Traffic flow prediction model for realizing co-mining of traffic point semantics and network topology

Abstract

The invention belongs to the technical field of traffic, and particularly relates to a traffic flow prediction model for realizing co-mining of traffic point semantics and network topology. The name of the model is STIGNN, the STIGNN focuses on coupling of topological relations between traffic point features and traffic points in the aspect of spatial feature aggregation, and potential and existing spatial dependencies are comprehensively considered. The STIGNN can process a long-time sequence through a time convolution layer based on gating, and the model can be directly applied to an inductive learning task by introducing a dual attention mechanism of node features and a topological relation into a space-time framework, and can be popularized to any network with complex entities and relations. Experimental results on two public real traffic network data sets METR-LA and PEMS-BAY show that the STIGNN is superior to an advanced baseline model.

Description

Technical Field

[0001] The invention belongs to the field of traffic technology, and in particular relates to a traffic flow prediction model for realizing co-mining of traffic point semantics and network topology. Background Art

[0002] Traffic flow prediction is a key task in urban traffic control and dispatching in the field of transportation, which is highly nonlinear and complex. The transportation system is an important infrastructure of modern cities, supporting the daily communication and travel of millions of people. With the rapid development of urbanization and population, the transportation system has become more and more complex, and frequent traffic jams, traffic accidents and other safety issues have emerged. Intelligent transportation systems have emerged and developed rapidly, aiming to achieve efficient coordination of elements such as people, vehicles and roads by making full use of cutting-edge technologies such as artificial intelligence, the Internet of Things and mobile communications. Traffic flow prediction is a core component of the intelligent transportation system. Its task is to use given historical speed data and basic road networks to predict future traffic flow speeds, effectively avoid traffic congestion, and improve the operating efficiency of the transportation system. However, traffic flow prediction is a complex spatiotemporal prediction problem. The traffic status of a specific area is not only affected by the distribution of surrounding traffic routes, but also changes dynamically according to different time periods. The spatial and temporal dependence makes traffic flow prediction face huge challenges.

[0003] Researchers have proposed many algorithm models to solve the problem of traffic flow prediction. Traditional statistical models such as historical average (HA) model, auto-regressive integrated moving average (ARIMA) model, Kalman filter model, etc., are limited by the reliance on the stationary assumption and cannot capture the complex dynamic relationship between traffic data, and their prediction ability is very limited. Traditional machine learning algorithms such as support vector regression (SVR) and linear regression (LR) have been further improved on the basis of traditional statistical models by mining historical traffic data, but they still cannot break through the complexity of the model in time and space. Inspired by the great success of deep learning technology in the fields of computer vision and natural language processing, researchers introduced recurrent neural network (RNN) models that can mine time dependencies and convolutional neural networks (CNN) that can capture spatial dependencies into traffic prediction, and the traffic flow prediction task has achieved unprecedented success. During this period, excellent traffic flow prediction models such as LSTMNN and GBM emerged. However, CNN and RNN can only be used for Euclidean data (i.e., images, text, and videos), and the spatial dependency of the road network is represented by a non-Euclidean graph structure. Serializing the graph structure will inevitably lead to information loss, which is not conducive to the mining of spatial topological information.

