Segmented iteration long-term traffic flow prediction method based on space-time diagram convolutional network

Abstract

The invention discloses a segmented iteration long-term traffic flow prediction method based on a space-time diagram convolutional network. Firstly, entities of sensors involved in an application scene are mapped into node weight information of a graph theory to achieve problem modeling, and a graph model is obtained; dividing the traffic characteristic matrix to serve as the input of the space-time diagram convolutional network; and finally, predicting the prediction result of each segment in parallel by using a parallel multi-step PMF prediction method, and embedding the predicted segments in combination with positions to obtain the traffic information of the next moments. According to the space-time diagram convolutional network, segmentation iteration is adopted to replace point-by-point iteration, and parallel multi-step prediction PMF is adopted to replace recursive multi-step prediction RMF during prediction. According to the method, the recursive iteration times required for extracting the time features are remarkably reduced through a segmented iteration strategy, so that the challenge of effectively training the GRU on a long sequence is solved. In addition, due to the adoption of the PMF, the inherent error accumulation problem in the traditional RMF method is further relieved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation, and in particular relates to a piecewise iterative long-term traffic flow prediction method based on a spatiotemporal graph convolutional network. Background Art

[0002] With the rapid development of the Internet of Things and 5G communication technology, people can obtain extensive traffic information through various devices or systems, which has promoted the development of intelligent transportation systems. Intelligent transportation systems are effective ways to relieve congestion, provide travelers with reliable travel routes, and provide reference suggestions for urban traffic management. As an important research topic in the field of intelligent transportation systems, the use of deep learning methods for traffic flow prediction has attracted much attention in recent years. Through deep learning methods, we can better understand the hidden information hidden in complex traffic networks, which greatly improves the accuracy of our traffic flow prediction. Traffic flow prediction is divided into short-term prediction and long-term prediction, and the basis for the division is based on the size of the lookback window and the prediction range. Short-term traffic flow prediction uses a shorter lookback window, that is, a small amount of historical data is used to predict recent changes, and long-term traffic flow prediction uses a longer lookback window, that is, a large amount of historical data is considered for long-term prediction. However, many existing studies focus on short-term traffic flow prediction. Long-term traffic flow prediction faces challenges when dealing with too long lookback windows and prediction ranges. Summary of the invention

[0003] In view of the deficiencies in the prior art, the present invention provides a segmented iterative long-term traffic flow prediction method based on a spatiotemporal graph convolutional network. The present invention aims to develop a new deep learning method for segmented iterative long-term traffic flow prediction based on a spatiotemporal graph convolutional network to predict the vehicle flow at each observation point in a road network. The technology uses a spatiotemporal graph convolutional network to solve the problems in the above-mentioned technology.

[0004] The implementation process of the segmented iterative long-term traffic flow prediction method based on the spatiotemporal graph convolutional network is as follows:

[0005] Step (1) Spatial dependency modeling:

[0006] The sensor entities involved in the application scenario are mapped into the node weight information of graph theory to realize problem modeling, and the graph model, i.e., the weighted undirected fully connected graph, is obtained to obtain the weighted adjacency matrix W;

[0007] Step (2) Time-dependent modeling:

[0008] The traffic feature matrix is ​​obtained based on the traffic volume collected by each sensor within time P. The sample normalization strategy is used to alleviate the distribution offset problem of time series data.

[0009] Step (3) Encoding:

[0010] The traffic feature matrix processed by the sample normalization strategy is divided into multiple segments as the input of the spatiotemporal graph convolutional network to obtain the final hidden features containing spatial features and temporal features.

[0011] First, the divided traffic feature matrix and weighted adjacency matrix W are input into the spatial correlation extraction module to extract deep spatial features, and then input into the temporal correlation extraction module to extract deep temporal features. Finally, the feature hidden feature h containing spatial features and temporal features is obtained. n .

[0012] Step (4) Decoding:

[0013] Using the parallel multi-step prediction PMF method, based on the obtained hidden feature h n All segments are predicted in parallel to obtain the prediction results of each segment. The prediction results of each segment are combined with the position embedding to obtain the traffic information for the next T moments.

[0014] Step 5: Train the segmented iteration-based spatiotemporal graph convolutional network model SegT-GCN (Encoding+Decoding).

[0015] Furthermore, the specific method of step (1) is as follows:

[0016] For N sensors set on the road network, a weighted undirected fully connected graph G = (V, E, W) is used to describe the topological structure of the road network. Each sensor is regarded as a node v, V = {v_1, v_2, ..., v_N} represents the set of sensors, where N is the number of nodes. The nodes are connected in pairs to form a set of edges E, E = n*n-1) / 2, and the weighted adjacency matrix W is used to represent the proximity between nodes.

