A Lanczos spatiotemporal network method for real-time traffic prediction
Through the Lanczos spatiotemporal network method, combined with the spatiotemporal convolution and graph convolution module, the problem of low real-time traffic prediction accuracy on the highway is solved, and real-time traffic prediction with higher accuracy and robustness is achieved, suitable for large spans, Long range of highways.
Patent Information
- Application Number
- CN202210273493.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-18
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-03-18
AI Technical Summary
The existing real-time traffic prediction methods have low prediction accuracy on highways with large spans and long range fluctuations, and the data is unreliable, making it difficult to meet the needs of real-time traffic management and travel behavior guidance.
By using the Lanczos spatiotemporal network method, by constructing multi-dimensional structured feature data and road network data, using the spatiotemporal convolution layer and the time convolution layer combined with the graph convolution module, combined with the Lanczos algorithm and GRU model, real-time prediction of traffic is achieved and long-distance capture Dependencies and optimize prediction models.
It significantly improves prediction accuracy and robustness. It is suitable for highways with large spans and long range fluctuations. It provides reliable real-time traffic prediction capabilities and can be directly applied on intelligent control platforms without repeated training.
Smart Images

Figure CN114860715B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of deep learning, and in particular to a Lanczos spatiotemporal network method for real-time traffic prediction in an intelligent prediction management system for highways. Background Art
[0002] With the continuous innovation and increasing demand for smart transportation, the use of intelligent prediction algorithms to predict future traffic conditions and perceive traffic conditions in advance can not only prompt traffic managers to take effective management and control measures, but also guide traffic users to change their travel behaviors or demands, thereby improving transportation efficiency and travel experience.
[0003] Real-time traffic flow prediction refers to linking the current traffic flow status with the status in the future to predict the real-time status of traffic flow. There are currently some real-time traffic flow prediction methods, but these methods have some shortcomings, such as poor prediction accuracy and poor effectiveness, especially for highways with large spans and long-range fluctuations, the prediction accuracy is low and the prediction data is unreliable. Summary of the invention
[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention proposes a Lanczos spatiotemporal network method for real-time flow prediction.
[0005] In order to achieve the above object, the present invention provides the following technical solutions:
[0006] A Lanczos spatiotemporal network method for real-time traffic prediction includes the following steps:
[0007] (1) Construct multi-dimensional structured feature data and road network data;
[0008] Obtain flow and traffic situation data in real time;
[0009] The multi-dimensional structured data includes traffic situation and meteorological data. The traffic situation data includes but is not limited to flow, speed, density, and vehicle type ratio data; the meteorological data includes but is not limited to weather conditions and temperature; the road network data is composed of node data of detectors of the predicted target road network;
[0010] (2) Determine the network structure;
[0011] Data input layer: Use the spatiotemporal module as the feature input encoder; Spatiotemporal convolution layer: Input features are input into the spatial convolution layer and the temporal convolution layer respectively, and then the residual operation is performed, and the results are returned to the output layer; Output layer: Use the predictor to output the last layer of data;
[0012] (3) Construct spatial convolution layer;
[0013] Spatial convolution is implemented as the graph convolution kernel based on the Lanczos method. First, the eigenvalues and eigenvectors of the similarity transformation matrix are calculated according to the road network data. The isomorphic transformation between the similarity transformation matrix and the road network matrix is realized through a multi-layer perceptron to obtain different matrices for mining spatial association rules. The matrices are then combined with multi-dimensional data through a graph convolution module to obtain feature sets of different association rules. The spatial module output is obtained through a fusion of linear transformations.
[0014] (4) Construct a temporal convolution layer;
[0015] Use GRU as the temporal convolution module, input the empty sequence feature data of step (3) into the gated recurrent unit, use the gating mechanism to filter the propagation features of long and short time series, and obtain the time series output through weight parameterization;
[0016] (5) Achieve real-time prediction;
[0017] Perform a two-dimensional convolution operation on the data obtained in step 4 to obtain the prediction result, and pass it into the loss function together with the predicted target data. After passing through the iterator, the loss function value reaches an acceptable threshold. That is, the training parameters are changed, and the prediction results under different parameters are compared to obtain the best prediction model.