[0004] In recent years, graph neural networks (GNNs) have become the forefront of deep learning research and have shown state-of-the-art performance in various applications. It can capture complex relationships between objects and reason based on data described by graphs. GNNs have been proven to be effective in various node-level, edge-level, and graph-level prediction tasks. In fact, GNNs are very suitable for traffic prediction problems. The road network is essentially a graph, the sensors on the road are nodes, and the connections between sensors are edges. With the graph as input, benefiting from the completion of the topological information of the traffic network, the GNN-based model will show superior performance than previous methods in the road traffic flow prediction task. The STGCN model first applied the GNN network to traffic network prediction. By using the graph convolutional neural network (GCN) to mine the topological information of the traffic network, and using the gated temporal convolutional layer (Gated TCN) module based on the gated mechanism instead of the RNN model to capture the time period characteristics, the model achieved excellent prediction results in the METR-LA and PEMS-BAY datasets. After that, researchers focused their attention on the migration of graph neural network algorithms to traffic prediction, and a large number of traffic prediction models based on the combination of GNN and RNN emerged. For example, the DCRNN model using diffusion convolutional neural networks (DCNNs) and gated recurrent unit (GRU), the STGAT model using graph attation neural network (GAT) and Gated TCN modules, and the GC-LSTM model using GCN and long short-term memory (LSTM). It can be clearly observed that the current traffic flow prediction model based on spatiotemporal graph structure focuses on the optimization of graph neural network and recurrent neural network modules, and different combinations can always bring surprises to traffic flow prediction. Of course, many strategies have also been introduced during this period, such as the learnable adjacency matrix introduced in the GraphWaveNet model. The road structure information mined by researchers from the map is subject to the granularity of manual processing, and some key network information will still be ignored. By allowing the network to adaptively learn graph structure information and make joint predictions with known graph structures, the problem of missing graph structures can be effectively solved. Graph WaveNet is also based on GCN and Gated TCN modules, but it has achieved higher prediction accuracy than DCRNN. The learnable matrix strategy also appears in the STGAT model, and experiments again prove the feasibility of this strategy.

[0005] GNN has made significant progress by considering the underlying graph structure. However, there are still some challenges; for example, the left and right lanes as nodes are close, but their correlation is not necessarily stronger than the same-direction lanes at a greater distance. For another example, traffic sections with similar infrastructure will have reference value even if they are far away. It is hoped that the road's own information and the road network topology information can be further coupled, but there is still a certain gap in the joint mining of topological information and node semantic information in the graph neural network. Whether it is the GCN model, the GAT model, the DCNN model, etc., they are all node-level information aggregation methods. The target node can only mine the second-order neighbor information through the first-order neighbor, and it is impossible to fully mine the network topology information. Although the SGC model and ChebNet model in the model can mine multi-order topological information at the same time, their mining of semantic associations between nodes is not in place, and they only formulate graph information aggregation strategies based on the graph structure. In addition, only a small number of models have learnable information aggregation rules, and the model cannot filter which information is important and which is not important by itself. In the model where aggregation rules can be learned, GAT adjusts the information aggregation method between nodes by calculating the dynamic attention between nodes, but it is limited to the first-order graph structure, and the information dependency mining is more based on the semantic information of the node itself, and the relationship between neighbors of different orders in the network cannot be well mined. ChebNet applies static attention to the node sets of neighbors of different orders, and judges which order of neighbor sets of each node is more important through learning, but the learning parameters are set in the network topology, and the semantic association of nodes cannot be fully extracted. The fusion of the two models may help find a traffic flow prediction solution that simultaneously mines the topological information of the traffic network and the semantic information of the traffic nodes.

[0006] In order to better explore the spatiotemporal relationship of the traffic network and improve the accuracy of traffic flow prediction, the present invention proposes a spatiotemporal graph neural network that combines the semantic information of traffic nodes and the network topology information. Summary of the invention

[0007] The present invention proposes a traffic flow prediction model that realizes the co-mining of traffic point semantics and network topology. The model solves the problem that the existing methods ignore the importance of coupling traffic point characteristics with traffic network information for traffic flow prediction, and realizes effective traffic flow prediction tasks.

[0008] The technical solution of the present invention is achieved in this way: A traffic flow prediction model that realizes the co-mining of traffic point semantics and network topology is named STIGNN. The STIGNN model includes two parallel computing ChebGAT modules and three-layer serially connected GateTCN modules. The ChebGAT module couples the ChebNet static attention and the GAT dynamic attention to fully mine the spatial correlation of the traffic network and realize the simultaneous mining of topological features and semantic information. The GateTCN module integrates the time series model of the gating mechanism to mine the temporal correlation of road information detected at different times, and then obtains the road features of the fused time series information. The road features of the fused time series and the sensor network are then input into the ChebGAT module to deeply mine the spatiotemporal characteristics of the road network for the final traffic flow prediction.

