A traffic flow prediction method based on adaptive dynamic fusion graph convolution network

Abstract

The application discloses a traffic flow prediction method based on an adaptive dynamic fusion graph convolution network, wherein in the case of a given traffic sequence input, an adaptive dynamic graph generator captures a dynamic time graph and a dynamic space graph of a traffic network, and respectively captures dynamic time and space trends. Then, a space-time fusion graph module is used to capture a dynamic space-time fusion graph of the traffic network, the collected traffic sequence and the dynamic space-time fusion graph are input into a space-time block, a nonlinear dependence relationship and a dynamic trend are enhanced, space-time features of the traffic sequence are obtained, the collected traffic sequence is input into an intrinsic semantic enhancement module, a traffic sequence after semantic enhancement is obtained, finally, the space-time features of the traffic sequence are integrated, then the traffic sequence after semantic enhancement is input into a multilayer perceptron, and a prediction result is obtained. The model provided by the application has better effect and can more accurately predict traffic flow.

Description

Technical Field

[0001] This application belongs to the field of traffic flow prediction technology, and in particular relates to a traffic flow prediction method based on an adaptive dynamic fusion graph convolutional network. Background Technology

[0002] With the rapid development of Intelligent Transportation Systems (ITS) and the continuous increase in urban traffic data, accurate traffic flow prediction plays an increasingly important role in real-world transportation networks. To capture the complex spatiotemporal dependencies in transportation networks and achieve accurate traffic prediction, scholars have proposed a large number of models to date.

[0003] Traditional traffic prediction model research utilizes Support Vector Regression (SVR), Vector Autoregression (VAR), or K-Nearest Neighbors (KNN) algorithms to extract temporal dependencies, aiming to achieve reliable short-term traffic flow predictions. In recent years, thanks to continuous advancements in computing resources and big data analytics methods, deep learning (DL) technology has been widely applied, and some deep learning-based models have achieved great success in feature extraction. For example, Long Short-Term Memory (LSTM) networks, a variant of RNNs, utilize gating to overcome the gradient vanishing problem during feature extraction. Temporal Convolutional Networks (TCNs), a CNN-based architecture, uses dilated convolutions to capture long-range temporal dependencies. However, all of the above models neglect non-Euclidean spatial dependencies in traffic networks.

[0004] Because Generative Graph Convolutional Networks (GCNs) can effectively capture non-Euclidean spatial dependencies in transportation networks, several spatiotemporal models based on GCNs have been proposed in recent years. For example, the Diffused Convolutional Recurrent Neural Network (DCRNN) combines the advantages of RNNs and GCNs, utilizing bidirectional diffuse graph convolutions to extract spatial dependencies. The Spatiotemporal Graph Convolutional Network (STGCN) uses one-dimensional gated convolutions to capture temporal dependencies and uses GCNs to construct spatial dependencies. However, these models all utilize static methods to capture spatiotemporal dependencies, neglecting the dynamic trends of transportation networks. Although some models capture some dynamic trends in transportation networks, the captured dynamic spatiotemporal trends are incomplete. They typically rely on prior knowledge such as predefined graphs and embedded node vectors to construct dynamic trends, which are static to some extent, limiting the real-time discovery of dynamic trends in transportation networks.

[0005] To date, some existing models have been able to capture nonlinear dependencies in traffic networks. For example, random forest models have been used to capture nonlinear temporal dependencies in traffic flow prediction; extreme gradient boosting techniques have been used to extract short-term nonlinear temporal dependencies in traffic networks; temporal attention mechanisms have been used to extract nonlinear temporal dependencies to improve the accuracy of traffic speed prediction; and self-attention mechanisms have been used to capture nonlinear temporal dependencies. However, all of these methods neglect nonlinear spatial dependencies, or the extracted nonlinear spatiotemporal dependencies are incomplete.

[0006] Therefore, extracting the nonlinear dependencies between sensors remains a challenge. Furthermore, most existing methods fail to fully capture the inherent semantics of traffic sequences, making it difficult for models to extract deep spatiotemporal dependencies within traffic networks. Summary of the Invention

[0007] The purpose of this application is to provide a traffic flow prediction method based on an adaptive dynamic fusion graph convolutional network. First, a novel adaptive dynamic graph generator is proposed to adaptively capture the dynamic spatiotemporal trends in traffic networks. Second, a new spatiotemporal fusion graph module is proposed to extract nonlinear spatiotemporal dependencies between sensors. Furthermore, a novel intrinsic semantic enhancement module is proposed to comprehensively mine the intrinsic spatiotemporal semantics of traffic sequences, thereby enhancing the representation capability of spatiotemporal dependencies in traffic networks.