[0018] Furthermore, in step (1), constructing multi-dimensional structured feature data and road network data is achieved by the following sub-steps:
[0019] (1.1) The traffic data is the historical traffic data of the Y period before time T on N road sections; the sampling time interval of the historical traffic data is uniform, the time interval is Δt, and the amount of sampled traffic data is Y / Δt;
[0020] (1.2) The traffic situation data is the historical data of the Y period before time T on N road sections and the corresponding predictions T+t1, T+t2, ..., T+t k The traffic situation data of K time points; the sampling time interval of historical data is uniform, the time interval is Δt, and the amount of sampled traffic data is Y / Δt;
[0021] (1.3) The topological relationship between the nodes in the road network is applicable to graph theory and is described as G t =(V t ,E,A), represents the state of the road network at time t, where V t =(v t:1 ;v t:2 ;...;v t:N ) represents the status of N road segments at time t, E represents the connectivity between these road segments, A is the adjacency matrix of the road network, and when the i and j road segments are connected, A i;j =1, otherwise 0, where A i;j =A j;i , that is, A is a symmetric matrix.
[0022] Furthermore, in step (2), determining the network structure is achieved by the following sub-steps:
[0023] (2.1) Construct the network data input-output layer; in the data input layer, use the spatiotemporal module as the feature input encoder; in the data output layer, use the two-dimensional convolutional basic unit as the predictor of the network output layer.
[0024] (2.2) Constructing the residual module: Perform residual operation on the results returned by the spatiotemporal convolution layer, that is, the input of the network and the input of the spatiotemporal convolution layer are re-fused each time, and two-dimensional convolution is used as the fusion method of the residual module to achieve perception between the hidden layer and the original data.
[0025] Furthermore, in step (3), constructing the spatial convolutional layer is achieved by the following sub-steps:
[0026] (3.1) Constructing affinity matrix From graph theory, we know that the degree of road section i is N represents the number of road sections, and the degree matrix D of the road network is obtained, where D ii =d i , and then we get the Laplace matrix L = DA, and after symmetric normalization of L, we get the affinity matrix:
[0027]
[0028] Here, I represents the identity matrix of the same type as A;
[0029] (3.2) Take the initial non-zero vector s and a positive integer M less than N, and perform the Lanczos similarity transformation algorithm. According to the results of the Lanczos algorithm, the matrix V can be obtained from this set of basis vectors M =[v1,…,v M ], and a triangular symmetric matrix H M , whose diagonal and superdiagonal elements are {α1,α2,…,α M} and {β1,β2,…,β M}, its H M The construction rules are as follows:
[0030]
[0031] According to the nature of the algorithm, V M * V M The transpose of
[0032] (3.3) to get H M Then, by its triangular symmetric matrix H MPerform eigendecomposition H M =BRB * , R, B, B * H M The decomposed eigenvalue matrix, eigenvector matrix and their transpose; the approximate eigenvector matrix V = V M B, therefore The approximate eigenvalues and eigenvectors of are
[0033] (3.4) For the graph signal, i.e., multidimensional structured feature data, let the multidimensional data feature input and the spatial convolution layer output be X t , S t , then the graph convolution of the road network G can be approximated as:
[0034]
[0035] The parameterized graph convolution is:
[0036] S t =[VR u V * X t ]W
[0037] Here, u represents the uth power, and W represents the learnable parameter weights of the graph signal propagation in the spatial convolutional layer;
[0038] (3.5) Then, by expanding the eigenvalues, we can capture long-distance dependencies. Implement enhanced convolution filtering:
[0039]
[0040] Where W is the learnable parameter weight of the graph signal propagation in the spatial convolution layer, and E is the length of the associated node size set l, which is intended to capture the features of adjacent nodes with a distance of li. This way of capturing long-distance dependencies can effectively save space computing power.
[0041] Furthermore, constructing the temporal convolutional layer in step (4) is implemented by the following sub-steps:
[0042] The result of the spatial convolution layer S t As the input of the time module, a gated recurrent unit is used as its forward propagation method, and the forward propagation model formula is:
[0043]
[0044] Where σ is the activation function, [] represents tensor connection, * represents element multiplication, and · represents matrix product. z , Wr , Represent the update gate, reset gate and learnable weights of the hidden layer of the temporal convolutional layer respectively.