[0009] Optionally, the ChebNet model aggregates features as follows: , in, represents a nonlinear activation function, It is The learnable static weights of the neighbors of order , When the value is expressed, it indicates the importance of the information itself. Representation Node The degree, is a learnable weight matrix, Representation Node The traffic point information of the first-order neighbors, Representation Node Traffic point information of the second-order neighbors, Representation Node Traffic point information of the third-order neighbors, Representation Node The degree, Representation Node The degree.

[0010] Optionally, GAT performs the following feature aggregation process: , in, Representation Node and The degree of semantic relevance of , in, Represents a splicing operation, is a learnable weight vector, Representation Node The first-order neighbor set of .

[0011] Optionally, the neighborhood perception range of the ChebGAT module is fixed to 3rd order, and the process of aggregating features is as follows: .

[0012] Optionally, the STIGNN model transposes the sensor feature matrix , while setting the dilated causal convolution kernel ,in, Represents the size of the convolution kernel, and the process of dilated causal convolution can be expressed as: , in, is the dilation factor of the dilated causal convolution, Indicates the time series length of the intercepted traffic data, It is a traversal The numerical value of Indicates the use of dilated causal convolution kernels right Perform convolution, Represents the sth parameter of the convolution kernel f.

[0013] Optionally, the GateTCN module data feature processing process is as follows: , in, Indicates GateTCN No. The update gate of the layer is used to control the information ratio of longer time series and shorter time series. Represents the reset gate, To obtain long time series information, , and are all learnable convolution kernels. , and are all learnable feature biases, Represents the dot product.

[0014] Optionally, the ChebGAT module introduces an adaptively learned traffic graph while capturing spatial features. , The edges and edge weights are learned end-to-end using the stochastic gradient descent algorithm. The adjacency matrix of , using a randomly initialized matrix study : , in, Named as source node embedding, Name the target node embedding.

[0015] After adopting the above technical solution, the beneficial effects of the present invention are: The traffic flow prediction model in the present invention innovatively combines the key elements of the ChebNet model and the GAT model, so that the graph neural network can simultaneously calculate the feature associations between traffic points and the dependencies between adjacent traffic points of different orders. The traffic flow prediction model in the present invention is more comprehensive in mining the spatial information of the traffic network. At the same time, the addition of the GatedTCN module enables the model to have the ability to mine the temporal characteristics of traffic flows. The coupling of node semantic information and the topological structure of the traffic network allows the traffic points to be predicted to fully refer to the status of nearby traffic points and the structure of the traffic network, so that the traffic flow prediction model in the present invention is more comprehensive in mining the spatial information of the traffic network. In view of the dynamic temporal nature of the traffic flow network, the present invention designs a Gated TCN module, which adds a gating function to learn and control the dependencies of traffic information in different time periods, so that the present invention has the ability to predict traffic flow changes with time series. Through the combination of two-way mining of spatial and temporal information, the present invention can achieve more accurate traffic flow prediction compared to existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0017] Figure 1 It is the process of aggregating neighbor information between the GAT model and the ChebNet module; Figure 2 It is the structure of the RNN model; Figure 3 It is a dilated causal convolution with a convolution size of 2; Figure 4 This is the structural diagram of the GateTCN module; Figure 5 This is the overall architecture diagram of STIGNN. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The embodiment of the present application discloses a traffic flow prediction model that realizes co-mining of traffic point semantics and network topology.

[0020] Embodiment, according to Figures 1 to 5 As shown, a traffic flow prediction model that realizes co-mining of traffic point semantics and network topology.

