal self-attention mechanism to jointly extract temporal features, and adopting a spatial self-attention mechanism to extract spatial features; Inputting the temporal features and the spatial features into a bidirectional spatiotemporal interaction fusion module to obtain spatiotemporal fusion features; the bidirectional spatiotemporal interaction fusion module is used to achieve dual-path attention mapping of time-guided space and space-guided time by establishing a directional attention path between dimensions; The spatiotemporal fusion features, the traffic flow data, the time tags, and the adjacency matrix are input into a fine-grained spatiotemporal graph convolution module to obtain fine-grained dynamic spatiotemporal features. The fine-grained spatiotemporal graph convolution module is used to combine static adjacency features with dynamic graphs to perform graph convolution on the time-of-day features, weekly features, and traffic flow data, respectively, to extract fine-grained dynamic spatiotemporal features at multiple scales. The fine-grained dynamic spatiotemporal features are input into a prediction module to obtain a traffic flow prediction result.

2. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 1 is characterized in that: The multi-channel spatiotemporal encoder includes: a causal convolutional attention module, a temporal self-attention module, and a spatial self-attention module; The high-dimensional embedded features are input into the encoding module to obtain temporal features and spatial features, including: Reshaping the high-dimensional embedded features and inputting them into the causal convolutional attention module and the temporal self-attention module of the first multi-channel spatiotemporal encoder respectively to obtain outputs of the causal convolutional attention module and the temporal self-attention module; The outputs of the causal convolutional attention module and the temporal self-attention module of the first multi-channel spatiotemporal encoder are fused using a gated fusion mechanism to obtain a first temporal feature; Inputting the high-dimensional embedding feature into the spatial self-attention module of the first multi-channel spatiotemporal encoder to obtain a first spatial feature; The first temporal feature is used as the input of the causal convolutional attention module and the temporal self-attention module of the second multi-channel spatiotemporal encoder, and the first spatial feature is used as the input of the spatial self-attention module of the second multi-channel spatiotemporal encoder, and so on, until the last multi-channel spatiotemporal encoder, to obtain temporal features and spatial features.

3. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 2 is characterized in that: In the causal convolutional attention module: Perform layer normalization on the features of the input causal convolutional attention module to obtain normalized features; in, is the normalized feature, The features of the input causal convolution attention module; According to the normalized features, the query, key and value required for the attention mechanism are respectively calculated; the query, key and value are respectively: in, 、 、 are the query, key, and value in the causal convolutional attention mechanism, respectively. 、 and are the three learnable parameter matrices in the causal convolutional attention mechanism; The query, the key, and the value are reshaped to obtain the following reshaped results: in, 、 and They are 、 、 The reshaping result, is the number of attention heads, is the number of channels per attention head, and the total dimension of the data embedding is ; is the number of sensor nodes, is the number of time steps of input data, is the local window size, is the set of real numbers; Based on the reshaped result of query, key and value, the window attention score matrix under the local window is determined as: in, is the window attention score matrix, is the dimension of the key vector; According to the preset The matrix selectively expresses the window attention score matrix and normalizes it through Softmax. The masked window attention score matrix is: in, is the window attention score matrix after mask processing, is the mask matrix of the window attention score matrix, is negative infinity; The masked window attention score matrix is ​​weighted and summed with the values ​​in the causal convolution attention mechanism, and then layer normalization is performed. Then, it is processed through the fully connected layer MLP to obtain the output of the causal convolution time attention module: in, is the output of the causal convolutional attention module, Indicates the traffic flow characteristic number.

4. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 2 is characterized in that: In the temporal self-attention module: The input features are mapped to the attention space through the learnable parameter matrix to obtain the query, key, and value required by the self-attention mechanism; the query, key, and value required by the self-attention mechanism are: in, , , are the query, key, and value in the temporal attention mechanism, respectively. , , They are three independently learnable parameter matrices in the temporal self-attention mechanism, is the input feature; Based on the query, key, and value required by the self-attention mechanism, the output of the temporal self-attention module is determined to be: in, is the output of the temporal self-attention module, is the dimension of the key vector, is the transpose of the key in the temporal attention mechanism.

5. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 2 is characterized in that: In the spatial self-attention module: The input features are mapped to the attention space through the learnable parameter matrix to obtain the query, key, and value required by the spatial self-attention mechanism; the query, key, and value required by the spatial self-attention mechanism are: in, , , are the query, key, and value in the spatial self-attention mechanism, respectively. is the input of the spatial attention mechanism, is the set of real numbers, Embed the total dimensions for the data; is the number of sensor nodes, is the number of time steps of input data, 、 and are the three learnable parameter matrices in the spatial self-attention mechanism; Based on the query, key, and value required by the spatial self-attention mechanism, the output of the attention head is determined to be: in, is the self-attention mechanism, is the dimension of the key vector; According to the output of the spatial self-attention mechanism and the multi-head attention mechanism, the output of the spatial self-attention module is: in, is the output of the spatial self-attention module, , , Respectively i The query, key, and value of each attention head, For the i An attention head, It is a multi-head attention mechanism. is the number of attention heads.

6. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 2 is characterized in that: The outputs of the causal convolutional attention module and the temporal self-attention module are fused using a gated fusion mechanism to obtain the first temporal feature: in, is the time series feature, is the output of the causal convolutional attention module, is the output of the temporal self-attention module, is the weight, is the output of the spatial self-attention module, for The weight of for The weight of .

7. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 1 is characterized in that: Inputting the temporal features and the spatial features into a bidirectional spatiotemporal interactive fusion module to obtain spatiotemporal fusion features, including: Establish a dual-path mapping of time-guided space and space-guided time; among them, the attention mapping path of time-guided space is: in, is the query projected by the time feature, is the spatial feature, is the time feature, and are the keys and values ​​projected by spatial features, is the time query projection matrix, is the spatial key projection matrix, is the spatial value projection matrix; The spatially guided temporal attention mapping pathway is: in, is the query projected by spatial features, and The keys and values ​​projected by the time feature, is the time query projection matrix, is the time key projection matrix, is the time value projection matrix; According to the query projected by the time feature, the key and value projected by the spatial feature, the query projected by the spatial feature, the key and value projected by the time feature, dual-path parallel computing is used to make the time and space dimensions interact, and residual connection is introduced to fuse the obtained time-guided spatial path features and space-guided time path features, and the obtained spatiotemporal fusion features are: in, is the spatiotemporal fusion feature, is the time-guided spatial pathway feature, is the spatial guidance time path feature, is the self-attention mechanism, Normalizes operations for a layer.

8. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 1 is characterized in that: The time tag includes: time of day feature and week feature; The spatiotemporal fusion features, the traffic flow data, the time tags, and the adjacency matrix are input into a fine-grained spatiotemporal graph convolution module to obtain fine-grained dynamic spatiotemporal features, including: Performing high-dimensional embedding on the traffic flow data to obtain a traffic flow embedding matrix; Performing high-dimensional embedding and expansion on the time-of-day feature and the week-of-the-week feature, respectively, to obtain a time-of-day feature embedding matrix and a week-of-the-week feature embedding matrix; The traffic flow embedding matrix, the time-of-day feature embedding matrix, and the weekday feature embedding matrix are processed using a self-attention mechanism to obtain traffic flow attention features, time-of-day attention features, and weekday attention features; Extracting a static graph of the traffic network based on the adjacency matrix and constructing a dynamic graph, fusing the static graph and the dynamic graph and processing them through a fully connected layer to obtain a static-dynamic fusion graph; Perform graph convolution on the traffic flow attention feature, the time of day attention feature, and the weekly attention feature according to the dynamic and static fusion graph, and fuse the obtained graph convolution result with the spatiotemporal fusion feature to obtain a fused dynamic graph convolution feature; The fused dynamic graph convolution features are processed by a self-attention mechanism to obtain fine-grained dynamic spatiotemporal features.

9. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 8 is characterized in that: The process of building a dynamic graph includes: Define a node embedding matrix for dynamically encoding the semantic features of nodes, and define an edge embedding matrix for dynamically modeling the spatial dependencies between nodes; The node embedding matrix and the edge embedding matrix are used to generate a dynamic graph: in, For dynamic graph, is the node embedding matrix, is the edge embedding matrix, is the transpose of the node embedding matrix.

10. The traffic flow prediction method based on multi-channel spatiotemporal interaction and fine-grained graph convolution according to claim 1 is characterized in that: The prediction module includes fully connected layers.

Images