With the rapid growth of urban trafic, the importance of trafic flow prediction has increased. However, existing models still have significant limitations in capturing complex spatio-temporal dependencies: 1) The data collected by sensors is often incomplete, which can seriously afect the accuracy and reliability of trafic flow prediction. However, most existing models for processing missing values focus only on temporal dependencies, ignoring the complexity of spatio-temporal correlations in trafic data; 2) Relying only on predefined graph structures based on a priori knowledge or adaptive graph structures generated by node embedding often makes it dificult to accurately reflect the complexity and temporal variability of trafic flow data. However, most current models usually adopt only a single graph structure, fail to fully exploit the complementary advantages of multiple graph structures, and thus have certain limitations in capturing spatio-temporal features. To this end, this paper proposes a generative adversarial dual graph network for trafic flow prediction, viz, GADGN. In this paper, we use the GAN-TCT (Generative Adversarial Network-Transformer-CNN Fusion Module) module to achieve efective missing value filling. The generator of GAN-TCT combines a long short-term memory network (LSTM), a convolutional neural network (CNN) and a transformer to efectively extract the time-series features and spatial correlations in the input data, thus enhancing the realism of the generated data. In addition, this paper designs the dual graph convolutional fusion coding mechanism (DSGCE), which simultaneously considers the information contained in the node features and the adjacency matrix, generates an adaptive graph structure using node embeddings, constructs a predefined graph structure based on a priori knowledge through the adjacency matrix, and extracts the features contained in the two graph structures by convolution of the two graphs, and then extracts a more comprehensive graph structure from the two graph structures through the graph features learned by the self-encoder. It then extracts a more comprehensive graph structure from the two graph structures using the graph features learned by the self-encoder, and then extracts a more comprehensive spatio-temporal information from the two graph structures using the graph features learned by the self-encoder, and then combines this information to generate a new feature representation containing more potential patterns and dependencies, and ifnally captures the temporal features through a gating mechanism to obtain the final prediction results. Extensive experiments on real road datasets PeMSD4 and PeMSD8 have shown that the GADGN model reduces the MAE by about 3.5%, the RMSE by about 1.5% and the MAPE by about 1.8% compared to the existing state-of-the-art method (SOTA). These results fully demonstrate the superiority and efectiveness of the GADGN model in trafic lfow prediction tasks.
With the acceleration of urbanisation, the transport system is facing serious problems such as congestion and uneven public transport services. Trafic flow prediction has become a key issue in urban trafic management. Most of the traditional trafic flow prediction methods rely on cloud-based data centres for computation, but with the rapid development of sensor networks and the application of edge computing, more and more real-time data needs to be processed on edge devices. Edge computing, as a distributed computing architecture, can efectively reduce latency and improve data processing eficiency by distributing computing tasks to edge nodes close to the data source. In this context, this paper proposes a trafic flow prediction model (GADGN) based on generative adversarial network and dual graph convolutional architecture. The model combines the GAN-TCT module with the dual-graph convolutional fusion mechanism and is optimised for eficient data gap filling and trafic flow prediction in an edge computing environment.
Missing values in the data not only weaken the validity of the data, but also directly afect the reliability of model training, leading to inaccurate prediction results. Traditional filling methods, such as mean-filling, linear interpolation or forward filling, although simple and easy to use, often fail to capture the true distribution of the data. In addition, the relationships between nodes in urban trafic networks change dynamically, and trafic flows vary significantly depending on time, location and environment. Traditional models that rely only on static modelling are dificult to adapt to complex changes in trafic flows. Static neighbourhood matrices reflect the fixed connectivity of nodes in a transport network, and understanding these relationships is critical to capturing the underlying trafic patterns. Relying solely on dynamic modelling can result in models that do not fully capture changing trafic flow patterns. Therefore, the construction of efective trafic flow prediction models must focus on both static and dynamic spatio-temporal features to better adapt to complex trafic flow changes.