[0021] 1. Related Work 1.1 Problem Modeling The goal of traffic forecasting is to The traffic flow observed by a sensor is used to predict the future traffic flow. Sensors are regarded as nodes, and sensors with close distance or upstream and downstream relationships are regarded as edges. A traffic network can be defined as an undirected graph ,in represents a node set, Denotes an edge set. Representation diagram The adjacency matrix of Indicates sensor and There is a connection between . Using the same sampling frequency, the time step of each interval is , the traffic data is sampled times. In each step, the graph There is a sensor information matrix. The sensor information matrix recorded each time is concatenated and recorded as the feature matrix ,in represents the dimension of the information recorded by the sensor, including traffic flow, vehicle speed, road occupancy, etc. Indicates The information recorded by sensors in the traffic data collected, Indicates The traffic flow prediction problem aims to learn a function , this function combines the graph information with the sensor status at the historical moment and the predicted traffic flow to infer the traffic flow at the next moment. The process is as follows: (1), In fact, according to the interval time step Traffic flow prediction can be divided into many types according to different For short-term forecasts, For the medium-term forecast, For long-term forecasting.

[0022] 1.2 Graph Neural Networks In order to mine the topological information of the graph, Scarselli et al. proposed the concept of graph neural network, which combines recursive neural network and random walk model based on Banach's fixed point theory to achieve the fusion of node features and topological information. Based on this, the field of graph neural network has produced models such as GraphSAGE, DCNNs, GAT, etc. These models intuitively set the dependency of the target node on the features of its neighboring nodes in the problem space, so they are called spatial domain models. Among them, the GAT model introduced the attention mechanism into the graph neural network for the first time, calculated the dependency between nodes through the features of adjacent nodes, and fully involved the node semantic information in the aggregation of network information. Its aggregation process is as follows: Figure 1 As shown on the left. The black line represents the edge, and the arrow represents the direction of information aggregation. Representation Node and The attention score is calculated as follows: (2), in, Represents a splicing operation, is a learnable weight vector, is the learnable weight matrix, Representation Node The first-order neighbor set of . You can clearly see the node Only the information aggregation within the first-order neighborhood can be controlled, and the information aggregation weight is controlled by the node semantic information.

[0023] Convolutional neural networks have achieved great success in the field of images, but due to the irregularity of graph structures, convolutional neural networks cannot be migrated to the graph field. Bruna et al. proposed the first generation of graph convolutional neural network Spectral CNN in combination with graph signal processing. By orthogonal diagonalizing the graph Laplacian matrix, the Fourier basis and frequency of the graph can be obtained and a complete graph Fourier domain (also called graph spectral domain) can be formed. The graph signal is mapped from the spatial domain to the spectral domain. The mapping of the graph signal in the graph Fourier basis is serialized data, which can be effectively captured by convolution. Based on this, models such as ChebNet, GCN, and GraphHeat have emerged in the field of graph neural networks. These models all achieve graph tasks by adjusting the graph frequency in the frequency domain and amplifying the effective graph signal. Among them, ChebNet first proposed that the convolution kernel can be fitted by a frequency polynomial, which reduces the resource consumption of convolution kernel learning. There is a direct correlation between the frequency polynomial function in the frequency domain and the polynomial function of the adjacency matrix in the spatial domain, which means that while adjusting the order of the frequency, it also adjusts the weights of neighbors of different orders. Its aggregation process is as follows Figure 1 As shown on the right. The black line represents the edge, and the arrow represents the direction of information aggregation. Indicates The weight of a neighbor node is a learnable static parameter.

[0024] The GAT model focuses on the semantics of nodes and adjusts the contribution of first-order neighbors to the target node through the semantics between nodes. ChebNet focuses on the order of neighbor nodes and sets different learnable parameters for neighbors of different orders. Both models have their own focus and have achieved excellent results.

[0025] 1.3 Timing Network The architecture of the RNN model is as follows Figure 2 As shown in the figure, it stores the hidden features of traffic information at historical moments through unique memory cells and uses them for prediction at the next moment. RNN has outstanding performance in dealing with time series modeling problems. In order to further optimize the RNN algorithm, Graves et al. proposed LSTM. LSTM adds a gate structure based on RNN. It uses the forget gate to decide which traffic information to discard, and the input gate to determine which new input information to pass to the next step to update the information stored in the old memory cell, and the output gate to predict the traffic flow at that moment. Chung et al. further proposed the GRU model based on LSTM. GRU only uses the update gate and reset gate, which is simpler than LSTM and saves more time to train.