[0008] To achieve the above objectives, the technical solution of this application is as follows:

[0009] A traffic flow prediction method based on an adaptive dynamic fusion graph convolutional network, wherein the adaptive dynamic fusion graph convolutional network includes an adaptive dynamic graph generator, a spatiotemporal fusion graph module, a spatiotemporal block, an intrinsic semantic enhancement module, and an output module. The traffic flow prediction method based on the adaptive dynamic fusion graph convolutional network includes:

[0010] The collected traffic sequences are input into an adaptive dynamic graph generator to capture the dynamic temporal and dynamic spatial graphs of the traffic network.

[0011] The spatiotemporal fusion graph module maps the dynamic temporal graph and the dynamic spatial graph into nonlinear vectors representing nonlinear dependencies in time and space, and uses a gating fusion mechanism to fuse them in order to capture the dynamic spatiotemporal fusion graph of the traffic network.

[0012] The collected traffic sequences and dynamic spatiotemporal fusion maps are input into the spatiotemporal block to enhance nonlinear dependencies and dynamic trends, thereby obtaining the spatiotemporal characteristics of the traffic sequences.

[0013] The collected traffic sequences are input into the internal semantic enhancement module to obtain semantically enhanced traffic sequences;

[0014] In the output module, the spatiotemporal features of the traffic sequence are integrated, and then input into a multilayer perceptron along with the semantically enhanced traffic sequence to obtain the prediction result.

[0015] Furthermore, the adaptive dynamic graph generator includes a dynamic temporal graph generator and a dynamic spatial graph generator. The dynamic temporal graph generator includes a time-position encoder, a time block, a skip connection, and a multilayer perceptron. The dynamic temporal graph generator performs the following operations:

[0016] A time-location encoder is used to encode an input traffic sequence and capture its temporal dependencies;

[0017] By stacking individual time blocks to capture the dynamic temporal trends of the traffic network, a gated temporal convolutional network is used to capture temporal features in each time block, and then layer regularization is used to regularize the captured temporal features.

[0018] After regularization, the temporal features are processed by skip connections and then input into a multilayer perceptron to obtain a dynamic temporal map.

[0019] The dynamic spatial graph generator includes an adaptive graph module, a spatial node embedding module, a layer regularization module, a linear layer, and a value retrieval module. The dynamic spatial graph generator performs the following operations:

[0020] An adaptive graph module is used to capture implicit spatial dependencies in the traffic network;

[0021] In the spatial node embedding module, a diffusion graph convolutional network is used to adaptively capture dynamic spatial features;

[0022] Then, after layer regularization and linearization, and after mean and normalization processing, the maximum K values ​​are obtained through the value retrieval module to obtain the dynamic space graph.

[0023] Furthermore, the step of inputting the collected traffic sequences and dynamic spatiotemporal fusion graph into the spatiotemporal block to enhance nonlinear dependencies and dynamic trends includes:

[0024] Within each spatiotemporal block, a gated temporal convolutional network is used to enhance nonlinear dependencies in the traffic sequences, which are then fused with dynamic spatiotemporal data. Figure 1 The dynamic trend is enhanced by using a diffusion graph convolutional network.

[0025] Furthermore, the intrinsic semantic enhancement module performs the following operations:

[0026] Daily and weekly embeddings are used to enhance the temporal periodicity of traffic sequences, and linear layers are employed to capture hidden temporal features.

[0027] The spatial directionality of traffic sequences is captured using diffusion graph convolutional networks and dynamic temporal graphs.

[0028] Gated temporal convolutional networks are used to change the dimensionality of the input data.

[0029] Furthermore, the output module performs the following operations:

[0030] The spatiotemporal features output by the spatiotemporal block are truncated in time length, and then skip connections are performed.

[0031] The features after skip connections and the semantically enhanced traffic sequence are input into a multilayer perceptron to obtain the prediction results.

[0032] This application proposes a traffic flow prediction method based on an adaptive dynamic fusion graph convolutional network. First, it introduces a novel ADGG method, which includes DTGG and DSGG to capture the dynamic temporal and spatial trends of the traffic network in real time, respectively. Second, it proposes a novel STFGM method for deep extraction of nonlinear spatiotemporal dependencies between sensors. Third, it proposes a novel ISEM method to capture the inherent spatiotemporal semantics of traffic sequences, thereby enhancing the expressive power of the spatiotemporal correlation of the traffic network. Compared to baseline models, the proposed model performs better and predicts traffic flow more accurately. Attached Figure Description

[0033] Figure 1 This is a flowchart of a traffic flow prediction method based on an adaptive dynamic fusion graph convolutional network, as proposed in this application.

[0034] Figure 2 This is a schematic diagram of the adaptive dynamic fusion graph convolutional network structure of this application.

[0035] Figure 3 This is a comparison chart of the actual and predicted values ​​of the sensor in the experiment.