[0017] In order to represent the spatial correlation between two points, the weighted adjacency matrix W based on Gaussian similarity is calculated by the equation, where the weight w(I, j) between node i and node j is calculated as follows:

[0018] w(i,j)=exp(-dist(i,j) 2 / σ 2 )

[0019] where dist(I,j) represents the distance between node i and node j, and the width of the neighborhood is controlled by the standard deviation σ of the distance.

[0020] Calculate the weights between all nodes in the weighted undirected fully connected graph to obtain the weighted adjacency matrix W. The greater the correlation between two points, the greater the weight. The weighted adjacency matrix W is as follows:

[0021]

[0022] Furthermore, the specific method of step (2) is as follows:

[0023] Based on the traffic volume collected by each sensor within P time, the traffic feature matrix X∈R is obtained. N×P , the traffic feature matrix X is as follows:

[0024]

[0025] Where P represents the number of node time series features (i.e., the length of the historical time series), and N represents the number of sensors. i ∈R N×i , represents the traffic volume at each sensor at time i.

[0026] Use sample normalization strategies to alleviate the distribution shift problem of time series data.

[0027] The sample normalization strategy consists in subtracting the last value of the sequence from the value collected by the kth sensor. This value is then added back after the decoding phase, using the following formula:

[0028]

[0029] Where P represents the length of the historical time series, T represents the length of the predicted time series, and k represents the kth sensor. Represents the value collected by the kth sensor. Represents the value predicted by k sensors.

[0030] After defining the weighted adjacency matrix W and the traffic feature matrix X, the traffic volume prediction problem is converted into the problem of learning the mapping function f, through which the traffic information of the next T moments can be calculated. The mapping function f is shown as follows:

[0031] f(W,X)=[X P+1 ,X P+2 ,...,X P+T ]

[0032] Where W represents the weighted adjacency matrix, X represents the traffic feature matrix, and X P+1 ,X P+2 ,…,X P+T Represents the traffic volume at each sensor in the next T moments.

[0033] Furthermore, the specific implementation process in step (3) is as follows:

[0034] 3.1 Data preparation: In long-term time series forecasting, when dealing with long look-back windows and forecast ranges, a large number of repeated iterations will occur when forecasting based on RNN methods, resulting in unsatisfactory forecast accuracy and inference speed. Use segmented iteration and parallel multi-step forecasting to reduce the number of iterations and obtain better forecast accuracy and inference speed.

[0035] The traffic feature matrix X∈R N×P , divided into n segments according to the set window length l, the traffic characteristic matrix of a single segment is expressed as Where l represents the window length of each segment, Indicates the number of segments.

[0036] 3.2 Extract deep spatial features through spatial correlation extraction module:

[0037] The spatial correlation extraction module uses a 2-layer GCN model to learn spatial features from traffic data. The 2-layer GCN model is represented as follows:

[0038]

[0039] Where W represents the weighted adjacency matrix, represents the preprocessing step, is a matrix with a self-connected structure, I N is the identity matrix, is a degree matrix, W0 and W1 represent the weight matrices of the first and second layer GCN models respectively, and σ( ) and Relu( ) represent activation functions.

[0040] 3.3 Extracting deep temporal features through the temporal correlation extraction module:

[0041] Segmented traffic feature matrix After the spatial correlation is extracted by the 2-layer GCN model, it is input into the temporal correlation extraction module through linear projection W pri ∈R l×d The converted output is and hidden features h i It is sent to the gated recurrent unit GRU to capture temporal features. The entire process in the GRU unit can be formulated as:

[0042]

[0043]

[0044]

[0045]

[0046] Among them, z i represents the update gate and r i represents the reset gate, b r and b z Indicates the offset, W z , W r , W h represents the corresponding weight matrix, h i represents the hidden features of the current time step i, h i represents the candidate hidden features containing the information of the current time step, Represents the convolution operation on a graph.

[0047] After n loop iterations, the final hidden feature h obtained from the last step is n All temporal features of the original sequence X have been encapsulated. This hidden feature will be passed to the decoding part for subsequent inference and prediction steps.

[0048] Furthermore, the specific method of step 4 is as follows:

[0049] 4.1 Position Embedding: During the decoding phase, the order between segments is lost due to the interruption of the loop recursion. To solve this problem, m corresponding position embeddings are generated, each denoted as PE (j) ∈R m×d , to identify the location of the segment. It means that the prediction range is divided into m segments, T represents the prediction range, and l represents the window length of each segment.