[0045] Furthermore, in step (5), the real-time prediction is achieved by the following sub-steps:
[0046] (5.1) Design the optimizer and loss function of the model; the optimizer is the Adam optimizer, and the loss function uses the MAE (mean absolute error) suitable for the regression prediction function. The specific discriminant formula is:
[0047]
[0048] Where m represents the total number of samples, y n Represents the actual flow value of sample n; is the flow prediction value of sample n;
[0049] (5.2) Obtain prediction results; train the prediction model to minimize the model loss function, iterate repeatedly until the model is fully converged, and optimize the model hyperparameters based on the test set performance;
[0050] (5.3) Use the trained model to predict the traffic flow on the actual road at K future time points.
[0051] The beneficial effects of the present invention are:
[0052] (1) The Lanczos spatiotemporal network method for real-time traffic prediction of the present invention uses the Lanczos algorithm as a spectral domain convolution method in order to better capture long-distance dependencies, which is more adaptable to large-scale road networks with different spans and sparse distribution. The Lanczos algorithm transforms the spatial correlation matrix in graph theory into a low-dimensional correlation matrix based on similarity transformation. While reducing spatial complexity, it constructs a prediction model with stronger plasticity and better interpretability, significantly improving the prediction accuracy of the model.
[0053] (2) The method of the present invention can be applied to highways with large spans and long-range fluctuations. Under real-time real-environment data collection, the real-time traffic flow prediction can achieve reliable accuracy. The present invention trains the prediction network with the data collected in real time; it has strong robustness, and the graph neural prediction network used has strong robustness, strong stability and practicality; it has strong applicability. After the graph neural prediction network is trained, it can be directly deployed on the intelligent management and prediction platform without repeated training.
[0054] (3) The method of the present invention uses the construction of multi-dimensional structured features and establishes the spatial topological structure of the highway based on graph theory, constructs the adjacency matrix of the road network spatial information, and expands the Lanczos algorithm to calculate its features and eigenvalues to approximate the spectral domain filter of the graph convolution. By diagonalizing the approximated triangular symmetric matrix to capture long-distance dependencies, it can effectively save space computing power; use a recurrent neural network to capture time series features; then use a staggered network to ensure the stability of the network; finally, use a two-dimensional convolution module to obtain the prediction result, and then iteratively optimize it through the MAE loss function to finally complete the real-time traffic prediction of the road network. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 Schematic diagram of iterative changes in the loss function during the training process in Example 1.
[0056] Figure 2 Schematic diagram of the training process of the network model in the present invention.
[0057] Figure 3 This is a diagram of the Lanczos spatiotemporal graph convolutional network architecture of the present invention. DETAILED DESCRIPTION
[0058] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings. It should be pointed out that the specific implementation methods are only detailed descriptions of the present invention and should not be regarded as limitations of the present invention.
[0059] The present invention proposes a Lanczos spatiotemporal network method for real-time flow prediction, comprising the following steps:
[0060] Step 1: Construct multi-dimensional structured feature data and road network data;
[0061] Obtain flow and traffic situation data in real time.
[0062] The multi-dimensional structured data includes traffic situation and meteorological data. The traffic situation data includes but is not limited to data such as flow rate, speed, density, vehicle type ratio, etc.; the meteorological data includes but is not limited to weather conditions, temperature, etc.;
[0063] The road network data is composed of node data of the detector of the predicted target road network, and the road network information includes connectivity and distance between nodes.
[0064] In multidimensional structured feature data, assuming that a current time point is T, the prediction targets are T+t1, T+t2, ..., T+t k The flow rate at K time points. Δt=t k -t k-1 , For example, take Δt = 5min, K = 6, and the prediction duration is 30min.
[0065] Specifically, the construction of multi-dimensional structured feature data and road network data in step 1 is implemented by the following sub-steps:
[0066] (1.1) The traffic data is the historical traffic data of the Y period before time T on N road sections; the sampling time interval of the historical traffic data is uniform, the time interval is Δt, and the amount of sampled traffic data is Y / Δt;
[0067] (1.2) The traffic situation data is the historical data of the Y period before time T on N road sections and the corresponding predictions T+t1, T+t2, ..., T+t k The traffic situation data of K time points; the sampling time interval of historical data is uniform, the time interval is Δt, and the amount of sampled traffic data is Y / Δt;
[0068] (1.3) The topological relationship between the nodes in the road network is applicable to graph theory and is described as G t =(V t ,E,A), represents the state of the road network at time t, where V t =(v t:1 ;v t:2 ;...;v t:N ) represents the status of N road segments at time t, E represents the connectivity between these road segments, A is the adjacency matrix of the road network, and when the i and j road segments are connected, A i;j =1, otherwise 0, where A i;j =A j;i , that is, A is a symmetric matrix.