[0026] However, the extended model based on RNN still faces the problem of requiring a lot of resources to train and cannot handle the prediction of longer sequences. Wu et al. used dilated causal convolution as the temporal convolution layer (TCN) to capture the timing characteristics of nodes. Dilated causal convolution obtains an exponentially growing receptive field by increasing the depth of the network. Unlike RNN-based methods, dilated causal convolution can correctly handle long time series in a non-recursive manner, which helps parallel computing and solves the problem of gradient explosion. The present invention maintains the causal order of time by filling zeros in the dilated causal convolution, so the prediction of the current time step only involves historical information. The convolution process of the dilated causal convolution is as follows: Figure 3 shown.

[0027] 2. STIGNN model 2.1 Spatial Correlation Mining Traffic networks are usually represented by graph structures. In order to fully explore the spatial correlation of traffic networks, the present invention simultaneously considers the distance between traffic points and the differences in road conditions, where topology refers to the traffic network structure and semantics refers to the information of traffic points. The present invention simultaneously mines both the traffic network structure and the information of traffic points. Graph neural networks have derived many algorithms for mining node information or topological structure information in graph data, but there are still deficiencies in the simultaneous mining of topological features and semantic information. In order to solve this problem, the present invention proposes a ChebGAT module that couples ChebNet static attention and GAT dynamic attention to fully explore the spatial correlation of traffic networks. The process of ChebNet aggregating features is as follows: (3), in, represents a nonlinear activation function, It is The learnable static weights of the neighbors of order , When the value is expressed, it indicates the importance of the information itself. Representation Node The degree is used for diagonal normalization to ensure that the node features of neighbors of different orders are on the same scale. is a learnable weight matrix, and a nonlinear activation function Combined with nonlinear transformation of nodes that have integrated network topology information, further features are extracted. Representation Node The traffic point information of the first-order neighbors, Representation Node Traffic point information of the second-order neighbors, Representation Node Traffic point information of the third-order neighbors, Representation Node The degree, Representation Node The aggregation feature process of the GAT model is as follows: (4), in, is calculated by formula (2), indicating that the node and The degree of semantic association. Combining formula (3) and formula (4) to obtain the ChebGAT module, considering that the distance of sensors after the third-order neighbors is too far and cannot form a strong correlation, the present invention fixes the neighborhood perception range of the ChebGAT module to the third order, and the process of aggregating features is as follows: (5), When two traffic sensors are close to each other but have very different monitoring data, or have similar detection data but are far apart, the two sensors cannot provide much effective information to each other. The coupling of node semantic information and traffic network topology allows the model to select the most suitable node for reference and fully capture the effective information in the traffic network.

[0028] The combination of formula (2) and formula (4) will increase the time complexity. The traditional GAT model first calculates the feature maps of all vertices. , and then calculate the attention coefficient. The subsequent feature aggregation stage is mainly weighted summation operation, which no longer involves high-complexity multiplication operations. Therefore, the time complexity of GAT is The ChebGAT module needs to calculate the attention scores of the target node and its first-order neighbors, second-order neighbors, and third-order neighbors. In the most complex case, the attention coefficients must be calculated between all nodes. Therefore, the time complexity of the ChebGAT module is , the complexity at this time mainly depends on the number of road sensors and the dimension of node features. However, due to the fact that there are not many main roads in a region, the time complexity brought by the number of sensors does not significantly affect the training time of the model. By comparing the prediction results of the ChebGAT module and the GAT model on the data sets METR-LA and PeMS-BAY, the present invention confirms the powerful ability of the ChebGAT module in spatial correlation mining.