[0036] Figure 4 This is a comparison chart of the actual and predicted values ​​of another sensor used in the experiment. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0038] In one embodiment, such as Figure 1As shown, a traffic flow prediction method based on an adaptive dynamic fusion graph convolutional network (ADFGCN) is provided. The ADFGCN includes an adaptive dynamic graph generator (ADGG), a spatiotemporal fusion graph module (STFGM), spatiotemporal blocks (ST-blocks), an intrinsic semantic enhancement module (ISEM), and an output module (OM). This embodiment of the traffic flow prediction method based on the ADFGCN includes:

[0039] Step S1: Input the collected traffic sequences into the adaptive dynamic graph generator to capture the dynamic time graph and dynamic spatial graph of the traffic network.

[0040] The traffic prediction problem can be defined by formula (1):

[0041]

[0042] in, Let represent traffic data at time step t, and D represent the number of input features. In this embodiment, the input features include traffic flow data, daily embedded data, and weekly embedded data. M(.) is a mapping function that uses P historical traffic data points to form a traffic sequence and predict Q future traffic data points. G t = (V,E,A,t) represents the graph state of the transportation network at time step t, where V represents a set of nodes, |V|=N, and N is the number of nodes; E represents a set of edges connecting the nodes. The adjacency matrix represents the edge weights; P represents the length of historical time to be considered, and Q represents the length of future time to be predicted.

[0043] like Figure 2 As shown, the Adaptive Dynamic Fusion Graph Convolutional Network includes an Adaptive Dynamic Graph Generator (ADGG), a Spatiotemporal Fusion Graph Module (STFGM), Spatiotemporal Blocks (ST-blocks), an Intrinsic Semantic Enhancement Module (ISEM), and an Output Module (OM).

[0044] Considering that prior knowledge (such as predefined graphs and embedded node vectors) can limit the model's ability to deeply explore the dynamic spatiotemporal trends of traffic networks, the ADGG proposed in this application, including the dynamic temporal graph generator DTGG and the dynamic spatial graph generator DSGG, can adaptively capture the dynamic spatiotemporal trends of traffic networks without any prior knowledge.

[0045] This embodiment employs a Time-to-Growth Generator (DTG) to adaptively generate a dynamic time map (DTG) to capture dynamic time trends. The DTGG generator includes a Time Position Encoder (TPE), time blocks, skip connections (SC), and a Multilayer Perceptron (MLP).

[0046] A time-position encoder (TPE) is used to encode an input traffic sequence and capture its temporal dependencies.

[0047] Given different time steps t, the encoded traffic sequences are shown in formulas (2) and (3).

[0048] LV=L(V) (2)

[0049]

[0050] in, Let L(.) represent the input traffic sequence, D be the number of features in the input traffic sequence, N be the number of nodes, and T be the length of the time dimension. L(.) represents a linear layer. This represents the traffic sequence after processing by the linear layer, where d is the number of features in the traffic sequence. c The selected feature dimension (even or odd), where k≥0 is a number for selecting either an even or odd dimension. This represents the encoded traffic sequence.

[0051] Next, DTGG captures the dynamic temporal trends of the traffic network by stacking several time blocks. Within each time block, temporal features are captured using a gated temporal convolutional network (G-TCN), which extracts dynamic temporal dependencies and preserves the sparsity of the traffic sequences. Then, layer regularization (LN) is applied to regularize the captured temporal features. Therefore, the i-th time block is defined as follows:

[0052] l i =ψ(X,η) i1 ,η i2 ,ε i1 ,ε i2 (4)

[0053] h i =LayerNorm(l i )+X (5)

[0054] Where ψ(.) represents G-TCN, η i1 η i2 ε i1 and ε i2 These are learnable parameters; LayerNorm(.) represents LN, l i and h i These are the outputs of the i-th G-TCN and the i-th time block, respectively.

[0055] As shown in formulas (6)-(8), this module uses SC and multilayer perceptron MLP to integrate the processing of the output of each time block and transform the shape of the input data into N×N.

[0056]

[0057] A DTG =Softmax(Mean(MLP(s))) (7)

[0058]

[0059] Among them, h i This represents the i-th time block, where 0 ≤ i <L t L t This represents the number of time blocks. ∑f(.) denotes SC (Structured Perceptron), where f(.) is the function for pruning the time length, and s represents the output of SC. Mean(.) is the averaging function used to average the current input. Softmax(.) is the regularization function used to regularize the current input. MLP(.) represents a multilayer perceptron used to integrate the outputs of different time blocks. λ(.) represents the ReLU function. and This indicates a linear layer. DTG stands for Dynamic Time Trends of Transportation Networks.

[0060] This embodiment employs a Dynamic Spatial Graph Generator (DSGG) to adaptively generate a dynamic spatial graph (DSG) to capture dynamic spatial trends. The DSGG includes an adaptive graph module (AG), a spatial node embedding module (SNE), a layer regularization module (LN), a linear layer (Linear), and a value retrieval module (TopK).