To solve these problems, this paper proposes a method that combines the GAN-TCT module with the dual graph convolutional fusion mechanism (GADGN) to improve the accuracy and robustness of trafic lfow prediction. In the missing value filling session, we design the GAN-TCT module to eficiently generate pseudo-data that is highly similar to the real trafic flow to fill the missing parts of the data. The generator extracts time-series features related to missing values using a Long Short-Term Memory (LSTM) network, a Convolutional Neural Network (CNN) captures local spatial dependencies, and global context modelling is performed by the Transformer module. The structure not only integrates local and global features, but also efectively enhances the spatio-temporal consistency of the data through the feature fusion layer. In addition, this paper designs a dual-image convolutional fusion coding mechanism, which adopts a dual-image convolutional structure to fully capture the multi-scale features of both static and dynamic images, thus enhancing the feature extraction capability of the model. Specifically, static graph convolution uses a predefined graph structure to extract fixed neighbourhood features, while dynamic graph convolution dynamically generates time-varying neighbourhood matrices through node embedding, efectively characterising dynamic spatio-temporal dependencies. However, since static and dynamic graph convolution may only reflect local relationships when fused, and may not adequately model the deeper features in the data, we consider introducing an adaptive feature reconstruction module (self-encoder), which is capable of extracting complex spatio-temporal features through a hidden space representation, and helps to better represent the global and non-linear properties of the data. This mechanism efectively combines the a priori information of the fixed topology with the flexibility of the time-varying topology, enabling the model to better capture the long- and short-term dependencies and multi-scale features in the data when dealing with complex spatio-temporal data (e.g. trafic flow prediction), thus improving the overall prediction performance.).
In the field of trafic flow prediction, researchers have developed a variety of deep learning models
to improve prediction accuracy and capture complex spatio-temporal dependencies, each proposing
innovative solutions to specific problems. For example, Graph WaveNet [
To address the above issues, this paper proposes a Dual Graph Convolutional Fusion Coding (DSGCE) mechanism that comprehensively captures both fixed relationships and dynamically changing features in trafic data by combining an adaptive graph structure and a predefined adjacency matrix. The mechanism enhances the adaptability to dynamic changes in spatio-temporal dependence while maintaining the model’s sensitivity to long-term stable features, providing a more comprehensive and eficient solution to the complex task of trafic flow prediction.
In the field of missing value processing, traditional methods such as mean-filling and k-nearest neighbour
algorithms [
To address the above shortcomings, this paper proposes a Generative Adversarial NetworkTransformer-CNN Fusion Module (GAN-TCT), which aims to comprehensively model the temporal variability characteristics and spatial dependence of trafic flow data. This multi-model fusion design efectively integrates local and global features, and achieves high robustness and accuracy in dynamic and complex trafic flow data.The GAN-TCT module not only provides an innovative method to fill in missing values, but also lays a solid foundation for improving the performance of trafic flow prediction.
The application of edge computing to trafic flow prediction is also gaining attention. In recent
years, many studies have attempted to combine edge computing techniques with trafic flow prediction,
aiming at real-time data processing and prediction using edge computing devices. For example, Yu et al.
[
The architecture of our proposed GADGN model is shown in figure 1, and consists of two main modules: the missing value filling module and the trafic prediction module. In the missing value filling module, we design the GAN-TCT module, which is diferent from the traditional GAN.The GAN-TCT module combines a Convolutional Neural Network (CNN) to extract local spatial features, a Long Short-Term Memory Network (LSTM) to capture time-series dependencies, and a Transformer module to model global spatio-temporal contextual information. Based on this structure, the Generative Adversarial Network (GAN) generates missing value filling results that are very close to the real data through adversarial training, thus achieving an important breakthrough in filling quality and spatio-temporal consistency, ensuring that the generator can simultaneously account for spatio-temporal dependencies and generate more realistic filling data. The filled data is divided into q time steps, and the data from each time step is used to generate a predefined knowledge-based adjacency matrix and a dynamic adjacency matrix based on node embeddings (without prior knowledge). These two adjacency matrices are then fed into two graph convolution modules to extract the spatial dependencies in the dynamic graph structure and the predefined knowledge-based graph structure, respectively. By combining the features from these two graph convolutions and feeding them into the self-encoder for feature learning, we are able to extract richer spatial feature information and complete comprehensive spatial feature extraction. Finally, by replacing the traditional update gate and reset gate with the extracted spatial features, we compute the gate control parameters to better capture the temporal dependencies in the data and obtain the final prediction results. The design not only improves the model’s ability to model complex spatio-temporal dependencies, but also significantly improves the accuracy of the prediction.