[0029] 2.2 Time Series Correlation Mining Wu et al. have demonstrated the powerful performance of dilated causal convolution in mining time series correlation. The present invention decides to use this model as a basis to further develop a time series feature mining model. In order to facilitate the study of the time series correlation of the traffic network, the present invention transposes the sensor feature matrix , while setting the dilated causal convolution kernel ,in Represents the size of the convolution kernel, and the process of dilated causal convolution can be expressed as: (6), in, is the dilation factor of the dilated causal convolution, Indicates the time series length of the intercepted traffic data, It is a traversal The numerical value of Indicates the use of dilated causal convolution kernels right Perform convolution, Represents the sth parameter of the convolution kernel f. The dilation factor increases with the stacking order of the dilated causal convolutional layers, which enables the dilated causal convolutional network to capture longer sequences with fewer layers, thus saving computational resources.

[0030] In RNN, the gating mechanism is crucial. Chung et al. have experimentally demonstrated the powerful ability of the gating mechanism in controlling time flow information. The present invention hopes to introduce update gates and reset gates in the dilated causal convolutional layer, so that the model can further establish long-term dependence of time correlation in the exponentially growing time receptive field. The present invention names the time series model that integrates the gating mechanism as the gated temporal convolutional layer, denoted as the GateTCN module. The structure of the GateTCN module is as follows: Figure 4 As shown, the data feature processing process is as follows: (7), in, Indicates GateTCN No. The update gate of the layer is used to control the information ratio of longer time series and shorter time series. Represents the reset gate, which determines which short time series information is retained for the calculation of long sequence information. To obtain long time series information, , and are all learnable convolution kernels. , and are all learnable feature biases, Represents the dot product.

[0031] 2.3 STIGNN model based on spatiotemporal feature mining The present invention Figure 5 The framework of STIGNN is presented in . It includes three layers of serially connected GateTCN modules and two parallel computing ChebGAT modules. The model input data is the road information detected by road sensors at different time periods and the sensor distribution map. Among them, the road sensor will detect the road information at different times and then mine the time correlation through the three layers of GateTCN modules to obtain the road features of the fused time series information. Then, the fused time series road features and the sensor network are input into the ChebGAT module to deeply mine the spatiotemporal characteristics of the road network for the final traffic flow prediction.

[0032] When the ChebGAT module captures spatial features, the present invention introduces an adaptively learned traffic map , The edges and edge weights are learned end-to-end using a stochastic gradient descent algorithm. By doing so, the present invention allows the model to discover hidden spatial dependencies. The adjacency matrix of , using a randomly initialized matrix study : (8), Will Named as source node embedding, Named as the target node embedding. and The multiplication can obtain the spatial dependency weight between the source node and the target node. The present invention uses the ReLU activation function to eliminate weak connections and uses the SoftMax function to normalize the adaptive adjacency matrix.

[0033] Through two ChebGAT modules, the present invention obtains the spatiotemporal feature vectors captured by the sensor network based on the real sensor network and the sensor network based on adaptive learning, which are recorded as and The feature fusion process of the two is as follows: (9), in, is a manually adjusted hyperparameter. Through experiments, it is found that When , the model can achieve the best prediction results.

[0034] 3. Experiment and analysis 3.1 Dataset and Preprocessing The present invention conducts experiments on two public transportation network datasets METR-LA and PEMS-BAY published by Li et al., where METR-LA includes traffic speed data from 207 sensors collected on Los Angeles County freeways from March 1, 2012 to June 30, 2012, and PEMS-BA is collected by the California Department of Transportation (CalTrans) Performance Evaluation System (PEMS), which collects 325 sensor data collected in the Bay Area from January 1, 2017 to May 31, 2017. Table 1 provides detailed dataset information.

[0035] Table 1 Dataset details

[0036] The original data includes the latitude and longitude coordinates of the sensor, the road map, and the data measured by the sensor. In order to determine which sensors should be connected to form a sensor network, the present invention introduces a Gaussian kernel function with a threshold, as shown in the following formula: , in, Representation Node and The spatial distance weight, is the shortest distance between sensors calculated by the Floyd algorithm, is the standard deviation of the sensor distance, is a predefined threshold. When the shortest distance between sensors is less than Only when , it is accepted as the edge of the sensor network and the weight is calculated.