[0061] Compared to using predefined graphs that rely on prior knowledge, DSGG uses adaptive graphs (AG) to capture implicit spatial dependencies in traffic networks, as shown in Equation (9):

[0062] A AG =Softmax(λ(V1V2) T (9)

[0063] Here, V1 and V2 represent randomly generated learnable parameters, λ(.) represents the ReLU function, and Softmax(.) represents the regularization function used to regularize the current input. AG represents the inherent spatial dependencies within a transportation network.

[0064] As shown in formulas (10)-(14), DSGG employs a Spatial Node Embedding Module (SNE) and adaptively captures dynamic spatial features using AG, which does not require prior knowledge, and D-GCN, a diffusion graph convolutional network that can capture dynamic spatial features. and This indicates the output of SNE.

[0065] e = E + V (10)

[0066] g1=Θ([A AG ],e,W1) (11)

[0067] g2=Θ([A AG ],e,W2) (12)

[0068]

[0069]

[0070] in, The input traffic sequence is represented by , where D is the feature dimension of the input traffic sequence, and T is the length of the time dimension. E, E1, and E2 represent learnable parameters. Θ(.) represents D-GCN, and W1 and W2 represent learnable parameters of D-GCN. ⊙ represents the Hadamard product. Let α represent the Tanh activation function, where α(.) is the control parameter of the Tanh activation function.

[0071] Then, use and Generate DSG as shown in formulas (15)-(17).

[0072]

[0073]

[0074]

[0075] Where LayerNorm(.) represents LN, L(.) represents a linear layer, and G represents the layer using... and A constructed dynamic graph. This represents the Tanh activation function, where α(.) is the control parameter of the Tanh activation function. Mean(.) represents the averaging function, used to average the current input. Softmax(.) represents the normalization function, used to regularize the current input. This represents the processed graph. TopK(.) is a function that takes the K largest values ​​in the graph to maintain spatial sparsity between nodes. This represents a dynamic spatial diagram, illustrating the dynamic spatial trends of the transportation network.

[0076] Step S2: The dynamic time graph and dynamic space graph are mapped into nonlinear vectors representing nonlinear dependencies in time and space through the spatiotemporal fusion graph module, and then fused using a gating fusion mechanism to capture the dynamic spatiotemporal fusion graph of the traffic network.

[0077] This embodiment proposes a spatiotemporal fusion graph module (STFGM) to capture nonlinear spatiotemporal dependencies in transportation networks. For example... Figure 2 As shown, STFGM takes DTG, representing the dynamic temporal trend of the traffic network, and DSG, representing the dynamic spatial trend, as inputs, and its definition is as follows:

[0078] First, a nonlinear mapping method is used to map DTG and DSG into nonlinear vectors V respectively. t and V s To capture nonlinear time and space dependencies, as shown in equations (18) and (19).

[0079] V t =λ(LayerNorm(A) DTG ×E t (18)

[0080] V s =λ(LayerNorm(A) DSG ×E s (19)

[0081] Among them, E t and E s The parameters are learnable, where LayerNorm(.) represents LN and λ(.) represents the ReLU function.

[0082] Next, a gate fusion mechanism will be used to merge V... t and V s The fusion is performed as shown in formulas (20) and (21).

[0083] z=δ(V t ×V s T (20)

[0084] A DSTFG =Softmax(z⊙A DSG +(1-z)⊙A DTG )+A AG (twenty one)

[0085] Where z represents the fusion of A DTG and A DSG The gating function is δ(.), which represents the Sigmoid function of the adaptive control output, ⊙ represents the Hadamard product, and Softmax(.) represents the regularization operation. The Dynamic Spatiotemporal Fusion Graph (DSTFG) represents both the dynamic spatiotemporal trends of the transportation network and its nonlinear spatiotemporal dependencies.

[0086] Step S3: Input the collected traffic sequence and dynamic spatiotemporal fusion map into the spatiotemporal block to enhance the nonlinear dependency and dynamic trend, and obtain the spatiotemporal characteristics of the traffic sequence.

[0087] like Figure 2 As shown, this embodiment uses ST-blocks to further process the dynamic trends and nonlinear dependencies captured by ADGG and STFGM, respectively. ST-blocks use... As input, D is the feature dimension of the input, and T is the time dimension. ST-blocks employ a gated temporal convolutional network G-TCN to enhance non-linear dependencies and uses A... DSTFG A Diffuse Graph Convolutional Network (D-GCN) is used to enhance the dynamic trends of the traffic network. The definition of the j-th ST-block is shown in equations (22) and (23):

[0088] t j =ψ(V,η j1 ,η j2 ,ε j1 ,ε j2 ) (twenty two)

[0089] st j =Θ([A DSTFG A DSTFG T ],t j W j )+V (23)

[0090] Where ψ(.) represents G-TCN, η j1 η j2 ε j1 and ε j2 This represents the learnable parameters. Θ(.) represents D-GCN, W j These are the learnable parameters of D-GCN. t j and st j These represent the outputs of the j-th G-TCN and ST-block, respectively, thus obtaining the spatiotemporal characteristics of the traffic sequence.