Missing value handling is a key issue in spatio-temporal data analysis. This study proposes a new GAN-TCT module for missing value filling in trafic flow data. The module combines the generator and discriminator of Generative Adversarial Networks (GAN) while introducing a spatio-temporal feature extraction mechanism. The generator part consists of LSTM, CNN and Transformer architectures, which can efectively capture the temporal dependence and spatial relationship of the data. The discriminator part adopts MLP (Multi-Layer Perceptron), which discriminates the diference between the generated data and the real data through multiple fully connected layers to ensure that the generated data has high authenticity and consistency. The generator and discriminator optimise each other through adversarial training, improving the quality of the filling results. The model not only fills in missing values, but also captures more accurate spatio-temporal dependencies when dealing with complex spatio-temporal data, thus generating more realistic and accurate missing value filled data.
Specifically, the generator first captures long-term dependencies in the time series using LSTM, and then extracts local spatial features using CNN. The CNN-processed features are then fed into the Transformer encoder, which models the global spatio-temporal context information through the selfattention mechanism, further enhancing the ability to model complex spatio-temporal dependencies. To obtain global features, the generator also uses a global pooling layer that fuses the global representation of spatial features with the spatio-temporal features output by the Transformer to produce more realistic data filled with missing values. The generator is formulated as follows: = () (1)
The LSTM process the output lstm, where x is the input data and lstm is subjected to 1D convolution to extract local spatial features: o = Re ( ()) (2)
= ( , , dim = − 1)
= − () where represents generated data. And the structure of the discriminator is based on several fully connected layers and its task is to distinguish between real and generated data. Its goal is to
The features output from the convolutional layer are mapped to the input dimensions of the Transformer by a Fully Connected Layer (FC):
= ( ) The transformer input is processed by the encoder to extract global spatio-temporal features: = ()
To further enhance the features, the generator extracts global features from the output of the convolutional layer (by average pooling) and passes them to a fully connected layer: = − (( , dim = 1))
Finally, the generator splices the features that come out of the transformer with the global features and fuses them together through a fully connected layer: (3) (4) (5) (6) (7) maximise the discrimination between real data and generated data . The loss function of the discriminator is defined as: = − [log ()] − [log(1 − ()] where (∙ ) is the discriminator output.
In addition, we introduce the Kullback-Leibler Dispersion (KLD) as a measure of the diference between the distribution of the generated data and that of the real data: (‖) = ∑︁ () log () () where and represent the probability distribution of real and generated data respectively.
The total loss function of the generator is: = − ⌊︀ log ( )︀⌋ + ∙ (︀ ‖ ︀) where the first term is the loss of the generator to fool the discriminator, the second term is the KLD between the distribution of the generated data and the distribution of the real data, and is the hyperparameter that controls the trade-of between the two.
During the training process, the generator and the discriminator improve each other through adversarial training, where the generator is continuously optimised to generate more realistic complementary missing value data, while the discriminator gradually improves its ability to discriminate between generated and real data. In addition, we design an adaptive data generation mechanism that enables GAN-TCT to dynamically adjust the generation process according to the spatio-temporal dependence of the data, in order to better capture patterns in complex trafic flow data and produce high-quality complementary results. This adaptive mechanism significantly reduces the filling error and improves the overall performance of the trafic flow prediction model.