[0037] 3.2 Experimental Setup All experiments were conducted on a single Tesla v100s GPU with 384GB RAM, and the Pytorch deep learning framework was used to reproduce the comparison model. The road information and sensor network detected by the sensors in the first 60 minutes were used as input to detect the traffic flow in the next 15, 30, and 60 minutes, where 15, 30, and 60 are in the short-term, medium-term, and long-term prediction intervals, respectively. The METR-LA and PEMS-BAY datasets involved in the model evaluation were randomly divided into three parts, 70% of which were used as training sets, 10% as validation sets, and the remaining 20% ​​as test sets. During the training process, STIGNN uses the Adam algorithm to optimize the gradient descent process, with the initial learning rate set to 0.0003, dropout set to 0.6, and epoch set to 200.

[0038] The present invention is compared with the following baselines: ARIMA: Autoregressive integrated average model using Kalman filter. FC-LSTM: Recurrent neural network traffic flow prediction model with fully connected LSTM hidden units. DCRNN: A traffic flow prediction model that uses bidirectional random walks on the traffic network to capture spatial dependencies and uses a predetermined sampled encoder-decoder structure to capture temporal dependencies. STGCN: A graph convolution model is directly constructed in a multi-layer traffic network based on time series, replacing the conventional graph neural network + recurrent neural network architecture, effectively reducing the number of parameters and achieving faster training speed. Graph WaveNet: A learnable adaptive dependency matrix is ​​used to capture the hidden spatial dependencies in the data, and long time series are processed through dilated one-dimensional convolutions with different receptive fields. ST-MetaNet: A sequence-to-sequence architecture consisting of an encoder that learns historical information and a decoder that makes predictions step by step is used to mine the spatiotemporal characteristics of the traffic network. STGRAT: An attention-based encoder-decoder architecture is proposed, which uses node attention to capture the spatial correlation between roads and uses temporal attention to capture the temporal dynamics of long sequences. STGAT: Processes long time series by stacking gated temporal convolutional layers and uses graph attention network to mine the spatial characteristics of traffic networks.

[0039] In order to verify the effectiveness of the STIGNN model, the present invention uses the following three indicators to test the performance of each model, including the root mean square error (RMSE): , Mean Absolute Error (MAE): , Mean Absolute Percent Error (MAPE): , in, represents the number of samples, Indicates The traffic flow prediction value of samples, Indicates The actual traffic flow of samples.

[0040] 3.3 Experimental Results Table 2 shows the overall performance of STIGNN and all the comparison models on the PEMS-BAY and METR-LA datasets. STIGNN has achieved excellent results in the three evaluation indicators (MAE, RMSE and MAPE) of these two datasets. The results show that the STIGNN in this invention can be applied to short-term, medium-term and long-term traffic flow prediction scenarios.

[0041] The model combining graph neural network and time series network has lower prediction error than traditional time series analysis models ARIMA and FC-LSTM in all evaluation indicators on the two data sets. Roads are interconnected, and the traffic flow prediction ability of ignoring the spatial topological information of the traffic network is limited, and it cannot handle complex traffic data well. At the same time, the model combining spatiotemporal information will also show different characteristics due to the difference between the spatial feature extraction component and the time series feature mining component. For example, the model STGRAT with dual attention of space and time has strong fitting ability in short-term prediction and achieves the best short-term prediction results in both data. However, the existence of attention weakens its generalization ability, which makes it perform slightly worse than the STIGNN model in long-term prediction. Other models such as Graph WaveNet accurately capture long-term dependencies in the data by developing a new adaptive correlation matrix and learning through node embedding, so that the model also shows high accuracy in long-term prediction. The STIGNN model combines the neighbor order and the correlation of node features in spatial feature mining, and uses the dilated causal convolution with fusion gate in the time series, so that the model achieves the best prediction results in both data sets.