[0050] 4.2 Parallel Multi-step Prediction PMF: The use of traditional recursive multi-step prediction RMF leads to error accumulation. At the same time, the sequential nature of recursion hinders parallel computation in training examples and limits the improvement of inference speed. To address these limitations, a new prediction strategy, namely parallel multi-step prediction PMF, is adopted. The specific operation is as follows: In the decoding stage, the same GRU unit used in the encoding stage is shared. Specifically, the final hidden feature h obtained from the encoding stage is n Copy m times and embed PE in m locations respectively (j) Combined as the input of the GRU unit.

[0051] Position Embedded PE (j) Reflects the position of each segment that needs to be predicted in the complete sequence. The GRU unit processes m inputs in parallel and generates m predicted segments. represents the jth segment of the prediction, which contains the predicted results of d time steps of N sensors. Parallel decoding realizes parallel computation within the sample, thereby improving the inference speed. In addition, the prediction error does not accumulate with the increase of the number of iterations, thereby improving the prediction accuracy.

[0052] 4.3 Prediction and sequence recovery: After the Dropout layer, 20% of the values ​​are randomly dropped for regularization purposes, and then a learnable linear prediction layer W is used. prd ∈R d×l Convert it to Since sample normalization is performed in time-dependent modeling, the value collected on the kth sensor is subtracted from the last value of the sequence, and all After combining them together, add the value back at the same time to get the final prediction result

[0053] Furthermore, the specific method of step 5 is as follows:

[0054] Based on the data collected by N sensors set up on the road network, a dataset is constructed to train and evaluate the prediction performance of the SegT-GCN model consisting of the encoding stage and the decoding stage.

[0055] The two loss functions of mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the experimental results, which are expressed as:

[0056]

[0057] Where T represents the length of the predicted time series, N represents the number of sensors, represents the true value of the kth sensor at time t, Represents the predicted value of the kth sensor at time t.

[0058] The beneficial effects of the present invention are as follows:

[0059] The temporal correlation extraction module of the present invention uses segmented iteration instead of point-by-point iteration, and uses parallel multi-step prediction PMF instead of recursive multi-step prediction RMF during prediction. The present invention significantly reduces the number of recursive iterations required to extract temporal features through a segmented iteration strategy, thereby solving the challenge of effectively training GRU on long sequences. In addition, the use of PMF further alleviates the error accumulation problem inherent in traditional RMF methods. By adopting these innovative strategies, SegT-GCN not only achieves better results in prediction accuracy, but also significantly improves efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1A schematic diagram of a system scenario of an embodiment of the method of the present invention;

[0061] Figure 2 This is a schematic diagram of the network structure of an embodiment of the present invention;

[0062] Figure 3 This is a diagram of the prediction results of the sensor S1 part of the embodiment of the present invention;

[0063] Figure 4 This is a diagram of the prediction results of the sensor S2 part of the embodiment of the present invention;

[0064] Figure 5 This is a partial prediction result diagram of the sensor S3 according to an embodiment of the present invention;

[0065] Figure 6 This is a partial prediction result diagram of sensor S4 according to an embodiment of the present invention;

[0066] Figure 7 This is a diagram of the prediction results of the sensor S5 part of the embodiment of the present invention;

[0067] Figure 8 This is a comparison diagram of errors in the method of the embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to enable relevant personnel to more clearly understand the technical content of the present invention, the technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0069] The present invention provides a segmented iterative long-term traffic flow prediction method based on a spatiotemporal graph convolutional network, and designs a SegT-GCN model. The traffic feature matrix is ​​segmented, and the segmented traffic feature matrix is ​​input into a spatial correlation extraction module to extract deep spatial features, and then input into a temporal correlation extraction module to obtain deep temporal features. Finally, the parallel multi-step prediction PMF method is used to parallelly predict the prediction results of each segment, and the predicted segments are combined with the position embedding to obtain the traffic information of the next T moments. SegT-GCN not only achieves good results in prediction accuracy, but also significantly improves efficiency. In addition, the present invention has good reusability in similar application scenarios, and the practical value of the invention is relatively strong.