The trafic prediction module is the core part of this paper, and the accuracy of trafic prediction is crucial for trafic management. In order to improve the prediction accuracy, this paper proposes a dual graph convolutional fusion (DSGCE) coding mechanism, which not only efectively captures the spatial dependence in the data, but also dynamically adjusts the gating parameters by the spatial features of the self-encoder output to capture the temporal dependence to achieve more accurate trafic prediction. Specifically, this paper uses graph convolution based on the Laplace operator to capture the time-varying relationships in the dynamic neighbourhood matrix that reflect the instantaneous changes in trafic flow. Meanwhile, Chebyshev polynomial-based graph convolution is used to process the predefined adjacency matrix to model the stable relationships between nodes, thus capturing the long-term dependence of trafic flow. On this basis, this paper extracts and learns the spatial features using a self-encoder to obtain a feature representation with comprehensive spatial information. Based on the spatial features output from the self-encoder, the model dynamically adjusts the gating parameters to better capture the time dependence in the trafic flow data. By combining the spatial features obtained by the dual-image convolutional fusion coding mechanism with the temporal features adjusted by the gating mechanism, the model is able to more accurately predict the changes in the spatio-temporal characteristics of the trafic flow.
We divide the filled data combined with temporal information into q time segments, and the input data x of each time segment is the combination of the input data of the time step and the output extraction, which maps the learned dynamic features to the feature matrix : features H− 1 of the previous time step, and pass x to the Multilayer Perceptron (MLP) layer for feature = (x)
Finally, the resulting dynamic graph is fed into the Laplace graph convolution for adaptive graph feature extraction:
D− 21 AD− 21 = D− 21 (︁ Re LU ︁( GGA )︁ i where D is the diagonal matrix of time step i , whose diagonal elements are the degrees of each node. The dynamic graph convolution operation is performed by constructing a normalised Laplace matrix to generate the output feature V:
V = ︂(
I + D− 21 AD− 12 )︂
X + represents the number of nodes of the matrix, D is the trafic feature captured by the sensor, and where I is the unit matrix, X ∈ R× is the input feature for dynamic graph convolution, N and term = ∙ , which is obtained by combining the above two formulas: b denote the weight and bias, respectively. And in this paper, the weight and bias matrices are also done an update operation through the node matrix to get the weight tensor W = ∙ and the bias V = ︂(
I + D− 21 (︁ Re LU )︁ ︁(
D− 21 )︂
XW + combined with temporal information:
Similarly, our predefined-based adjacency matrix is generated from the original adjacency matrix
The obtained dynamic feature matrix is combined with the node information matrix containing time information to generate the dynamic graph G for that time period. The first filled node embedding matrix and the time information are subjected to an element-by-element product operation to obtain the node information matrix with periodicity: where denotes the daily embedding and denotes the weekly embedding. Secondly, the correthe dynamic feature matrix and the periodic node matrix for that time period: sponding dynamic graph G is obtained by a further element-by-element multiplication operation of = ⊗ ⊗
G = tanh (︀ ⊗ )︀
The generated pre-defined neighbourhood based matrix is fed into the Chebyshev polynomial based graph convolution for feature extraction operation and finally outputs the pre-defined graph feature E:
= ⊗ ⊗ = 2− 1 − − 2, ≥ 2
E = ∑︁ − 1 =0 X input features of the Chebyshev graph convolution. where 0 and 1 are the initial polynomials, usually taken as 0 = I and 1 = , i.e. the unit matrix I and the adjacency matrix of the predefined graph. In this paper, = 2 and X are the
Next, the two graph features are combined and fed to the autocoder to further learn the spatial features and obtain a more comprehensive feature representation R. Specifically, the coder compresses the input features into the potential space Z via a two-layer linear transformation (weight matrices 1 and 2) and an activation function, and the decoder starts from the potential space Z and reduces the feature matrices by using the two-layer linear transformation (weight matrices 1 and 2) and activation functions to increase the similarity of the reconstructed features to the original (12) (13) (14) (15) (16) (17) (18) (19) input, reduces the feature matrix, and finally adds a residual join term of the reconstructed features to the original input: Rcorv to increase the similarity where denotes sigmoid activation, W , W , W , , , denote learnable parameters, [∙ ] denotes splicing operation, R denotes reset and update gates, is an intermediate temporary variable, and H denotes the spatio-temporal characteristics of the output of the gating mechanism at the i -th time step.