[0091] Step S4: Input the collected traffic sequence into the internal semantic enhancement module to obtain the semantically enhanced traffic sequence.

[0092] like Figure 2 As shown, this embodiment proposes ISEM to extract complex intrinsic semantics (including temporal periodicity and spatial directionality) in traffic networks.

[0093] ISEM employs daily and weekly embeddings to enhance the temporal periodicity of traffic sequences and uses linear layers to capture hidden temporal features. Therefore, it can effectively capture traffic sequences... As input (where D is the feature dimension of the input and T is the time dimension), the time embedding vector TEV can be obtained through formula (24).

[0094] TEV = T DE (V)+T WE (V)+L ISEM (V)+V (24)

[0095] Among them, T DE (.) and T WE (.) represent daily embedding and weekly embedding, respectively. L ISEM (.) indicates a linear layer. Denotes the temporal embedding vector, where D T It is the time dimension.

[0096] Then, using the Diffuse Graph Convolutional Network D-GCN and A DTG To capture the spatial directionality of traffic sequences, a gated temporal convolutional network (G-TCN) is used to change the dimension of the input data, as shown in equations (25) and (26).

[0097] g=Θ([A DTG A DTG T ],ΤEV,W) (25)

[0098] O ISEM =ψ(g,η1,η2,ε1,ε2) (26)

[0099] Where Θ(.) represents D-GCN, W represents the learnable parameters, and g represents the output of D-GCN. ψ(.) represents G-TCN, where η1, η2, ε1, and ε2 represent the learnable parameters, and O ISEM This indicates the output of ISEM.

[0100] Step S5: In the output module, the spatiotemporal features of the traffic sequence are integrated, and then input into the multilayer perceptron along with the enhanced traffic sequence to obtain the prediction result.

[0101] The output module OM is used to process the outputs of ISEM and ST-blocks and obtain the prediction results, as shown in formulas (27)-(29).

[0102]

[0103] predictions = MLP(s + O) ISEM (28)

[0104]

[0105] Among them, st j This represents the output of the j-th ST-block, 0 ≤ j <L st L st This represents the number of ST-blocks. ∑f(.) denotes the skip connections SC, f(.) is a function of the pruning time length, and s represents the output of the SC. ISEM This represents the output of ISEM. MLP(.) is used to transform the shape of the input data and obtain the prediction result. λ(.) represents the ReLU function. and This indicates a linear layer. "predictions" represents the output of OM, indicating the results of traffic flow predictions.

[0106] This application also experimentally tested the effectiveness and performance of the aforementioned Adaptive Dynamically Fusion Graph Convolutional Network (ADFGCN). Two medium-sized datasets, PEMSD4(M) and PEMSD8(M), were used to validate the effectiveness of ADFGCN. Furthermore, four large publicly available datasets, PEMS03, PEMS04, PEMS07, and PEMS08, were used to evaluate the performance of ADFGCN compared to benchmark models. Details of the benchmark models used for comparison in the experiments are as follows:

[0107] HA: Historical Average Method. This model uses the average of 12 past time steps to predict the next 12 time steps.

[0108] VAR: Vector Autoregression, a model that uses traditional mathematical methods to capture the temporal dependencies of traffic sequences.

[0109] SVR: Support Vector Regression, which represents a linear traffic prediction model.

[0110] LSTM: Long Short-Term Memory Network. This model is an RNN-based neural network that uses a gating mechanism to adaptively capture long-distance temporal dependencies in traffic data.

[0111] DCRNN: Diffusion Convolutional Recurrent Neural Network. This model uses D-GCN to capture spatial dependencies and RNN to capture temporal dependencies.

[0112] STGCN: Spatiotemporal Graph Convolutional Network. This model uses GCN to capture spatial dependencies and gated-TCN to capture temporal dependencies.

[0113] ASTGCN: Spatiotemporal Graph Convolutional Network Based on Attention Mechanism. This model uses a spatiotemporal attention mechanism to capture the spatiotemporal dependencies of traffic networks.

[0114] Graph WaveNet: Graph WaveNet uses a holed gated-TCN to capture long-distance temporal dependencies and a GCN with a predefined and adaptive adjacency matrix to capture spatial dependencies.

[0115] Graph WaveNet(a): Graph WaveNet without a predefined adjacency matrix. This model uses only an adaptive adjacency matrix to capture spatial dependencies and does not rely on any prior knowledge.