[0070] The segmented iterative long-term traffic flow prediction method based on spatiotemporal graph convolutional network includes the following steps:

[0071] Step 1: N sensors are set up on the road network, as shown in the following figure. Figure 1As shown in the system scenario diagram, we use N = 5 sensors. Each sensor is regarded as a node v, V = {v_1, v_2, ..., v_N} represents the set of sensors, where N is the number of nodes. The nodes are connected in pairs to form a set of edges E, E = n(n-1) / 2, and the weighted adjacency matrix W is used to represent the proximity between nodes. In order to represent the spatial correlation between two points, we calculate the weighted adjacency matrix W based on Gaussian similarity through the equation, the formula is as follows:

[0072] w(i,j)=exp(-dist(i,j) 2 / σ 2 )

[0073] where dist(I,j) represents the distance between node i and node j, and the width of the neighborhood is controlled by the standard deviation σ of the distance.

[0074] Thus, the weighted adjacency matrix W is calculated. The greater the correlation between two points, the greater the weight calculation. The weighted adjacency matrix W is as follows:

[0075]

[0076] Step 2: Based on the traffic volume collected by each sensor within P time, the traffic feature matrix X∈R is obtained N ×P , the traffic feature matrix X is as follows:

[0077]

[0078] Where P represents the number of node time series features (i.e., the length of the historical time series), and N represents the number of sensors. i ∈R N×i , represents the traffic volume at each sensor at time i.

[0079] After defining the weighted adjacency matrix W and the traffic feature matrix X, the traffic volume prediction problem is converted into the problem of learning the mapping function f, through which the traffic information of the next T moments can be calculated. The mapping function f is shown as follows:

[0080] f(W,X)=[X P+1 ,X P+2 ,...,X P+T ]

[0081] Where W represents the weighted adjacency matrix, X represents the traffic feature matrix, and X P+1 ,X P+2 ,…,X P+T Represents the traffic volume at each sensor in the next T moments.

[0082] Step 3, instance normalization: The present invention uses a sample normalization strategy to alleviate the distribution offset problem of time series data. The sample normalization strategy includes subtracting the last value of the sequence from the value collected on the kth sensor, and then adding the value back after the decoding stage. The formula is as follows:

[0083]

[0084]

[0085] Where P represents the length of the historical time series, T represents the length of the predicted time series, and k represents the kth sensor. Represents the value collected by the kth sensor. Represents the value predicted by k sensors.

[0086] Step 4: Transform the traffic feature matrix X∈R N×P , divided into X∈R N× n× l , where l represents the window length of each segment, Indicates the number of segments. Represents the traffic feature matrix of the i-th segment with segment length l.

[0087] Step 5: The spatial correlation extraction module extracts deep spatial features: The GCN model is used to learn spatial features from traffic data. The 2-layer GCN model can be expressed as:

[0088]

[0089] Where X represents the traffic feature matrix, W represents the weighted adjacency matrix, represents the preprocessing step, is a matrix with a self-connected structure, I N is the identity matrix, is a degree matrix, W0 and W1 represent the weight matrices of the first and second layers respectively, and σ( ) and Relu( ) represent activation functions.

[0090] Step 6: The time correlation extraction module extracts deep temporal features: segmented traffic feature matrix After the spatial correlation is extracted by the 2-layer GCN model, it is input into the temporal correlation extraction module through linear projection W pri ∈R l×d The dimension is converted by using the Relu() activation function, where d represents the dimension of the hidden state of the gated recurrent unit GRU. The converted output is and hidden features h iIt is sent to the gated recurrent unit GRU to capture temporal features. The entire process in the GRU unit can be formulated as:

[0091]

[0092]

[0093]

[0094]

[0095] Among them, z i represents the update gate and r i represents the reset gate, b r and b z Indicates the offset, W z , W r , W h represents the weight matrix, h i represents the hidden features of the current time step t, h i represents the candidate hidden features containing the information of the current time step, Represents the convolution operation on a graph.

[0096] After n loop iterations, the final hidden feature h obtained from the last step is n All temporal features of the original sequence X have been encapsulated. This hidden feature will be passed to the decoding part for subsequent inference and prediction steps.

[0097] Step 7: Parallel decoding: In the decoding phase, the same GRU units used in the encoding phase are shared. Specifically, the final state h obtained from the encoding phase n is copied m times and combined with (j) The position embedding PE (j) Reflects the position of each fragment that needs to be predicted in the complete sequence.

[0098] GRU units process in parallel and generate m prediction fragments It represents the jth predicted segment, which contains the predicted results of N sensors for d time steps.