The spatio-temporal dependent information captured at each time step is output through the gating mechanism. We take the gated feature output of the q -th time step as the final result, i.e., H , and further process it through the convolutional layer to produce the final prediction result yˆ: yˆ = (H)
The loss function plays a crucial role in the training process and is used to evaluate the performance of the model during the process. By minimising the loss function, the model can appropriately reduce the diference between the predicted values and the true labels, thus improving the accuracy of the prediction. We use L1 loss as the loss function to train the whole model. Where K is the historical time step, T is the future time step and is the absolute value of the diference between the true value Y+ and the predicted value Yˆ +: = 1 ∑=︁1 ⃒⃒⃒ Yˆ + − Y+⃒⃒⃒
= +
Finally, we use the spatial features of the self-encoder output to compute gating parameters to dynamically extract temporal features from the data. The specific gating mechanism can be expressed as follows:
Rconv = V + E
Z = (2 · ( (1 · Rconv))) R = (2 · ( (1 · Z))) + Rconv = ([, H− 1] RW + ) = ([, H− 1] RW + ) = tanh ([, H− 1 ⊙ ] RW + )
H = ⊙ H− 1 + (1 − ) ⊙
Finally, we combine the total loss function of the generator with the loss values trained on the whole model to obtain the total loss: where is a weighting parameter to balance the contribution of the two loss components.
To evaluate the performance of the GADGN proposed in this paper, we use two publicly available real trafic flow datasets from the PeMS data: the PeMSD4 and PeMSD8. The PeMS data are collected in real time from more than 40,000 detectors in California, USA.
Specific dataset information is provided below: • PeMSD4: This dataset consists of 307 detectors collecting data at 5-minute intervals for a total of 59 days of trafic flow data. The data was collected between January and February 2018. • PeMSD8: This trafic dataset consists of 170 detectors collecting data at 5-minute intervals for a total of 62 days of data, collected between July and August 2016.
Each collection of both datasets contains features in three dimensions: volume, average speed and average occupancy.
We ran our experiments on a Win10 PC with an NVIDIA GeForce RTX 2070 graphics card. Specifically, our data are divided in a ratio of 6:2:2, where 60% of the data set is used as the training set, 20% of the data set is used as the validation set, and the remaining 20% is the test set. Regarding the hyperparameters of the GADGN model, we set the hidden size to 64, the kth-order Chebyshev polynomial expansion to 2, the embedding dimension D of PEMSD4 to 10, and the embedding dimension D of PEMSD8 to 5. The model was trained using the Adam optimiser and the training time was set to 100. Due to the combination of several modules such as biplot convolution, generative adversarial and self-encoder, the model has a high memory requirement, the batch size of the PEMSD4 dataset was set to 8 and the learning rate to 0.00075 for the PEMSD4 dataset, and the batch size was set to 24 and the learning rate to 0.0015 for the PEMSD8 dataset. The MAE was used as the loss function for training the model. Three evaluation metrics were used to assess performance: (1) mean absolute error (MAE), (2) root mean square error (RMSE) and (3) mean absolute percentage error (MAPE). (1) Mean absolute error: (2) Root mean square error: (3) Mean absolute percentage error: =
=1 1 ∑︁ ⃒ ⃒ ⃒⃒ Y+ − Yˆ +⃒⃒ ⎯ = ⎷⎸⎸ 1 ∑︁ (︁ Y+ − Yˆ +︁) 2
=1 = 100% ∑i=︁1 ⃒⃒⃒⃒⃒ Ŷ︀+Y −+Y + ⃒⃒⃒⃒ ⃒ (30) (31) (32)
We compare the proposed GADGN with the following 10 representative trafic forecasting methods:
• ARIMA [
MAE