[0069] Step 2: Determine the network structure;
[0070] Data input layer: Use the spatiotemporal module as the feature input encoder; Spatiotemporal convolution layer: Input features are input into the spatial convolution layer and the temporal convolution layer respectively, and then residual operation is performed, and the results are returned to the output layer; Output layer: Use the predictor to output the last layer of data.
[0071] Specifically, the determination of the network structure described in step 2 is implemented through the following sub-steps:
[0072] (2.1) Construct the network data input-output layer; in the data input layer, use the spatiotemporal module as the feature input encoder; in the data output layer, use the two-dimensional convolutional basic unit as the predictor of the network output layer.
[0073] Table 1
[0074] Convolutional Layer Number of convolution kernels Convolution kernel size Whether to compensate for the bias Conv2D Prediction length + output dimension (1, hidden layer dimension) yes
[0075] (2.2) Constructing the residual module: Perform residual operation on the results returned by the spatiotemporal convolution layer, that is, the input of the network and the input of the spatiotemporal convolution layer are re-fused each time, and two-dimensional convolution is used as the fusion method of the residual module to achieve perception between the hidden layer and the original data.
[0076] Table 2
[0077] Convolutional Layer Number of convolution kernels Convolution kernel size Whether to compensate for the bias Conv2D Prediction length (1, input dimension + hidden layer dimension) yes
[0078] Step 3: Construct the spatial convolution layer;
[0079] Spatial convolution is implemented as the graph convolution kernel based on the Lanczos method. First, the eigenvalues and eigenvectors of the similarity transformation matrix are calculated according to the road network data. The isomorphic transformation between the similarity transformation matrix and the road network matrix is realized through a multi-layer perceptron to obtain different matrices for mining spatial association rules. The matrices are then combined with the multi-dimensional data through a graph convolution module to obtain feature sets of different association rules, which are then fused through linear transformations to obtain the output of the spatial module.
[0080] The spatial graph convolution layer based on the Lanczos method described in step 3 is implemented through the following sub-steps:
[0081] (3.1) Constructing affinity matrix From graph theory, we know that the degree of road section i is N represents the number of road sections, and the degree matrix D of the road network is obtained, where D ii =d i , and then get the Laplace matrix L, L = DA, and after symmetric normalization of L, get the affinity matrix:
[0082]
[0083] Here, I represents an identity matrix of the same type as A.
[0084] (3.2) Take the initial non-zero vector s and a positive integer M less than N and perform the Lanczos similarity transformation algorithm. The specific algorithm steps are as follows:
[0085]
[0086] Among them, z is an intermediate variable in the pseudocode and has no practical meaning.
[0087] According to the Lanczos algorithm results, the matrix V can be obtained from this set of basis vectors M =[v1,…,v M ], and a triangular symmetric matrix H M , whose diagonal and superdiagonal elements are {α1,α2,…,α M} and {β1,β2,…,β M}, its H M The construction rules are as follows:
[0088]
[0089] According to the nature of the algorithm, V M * V M The transpose of .
[0090] (3.3) to get H M Then, by its triangular symmetric matrix H M Perform eigendecomposition H M =BRB * , R, B, B * H M The decomposed eigenvalue matrix, eigenvector matrix and their transpose; the approximate eigenvector matrix V = V M B, therefore The approximate eigenvalues and eigenvectors of are
[0091] (3.4) For the graph signal, i.e., multidimensional structured feature data, let the multidimensional data feature input and the spatial convolution layer output be X t , S t , then the graph convolution of the road network G can be approximated as:
[0092]
[0093] The parameterized graph convolution is:
[0094] S t =[VR u V * X t ]W
[0095] Here, u represents the uth power, and W represents the learnable parameter weights for graph signal propagation in the spatial convolutional layer.
[0096] (3.5) Then, by expanding the eigenvalues, we can capture long-distance dependencies. Implement enhanced convolution filtering:
[0097]
[0098] Where W is the learnable parameter weight of the graph signal propagation in the spatial convolution layer, and E is the length of the associated node size set l. For example, l can be set to: l = [0, 1, 3, 5, 7, 10, 20, 50, 70]. The purpose is to capture the features of adjacent nodes with a distance of li. This way of capturing long-distance dependencies can effectively save space computing power.