[0116] STSGCN: Spatiotemporally Synchronous Graph Convolutional Network. This model uses a spatiotemporally synchronized graph convolutional module to synchronously capture spatiotemporal dependencies in a transportation network.

[0117] ASTGNN: A spatiotemporal graph neural network based on an attention mechanism. This model uses a time-trend-aware multi-head self-attention mechanism to extract temporal dependencies and constructs a dynamic graph to extract spatial dependencies.

[0118] GMAN: Graph Multi-Head Attention Mechanism Network. This model uses a spatiotemporal self-attention module to capture spatiotemporal dependencies, making it suitable for long-distance time modeling and effectively mitigating error propagation problems.

[0119] MTGNN: Multivariate Temporal Graph Neural Network. This model includes a self-learning graph module, a general graph convolution module, and gated-TCN, which can adaptively extract spatiotemporal dependencies.

[0120] AGCRN: Adaptive Graph Convolutional Recurrent Network, this model uses an adaptive graph and node adaptive parameter learning modules to capture spatial dependencies.

[0121] DGCRN: Dynamic Graph Convolutional Recurrent Network. This model uses dynamic graph convolutional modules to capture spatial dependencies, while using an RNN-based structure to capture temporal dependencies.

[0122] D2STGNN: Decoupled Dynamic Spatiotemporal Graph Neural Network. This model decouples implicit and inherent traffic information to capture dynamic spatial dependencies.

[0123] In the experiment, mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) were used to evaluate the performance of the model, and the ratio of training data, validation data, and test data was set to 6:2:2.

[0124] Tables 1 and 2 list the experimental results of ADFGCN and the baseline model on the PEMSD4(M) and PEMSD8(M) datasets, as well as the improvement ratio of the best results to the second-best results.

[0125] Table 1

[0126]

[0127]

[0128] Table 2

[0129]

[0130] Tables 1 and 2 demonstrate the superiority of ADFGCN compared to other benchmark models, as detailed below:

[0131] (1) The model based on GCN performs better than the model without GCN, which shows that spatial dependence in traffic networks is crucial, and that GCN, as a graph-based structure, can effectively capture spatial dependencies.

[0132] (2) The model based on GCN and with dynamic structure is better than the model based on GCN but without dynamic structure, which shows that capturing the dynamic characteristics of traffic network with dynamic structure is effective.

[0133] (3) Models based on GCN without prior knowledge (especially Graph WaveNet(a) and ADFGCN) outperform models based on GCN with prior knowledge on most evaluation metrics, indicating that prior knowledge (such as predefined graphs in GCN and node vectors in GMAN) can hinder the discovery of dynamic trends in traffic networks to some extent. It is particularly noteworthy that ADFGCN captures the dynamic spatiotemporal trends of traffic networks without any prior knowledge and exhibits the best predictive performance.

[0134] (4) ADFGCN outperforms models that do not have nonlinear spatiotemporal dependencies because these models do not take into account the nonlinear spatiotemporal correlations in the traffic network. For example, HA ignores nonlinear temporal dependencies. LSTM and GMAN use gating and attention mechanisms to construct nonlinear temporal dependencies, but ignore nonlinear spatial dependencies. In contrast, by utilizing the nonlinear mapping capability of STFGM, ADFGCN can capture the nonlinear temporal and spatial dependencies in the traffic network and fuse them through a gating fusion mechanism, thereby capturing the nonlinear spatiotemporal dependencies in the traffic network.

[0135] (5) ADFGCN outperforms models without an intrinsic semantic module because ISEM in ADFGCN can effectively mine the intrinsic spatiotemporal semantics of traffic sequences, thereby enhancing the ability to represent spatiotemporal correlations in traffic networks.

[0136] (6) Overall, ADFGCN has the best predictive performance for the next 15 minutes, 30 minutes, 60 minutes and average values ​​in terms of MAE, MAPE and RMSE, especially MAPE (mean>4%). Figure 3 The data shows the true values ​​of sensor #402762 in the PEMSD4(M) dataset on July 19, 2017, and the predicted values ​​from ADFGCN and Graph WaveNet. Figure 4 The graphs show the true values ​​of sensor #402104 in the PEMSD4(M) dataset on July 19, 2017, and the predicted values ​​from ADFGCN and Graph WaveNet. Both graphs demonstrate that ADFGCN outperforms Graph WaveNet in prediction.

[0137] Furthermore, Tables 3 and 4 show the performance improvement of ADFGCN compared to state-of-the-art benchmark models on four large public datasets: PEMS03, PEMS04, PEMS07, and PEMS08. It can be seen that ADFGCN outperforms the benchmark models on most metrics (especially on PEMS07 and PEMS08), thus demonstrating the superiority of ADFGCN.