A comparison of the results of our model (GADGN) with the other 10 benchmark models on the PeMSD4 and PeMSD8 datasets is shown in table 1. All models are experimented on the same dataset
MAPE 24.18% 17.24% 14.70% 14.18% 13.90% 17.29% 16.77% 12.97% 12.89% 12.36% 12.14%
MAE 31.09 19.19 19.00 18.25 17.13 18.28 16.94 15.95 15.50 15.23 14.27
RMSE 44.32 29.81 28.70 28.06 26.80 30.05 26.25 25.22 24.51 24.35 23.67
MAPE 22.73% 13.10% 11.32% 11.64% 10.96% 12.15% 10.60% 10.09% 9.95% 9.83% 9.38% and their performance is evaluated over 12 time steps. It can be seen that GADGN shows excellent performance in all three metrics, MAE, RMSE and MAPE, outperforming all the benchmark models. • Comparison with traditional methods: Traditional time series models such as ARIMA and VAR have significant advantages in modelling temporal correlation. However, they are limited in that they do not account for spatial heterogeneity in the transport network, making it dificult to deal with spatial dependencies in complex transport systems. This leads to poor performance in modelling non-linear features, especially on the PEMS04 and PEMS08 datasets, which have a large number of nodes and high data complexity, and significantly poor prediction performance. In contrast, our model GADGN is able to comprehensively capture spatio-temporal features by introducing a spatio-temporal dependency-based missing value filling method and a dualplot convolutional fusion coding mechanism, while retaining the ability to model temporal dependencies, which significantly improves the prediction accuracy and stability. • Comparison with fixed neighbourhood matrix: Models such as STGCN, which are based on a ifxed adjacency matrix, are able to capture spatial dependencies in trafic flows using a predefined spatial structure. However, the limitation of such models lies in the static nature of the adjacency matrix, which is unable to adapt to the dynamically changing characteristics of the trafic network and thus lacks flexibility in dealing with complex spatio-temporal relationships, limiting their predictive performance. Our GADGN model innovatively combines an adaptive adjacency matrix and a self-encoder, which can not only flexibly capture the static spatial relationships, but also dynamically adjust the spatial dependencies according to diferent time steps, comprehensively expressing the global and non-linear characteristics of the data, thus improving the sensitivity to trafic flow changes and the prediction accuracy. • Compared to Adaptive Neighborhood Matrix: Models such as Graph WaveNet and AGCRN are able to dynamically adjust the adjacency matrix by learning the spatial correlations in the data, thus outperforming models based on fixed adjacency matrices in capturing dynamic changes in the trafic network. However, these models rely too heavily on adaptive adjacency matrices, often ignoring the a priori knowledge already present in the trafic network and making it dificult to fully exploit fixed spatial relationships. Our GADGN model innovatively combines the adaptive adjacency matrix with a predefined graph structure and uses the Dual Graph Convolutional Fusion Coding (DSGCE) mechanism to achieve eficient fusion of static and dynamic spatial relationships. Through this mechanism, the model not only flexibly adjusts the adjacency matrix to adapt to dynamic changes, but also fully exploits the a priori knowledge to comprehensively capture the multi-scale spatio-temporal features in the trafic flow, which further improves the prediction performance.
To more intuitively demonstrate the performance diference between GADGN and other models, we visualise the above tabular data as shown in figure 2.