[0099] Step 8: Prediction and sequence recovery: After the Dropout layer, 20% of the values ​​are randomly dropped for regularization purposes, and then a learnable linear prediction layer W is used. prd ∈R d×l Convert it to Since sample normalization is performed in time-dependent modeling, the value collected on the kth sensor is subtracted from the last value of the sequence, and all After combining them together, add the value back at the same time to get the final prediction result The network structure is as follows Figure 2 The network diagram is shown.

[0100] Step 9: Train the segmented iteration-based spatiotemporal graph convolutional network model SegT-GCN (Encoding+Decoding).

[0101] Models such as ITransformer and TimesNet are selected as baselines. The two loss functions, mean absolute error (MAE) and root mean square error (RMSE), are used to evaluate the experimental results and calculate the prediction error, which is expressed as:

[0102]

[0103]

[0104] Where T represents the length of the predicted time series, N represents the number of sensors, represents the true value of the kth sensor at time t, Represents the predicted value of the kth sensor at time t.

[0105] We used a real dataset JX-cars to train and evaluate the prediction performance of the SegT-GCN model, which consists of the encoding stage and the decoding stage. JX-cars records the car traffic data of a certain place in Jiaxing from January 1 to December 31, 2023. 80% of the data is used as the training set and 20% of the data is used as the test set. We selected sensors at 5 observation points as the research area, such as Figure 1 The uniform configuration of SegT-GCN includes a lookback of 720, a segment length of 48, a single GRU layer, a hidden size of 512, 30 training epochs, a learning rate decay of 0.8 after the first 3 epochs, and early stopping with a patience value of 10. The batch size is set to 64. The prediction results of some sensors S1-S5 are shown in the attached figure. Figure 3-7 shown.

[0106] After completing the above steps, this method was used to predict the traffic flow data collected from a road network in Jiaxing. The prediction error results are shown in the attached figure. Figure 8 As shown, the results show that in most cases, the proposed method outperforms existing newer methods.

[0107] It should be known that the parts not elaborated in detail in this specification belong to the prior art. Relevant technical personnel should understand that the above embodiments are only to help readers understand the principles and implementation methods of the present invention, and the scope of protection of the present invention is not limited to such embodiments. All equivalent replacements made on the basis of the present invention are within the protection scope of the rights of the present invention.

Claims

1. A piecewise iterative long-term traffic flow prediction method based on spatiotemporal graph convolutional network, characterized in that: The implementation process is as follows: Step (1) Spatial dependency modeling: The sensor entities involved in the application scenario are mapped into the node weight information of graph theory to realize problem modeling, and the graph model, i.e., the weighted undirected fully connected graph, is obtained to obtain the weighted adjacency matrix W; Step (2) Time-dependent modeling: The traffic characteristic matrix is ​​obtained based on the traffic volume collected by each sensor within P time; Use sample normalization strategies to alleviate the distribution shift problem of time series data; Step (3) Encoding: The traffic feature matrix processed by the sample normalization strategy is divided into multiple segments as the input of the spatiotemporal graph convolutional network to obtain the final hidden features containing spatial features and temporal features; First, the divided traffic feature matrix and weighted adjacency matrix W are input into the spatial correlation extraction module to extract deep spatial features, and then input into the temporal correlation extraction module to extract deep temporal features. Finally, the feature hidden feature h containing spatial features and temporal features is obtained. n ; Step (4) Decoding: Using the parallel multi-step prediction PMF method, based on the obtained hidden feature h n Perform parallel prediction on all segments to obtain the prediction results of each segment, and embed the prediction results of each segment into positions to obtain the traffic information for the next T moments; Step 5: Train the segmented iteration-based spatiotemporal graph convolutional network model SegT-GCN.

2. According to claim 1, the segmented iterative long-term traffic flow prediction method based on spatiotemporal graph convolutional network is characterized in that: The specific method of step (1) is as follows: For N sensors set on the road network, a weighted undirected fully connected graph G = V, E, W is used to describe the topological structure of the road network. Each sensor is regarded as a node v, V = {v_1, v_2, ..., v_N} represents the set of sensors, where N is the number of nodes; the nodes are connected in pairs to form a set of edges E, E = n(n-1) / 2, and the weighted adjacency matrix W is used to represent the proximity between nodes; In order to represent the spatial correlation between two points, the weighted adjacency matrix W based on Gaussian similarity is calculated by the equation, where the weight w(I, j) between node i and node j is calculated as follows: w(i,j)=exp(-dist(i,j) 2 / σ 2 ) Where dist(I,j) represents the distance between node i and node j, and the width of the neighborhood is controlled by the standard deviation σ of the distance; Calculate the weights between all nodes in the weighted undirected fully connected graph to obtain the weighted adjacency matrix W; the greater the correlation between two points, the greater the weight; the weighted adjacency matrix W is as follows:

3. The segmented iterative long-term traffic flow prediction method based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: The specific method of step (2) is as follows: Based on the traffic volume collected by each sensor within P time, the traffic feature matrix X∈R is obtained. N×P , the traffic feature matrix X is as follows: Where P represents the number of node time series features, and N represents the number of sensors; Use sample normalization strategies to alleviate the distribution shift problem of time series data; The sample normalization strategy consists in subtracting the last value of the sequence from the value collected by the kth sensor. This value is then added back after the decoding phase, using the following formula: Where P represents the length of the historical time series, T represents the length of the predicted time series, and k represents the kth sensor; Represents the value collected by the kth sensor; Represents the value predicted by k sensors.

4. The segmented iterative long-term traffic flow prediction method based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: The specific implementation process in step (3) is as follows: 3.1 Data preparation: The traffic feature matrix X∈R N×P , divided into n segments according to the set window length l, the traffic characteristic matrix of a single segment is expressed as Where l represents the window length of each segment, i∈1,2,...,n, Indicates the number of segments; 3.2 Extract deep spatial features through spatial correlation extraction module: The spatial correlation extraction module uses a 2-layer GCN model to learn spatial features from traffic data; The 2-layer GCN model is represented as follows: Where W represents the weighted adjacency matrix, represents the preprocessing step, is a matrix with a self-connected structure, I N is the identity matrix, is a degree matrix, W0 and W1 represent the weight matrices of the first and second layer GCN models respectively, σ( ) and Relu( ) represent activation functions; 3.3 Extracting deep temporal features through the temporal correlation extraction module: Segmented traffic feature matrix After the spatial correlation is extracted by the 2-layer GCN model, it is input into the temporal correlation extraction module through linear projection W pri ∈R l×d The dimension is converted using the Relu() activation function, where d represents the dimension of the hidden state of the gated recurrent unit GRU; the converted output and hidden features h i It is sent to the gated recurrent unit GRU to capture temporal features; the entire process in the GRU unit can be formulated as: Among them, z i represents the update gate and r i represents the reset gate, b r and b z Indicates the offset, W z , W r , W h represents the corresponding weight matrix, h i represents the hidden features of the current time step i, h i represents the candidate hidden features containing the information of the current time step, Represents the convolution operation of the graph; After n loop iterations, the final hidden feature h obtained from the last step is n All temporal features of the original sequence X have been encapsulated; this hidden feature will be passed to the decoding part for subsequent inference and prediction steps.

5. The segmented iterative long-term traffic flow prediction method based on spatiotemporal graph convolutional network according to claim 4 is characterized in that: Step 4: 4.1 Position Embedding: During the decoding phase, the order between segments is lost due to the interruption of the loop recursion; to solve this problem, m corresponding position embeddings are generated, each denoted as PE (j) ∈R m×d , to identify the location of the segment; where Indicates that the prediction range is divided into m segments, T represents the prediction range, and l represents the window length of each segment; 4.2 Parallel multi-step prediction PMF: Parallel multi-step prediction PMF, the specific operation is as follows: in the decoding stage, share the same GRU unit used in the encoding stage; specifically, the final hidden feature h obtained from the encoding stage n Copy m times and embed PE in m locations respectively *j) Combined as the input of the GRU unit; Position Embedded PE (j( Reflects the position of each segment that needs to be predicted in the complete sequence; the GRU unit processes m inputs in parallel and generates m predicted segments represents the jth predicted segment, which contains the predicted results of d time steps of N sensors; 4.3 Prediction and sequence recovery: After the Dropout layer, 20% of the values ​​are randomly dropped for regularization purposes, and then a learnable linear prediction layer W is used. prd ∈R d×l Convert it to Since the sample normalization operation is performed in the time-dependent modeling, the value collected on the kth sensor is subtracted from the last value of the sequence, and all After combining them together, add the value back at the same time to get the final prediction result 6. The segmented iterative long-term traffic flow prediction method based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: Step 5: Based on the data collected by N sensors set up on the road network, a dataset is constructed to train and evaluate the prediction performance of the SegT-GCN model consisting of the encoding stage and the decoding stage. The two loss functions of mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the experimental results, which are expressed as: Where T represents the length of the predicted time series, N represents the number of sensors, represents the true value of the kth sensor at time t, Represents the predicted value of the kth sensor at time t.

Images