[0138] Table 3

[0139]

[0140]

[0141] Table 4

[0142]

[0143] Furthermore, this experiment also validated the effectiveness of ADFGCN through two ablation experiments (structure ablation experiment and graph ablation experiment) based on the PEMSD4(M) dataset. Both experiments used three metrics—MAE, RMSE, and MAPE—to evaluate the next steps at 15 minutes, 30 minutes, 60 minutes, and mean.

[0144] To verify the effectiveness of the key structure of ADFGCN, a structural ablation experiment was designed, and four variants of ADFGCN were defined.

[0145] • Variant 1: w / o ISEM, ADFGCN without ISEM.

[0146] • Variant 2: w / o ISEM & STFGM, ADFGCN without ISEM and STFGM, i.e. using DTG and DSG independently to capture spatiotemporal dynamic trends.

[0147] • Variant 3: w / o ISEM&STFGM&DTG, without ISEM, STFGM and DTG, ADFGCN.

[0148] • Variant 4: w / o ISEM&STFGM&DTG&DSG, without ISEM, STFGM, DTG and DSG, ADFGCN. Based on this, according to the graph construction method proposed by Li et al. (2017), a static graph based on latitude and longitude information is used to capture spatial dependencies in the transportation network.

[0149] The results of the structural ablation experiments are shown in Tables 5 and 6, and the conclusions are as follows:

[0150] Table 5

[0151]

[0152] Table 6

[0153]

[0154] From Tables 5 and 6 above, we can conclude that:

[0155] (1) ADFGCN performs better than ISEM, indicating that the intrinsic spatiotemporal semantics in traffic networks are very important, and ISEM can capture the intrinsic semantics well.

[0156] (2) The performance of w / o ISEM is better than w / o ISEM & STFGM, indicating that STFGM can effectively capture the nonlinear spatiotemporal dependence between sensors.

[0157] (3) The performance of w / o ISEM&STFGM is better than w / o ISEM&STFGM&DTG, indicating that capturing the dynamic time trend of the traffic network helps to improve the model performance.

[0158] (4) w / o ISEM&STFGM&DTG outperforms w / o ISEM&STFGM&DTG&DSG, indicating that capturing the dynamic spatial trends of the traffic network helps improve model performance.

[0159] A DSTFG It is a graph structure of ADFGCN used to represent dynamic spatiotemporal trends and nonlinear spatiotemporal dependencies in transportation networks. To verify A... DSTFG To assess its effectiveness, an experimental design was created to perform a graphical ablation experiment, where A... DSTFG A comparison was made with five graph construction methods. Furthermore, to eliminate the influence of intrinsic spatiotemporal semantics on predictions, ISEM was removed to ensure fairness in the experiment.

[0160] • w / o ISEM: ADFGCN without ISEM.

[0161] • w / o ISEM(dynamic): ADFGCN without ISEM. A dynamic graph is constructed to reflect the dynamic spatial trends of the traffic network.

[0162] • w / o ISEM(adaptive-1): ADFGCN without ISEM. A one-way adaptive graph was constructed to represent the hidden spatial trend.

[0163] • w / o ISEM(adaptive-2): ADFGCN without ISEM. An adaptive graph containing two learnable vectors is constructed to represent the hidden spatial trends in the traffic network.

[0164] • w / o ISEM(static): ADFGCN without ISEM. A static map containing latitude and longitude information is constructed to represent the static spatial trends in the transportation network.

[0165] The results of the ablation experiments are shown in Tables 7 and 8:

[0166] Table 7

[0167]

[0168] Table 8

[0169]

[0170] From the experimental data in Tables 7 and 8, the following conclusions can be drawn:

[0171] (1) w / o ISEM(static) shows poor performance, indicating that static structures that rely on prior knowledge cannot fully reflect spatial trends.

[0172] (2) w / o ISEM(adaptive-1) and w / o ISEM(adaptive-2) perform better than w / o ISEM(static), indicating that the adaptive graph can better represent the hidden spatial trends in the traffic network.

[0173] (3) w / o ISEM(dynamic) performed the worst. The reason is that although w / o ISEM(dynamic) can reflect the dynamic spatial trend of the traffic network, it has a large number of parameters, is difficult to converge, and ignores the dynamic time trend and nonlinear spatiotemporal correlation, resulting in a significant decrease in prediction performance.

[0174] (4) In contrast, due to w / o ISEM (i.e., A DSTFGIt can capture both dynamic spatiotemporal trends and nonlinear spatiotemporal dependencies, thus achieving the best experimental results among all graph construction methods.

[0175] This application proposes a novel ADFGCN model to achieve accurate traffic flow prediction. Specifically, ADFGCN contains three innovative modules, summarized below.