[0042] Due to the extremely complex traffic network in Los Angeles, the performance of the model in the METR-LA dataset is lower than that of PEMS-BAY. However, the STIGNN model significantly reduces the prediction error on the METR-LA data, which means that compared with other baseline methods, the STIGNN model can better handle complex network situations.

[0043] Table 2 Performance comparison of STIGNN and other models on PEMS-BAY and METR-LA datasets

[0044] The present invention proposes a spatiotemporal deep learning model STIGNN for traffic flow prediction. STIGNN is a general model for processing spatiotemporal prediction tasks. In order to comprehensively consider the distance between traffic points and the differences in road conditions, the present invention uses dual attention of road node features and distance between roads, and designs the ChebGAT module to mine potential and existing spatial dependencies. In order to effectively extract the time series information of the traffic network, STIGNN introduces a gating unit in the TCN module, named GateTCN module, to realize the mining and application of long time series features. Finally, STIGNN achieved good results on two public real traffic network datasets.

[0045] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. A traffic flow prediction model that realizes the co-mining of traffic point semantics and network topology, characterized in that: The name of the model is STIGNN. The STIGNN model consists of two parallel computing ChebGAT modules and three-layer serially connected GateTCN modules. The ChebGAT module couples the ChebNet static attention and the GAT dynamic attention to fully explore the spatial correlation of the transportation network and realize the simultaneous mining of topological features and semantic information. The GateTCN module integrates the time series model of the gating mechanism to mine the temporal correlation of road information detected at different times, and then obtains the road features of the fused time series information. The road features of the fused time series and the sensor network are then input into the ChebGAT module to deeply mine the spatiotemporal characteristics of the road network for the final traffic flow prediction.

2. A traffic flow prediction model for realizing co-mining of traffic point semantics and network topology according to claim 1, characterized in that: The process of aggregating features in the ChebNet model is as follows: ， in, represents a nonlinear activation function, It is The learnable static weights of the neighbors of order , When the value is expressed, it indicates the importance of the information itself. Representation Node The degree, is a learnable weight matrix, Representation Node The traffic point information of the first-order neighbors, Representation Node Traffic point information of the second-order neighbors, Representation Node Traffic point information of the third-order neighbors, Representation Node The degree, Representation Node The degree.

3. A traffic flow prediction model for realizing co-mining of traffic point semantics and network topology according to claim 2, characterized in that: The GAT aggregation feature process is as follows: ， in, Representation Node and The degree of semantic relevance of ， in, Represents a splicing operation, is a learnable weight vector, Representation Node The first-order neighbor set of .

4. A traffic flow prediction model for realizing co-mining of traffic point semantics and network topology according to claim 3, characterized in that: The neighborhood perception range of the ChebGAT module is fixed to 3rd order, and the process of aggregating features is as follows: 。 5. A traffic flow prediction model for realizing co-mining of traffic point semantics and network topology according to claim 4, characterized in that: STIGNN model transposes the sensor feature matrix , while setting the dilated causal convolution kernel ,in, Represents the size of the convolution kernel, and the process of dilated causal convolution can be expressed as: ， in, is the dilation factor of the dilated causal convolution, Indicates the time series length of the intercepted traffic data, It is a traversal The numerical value of Indicates the use of dilated causal convolution kernels right Perform convolution, Represents the sth parameter of the convolution kernel f.

6. A traffic flow prediction model for realizing co-mining of traffic point semantics and network topology according to claim 5, characterized in that: The GateTCN module data feature processing process is as follows: ， in, Indicates GateTCN No. The update gate of the layer is used to control the information ratio of longer time series and shorter time series. Represents the reset gate, To obtain long time series information, , and are all learnable convolution kernels. , and are all learnable feature biases, Represents the dot product.

7. A traffic flow prediction model for realizing co-mining of traffic point semantics and network topology according to claim 6, characterized in that: The ChebGAT module introduces an adaptively learned traffic map while capturing spatial features. , The edges and edge weights are learned end-to-end using the stochastic gradient descent algorithm. The adjacency matrix of , using a randomly initialized matrix study : ， in, Named as source node embedding, Name the target node embedding.

Images