To evaluate the performance of our proposed GADGN model, we conducted a series of ablation experiments using the PEMS04 and PEMS08 datasets. By removing diferent components of the model, we analysed the contribution of each module to the overall performance of the model with the following experimental design: • w/o Trans: Removal of the Transformer module from the GAN-TCT framework to assess its impact on the accuracy of missing value filling. • w/o CNN: Removal of Convolutional Neural Networks (CNNs) from the GAN-TCT module to evaluate the importance of CNNs in capturing spatially dependent features during missing value iflling. • w/o PG: Removing the neighbourhood matrix based on predefined knowledge and using only the dynamic neighbourhood matrix generated based on node embedding to assess the importance of a priori knowledge in the construction of the graph structure. • w/o AE: The self-encoder module is removed and the output features of the biplot convolution are passed directly to the following module without going through the learning and fusion process of the self-encoder.
The results of the experiments are summarised in table 2. As can be seen from the table, the results of the ablation experiments are inferior to the GADGN model for both the PEMSD4 and PEMSD8 datasets. In particular, the results of the w/o Trans experiments show that the removal of Transformer significantly reduces the model’s ability to capture global spatio-temporal dependent features, which further validates the important role of Transformer in the missing value filling task for global feature extraction. The w/o CNN experiments show that the removal of Convolutional Neural Network (CNN) significantly reduces the model’s ability to capture global spatio-temporal dependent features, which fully illustrates the key role of CNN in learning spatial dependencies in the GAN-TCT module. The w/o PG experiments show that using only adaptive adjacency matrices significantly weakens the model’s ability to capture complex spatio-temporal dependencies, leading to a decrease in prediction accuracy, highlighting the importance of predefined knowledge-based graph structures in modelling complex spatial dependencies. Finally, the w/o AE experimental results show that the removal of the self-encoder mechanism significantly degrades the model performance, further confirming the indispensability of the self-encoder in fusing and reconstructing graph convolutional features, which efectively enhances the model’s ability to capture and express spatial dependency features. In summary, the design and implementation of the individual components contribute significantly to the overall performance of the GADGN model, further demonstrating its efectiveness in spatio-temporal data prediction tasks.
The GADGN model proposed in this paper significantly improves the missing value filling and prediction ability of spatio-temporal data by introducing the GAN-TCT module and the dual graph convolutional fusion encoding mechanism (DSGCE). The missing value filling module adopts the Generative Adversarial Network (GAN), which combines the Long Short Term Memory Network (LSTM), Convolutional Neural Network (CNN) and Transformer to accurately capture spatio-temporal dependencies and produce filling results that are closer to the real data. The feature extraction module efectively integrates static and dynamic adjacency matrices through the synergistic action of the dual-image convolutional fusion coding mechanism and the gating mechanism, which not only strengthens the learning ability of complex spatial dependencies, but also dynamically captures temporal dependencies, thus improving
GADGN w/o Trans w/o CNN w/o PG w/o AE
GADGN w/o Trans w/o CNN w/o PG w/o AE
the prediction performance of the model.
Despite the excellent results achieved by the GADGN model in centralised computing architectures, its application in edge computing environments still faces challenges in terms of computational resources and real-time performance, especially in terms of inference speed and computational eficiency, which need to be further optimised. In addition, although the dual-image convolutional fusion coding mechanism can efectively capture spatio-temporal dependencies, the model may have dificulty dealing with anomalies caused by extreme events such as unexpected accidents or severe weather in certain complex trafic networks, which may afect the accuracy of the prediction. Future research will focus on improving the computational eficiency and inference speed of the model on edge devices, and consider the introduction of external influences to promote widespread adaptation and extension of the model in practical applications.
Acknowledgments: This work was supported in part by the National Natural Science Foundation of China under Grants 62206116 and Chunhui Plan Collaborative Research Project, Ministry of Education, China (HZKY20220350)& Key R&D Plan of Hubei Provincial Department of Science and Technology 2023BCB041.
Declaration on Generative AI: The authors have not employed any generative AI tools.
MAE