[0176] First, a novel ADGG model is proposed, which includes DTGG and DSGG to capture the dynamic temporal and spatial trends of traffic networks in real time, respectively. Second, a novel STFGM model is proposed for deep extraction of nonlinear spatiotemporal dependencies between sensors. Third, a novel ISEM model is proposed to capture the inherent spatiotemporal semantics of traffic sequences, thereby enhancing the expressive power of spatiotemporal correlations in traffic networks. Finally, multiple experiments on six real traffic flow datasets validate the superiority of ADFGCN.

[0177] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A traffic flow prediction method based on adaptive dynamic fusion graph convolutional networks, characterized in that, The adaptive dynamic fusion graph convolutional network includes an adaptive dynamic graph generator, a spatiotemporal fusion graph module, a spatiotemporal block, an intrinsic semantic enhancement module, and an output module. The traffic flow prediction method based on the adaptive dynamic fusion graph convolutional network includes: The collected traffic sequences are input into an adaptive dynamic graph generator to capture the dynamic temporal and dynamic spatial graphs of the traffic network. The spatiotemporal fusion graph module maps the dynamic temporal graph and the dynamic spatial graph into nonlinear vectors representing nonlinear dependencies in time and space, and uses a gating fusion mechanism to fuse them in order to capture the dynamic spatiotemporal fusion graph of the traffic network. The collected traffic sequences and dynamic spatiotemporal fusion maps are input into the spatiotemporal block to enhance nonlinear dependencies and dynamic trends, thereby obtaining the spatiotemporal characteristics of the traffic sequences. The collected traffic sequences are input into the internal semantic enhancement module to obtain semantically enhanced traffic sequences; In the output module, the spatiotemporal features of the traffic sequence are integrated, and then input into the multilayer perceptron along with the semantically enhanced traffic sequence to obtain the prediction result; The adaptive dynamic graph generator includes a dynamic temporal graph generator and a dynamic spatial graph generator. The dynamic temporal graph generator includes a time-position encoder, a time block, a skip connection, and a multilayer perceptron. The dynamic temporal graph generator performs the following operations: A time-location encoder is used to encode an input traffic sequence and capture its temporal dependencies; By stacking individual time blocks to capture the dynamic temporal trends of the traffic network, a gated temporal convolutional network is used to capture temporal features in each time block, and then layer regularization is used to regularize the captured temporal features. After regularization, the temporal features are processed by skip connections and then input into a multilayer perceptron to obtain a dynamic temporal map. The dynamic spatial graph generator includes an adaptive graph module, a spatial node embedding module, a layer regularization module, a linear layer, and a value retrieval module. The dynamic spatial graph generator performs the following operations: An adaptive graph module is used to capture implicit spatial dependencies in the traffic network; In the spatial node embedding module, a diffusion graph convolutional network is used to adaptively capture dynamic spatial features, including: ； ； ； ； ； in, This represents the implicit spatial dependencies in the traffic network captured by the adaptive graph module. This represents the input traffic sequence. , and Represents the learnable parameters. This represents a diffusion graph convolutional network. and This represents the learnable parameters of the diffusion graph convolutional network. Represents the Adama product. This represents the Tanh activation function. These are the control parameters of the Tanh activation function. and This indicates the output of the spatial node embedding module; Then, after layer regularization and linearization, and after mean and normalization processing, the maximum value is finally obtained through the value retrieval module. These values ​​yield a dynamic space graph, including: ； ； ； in, Representation layer regularization, Represents a linear layer. Indicates use and The constructed dynamic graph, Represents the average function, Represents the normalization function. This represents the processed image. This means taking the largest value in the graph. A function of values, Represents a dynamic spatial graph.

2. The traffic flow prediction method based on adaptive dynamic fusion graph convolutional network according to claim 1, characterized in that, The step of inputting the collected traffic sequences and dynamic spatiotemporal fusion map into the spatiotemporal block to enhance nonlinear dependencies and dynamic trends includes: Within each spatiotemporal block, a gated temporal convolutional network is used to enhance nonlinear dependencies in the traffic sequence, and then combined with a dynamic spatiotemporal fusion graph through a diffusion graph convolutional network to enhance dynamic trends.

3. The traffic flow prediction method based on adaptive dynamic fusion graph convolutional network according to claim 1, characterized in that, The intrinsic semantic enhancement module performs the following operations: Daily and weekly embeddings are used to enhance the temporal periodicity of traffic sequences, and linear layers are employed to capture hidden temporal features. The spatial directionality of traffic sequences is captured using diffusion graph convolutional networks and dynamic temporal graphs. Gated temporal convolutional networks are used to change the dimensionality of the input data.

4. The traffic flow prediction method based on adaptive dynamic fusion graph convolutional network according to claim 1, characterized in that, The output module performs the following operations: The spatiotemporal features output by the spatiotemporal block are truncated in time length, and then skip connections are performed. The features after skip connections and the semantically enhanced traffic sequence are input into a multilayer perceptron to obtain the prediction results.

Images