[0099] Step 4: Construct the temporal convolution layer;
[0100] Apply GRU as the temporal convolution module, input the empty sequence feature data of step 3 into the gated recurrent unit, use the gating mechanism to filter the propagation features of long and short time series, and obtain the time series output through weight parameterization.
[0101] The construction of the temporal convolutional layer described in step 4 is implemented through the following sub-steps:
[0102] (4.1) The result of the spatial convolution layer S t As the input of the time module, the gated recurrent unit is used as its forward propagation method, which can well solve the long-term dependency. The forward propagation model formula is:
[0103]
[0104] Where σ is the activation function, [] represents tensor connection, * represents element multiplication, and · represents matrix product. z , W r , Represent the update gate, reset gate and learnable weights of the hidden layer of the temporal convolutional layer respectively.
[0105] Step 5: Realize real-time prediction;
[0106] The data obtained in step 4 is subjected to a two-dimensional convolution operation to obtain the prediction result, and the predicted target data is passed into the loss function. The iterator is used until the loss function value reaches an acceptable threshold, that is, the trainable parameters are changed, and the prediction results under different parameters are compared to obtain the best prediction model.
[0107] The real-time prediction described in step 5 is achieved through the following sub-steps:
[0108] (5.1) Design the optimizer and loss function of the model; the optimizer is the Adam optimizer, and the loss function uses the MAE (mean absolute error) suitable for the regression prediction function. The specific discriminant formula is:
[0109]
[0110] Where m represents the total number of samples, y n Represents the actual flow value of sample n; is the traffic prediction value of sample n.
[0111] (5.2) Obtain prediction results; train the prediction model to minimize the model loss function, iterate repeatedly until the model is fully converged, and optimize the model hyperparameters based on the test set performance;
[0112] (5.3) Use the trained model to predict the traffic flow on the actual road at K future time points. Specific embodiment 1
[0114] The actual application case is a highway in Zhejiang Province. Its data set contains the road network data of 229 sections. The time span is four months, with three months as the training data set and the last month as the test data set. The entire data set is traffic situation data (flow, speed, maximum speed, minimum speed and temperature) with a granularity of five minutes. This example uses one hour of traffic situation data and road network data to predict the flow of the next half hour. The MAE loss function is used, and the number of iterations is 30. The iterative changes of the loss function during the training process are as follows: Figure 1 As shown, Figure 1 In the figure, the train line represents the training set, and the val line represents the validation set. According to the change in loss, it can be seen that the loss of the model decreases rapidly and reaches stability after the 10th iteration.
[0115] The test set uses five indicators as evaluation indicators: MAE (mean absolute error), RMSE (root mean square error), MAPE (mean absolute percentage error), RRSE (relative square root error), and CORR (empirical correlation coefficient). The test results of each indicator are shown in the following table.
[0116] Table 3 Evaluation indicators
[0117]
[0118] The prediction results show the prediction results of the test set for one month, where each prediction step is 5 minutes and half an hour is 6 prediction steps. It can be seen that the closer the prediction step is to the measured data, the better the performance of various indicators; since real-time predictions are mostly based on short-term predictions such as five minutes, the test results perform best within five minutes (i.e. the first prediction step).
[0119] In addition, the prediction error changes of MAE, RMSE, MAPE, RRSE, and CORR indicators within half an hour of prediction are controlled within 12.53%, 13.45%, 8.5%, 13.5%, and 1%, respectively. From the perspective of stability affected by duration, the prediction results are within an acceptable range.
Claims
1. A Lanczos spatiotemporal network method for real-time traffic prediction, Its characteristics are: The following steps are involved: (1) Construct multi-dimensional structured feature data and road network data; Obtain flow and traffic situation data in real time; The multi-dimensional structured data includes traffic situation and meteorological data. The traffic situation data includes but is not limited to flow, speed, density, and vehicle type ratio data; the meteorological data includes but is not limited to weather conditions and temperature; the road network data is composed of node data of detectors of the predicted target road network; (2) Determine the network structure; Data input layer: Use the spatiotemporal module as the feature input encoder; Spatiotemporal convolution layer: Input features are input into the spatial convolution layer and the temporal convolution layer respectively, and then residual operation is performed, and the results are returned to the output layer; Output layer: Use the predictor to output the last layer of data; (3) Construct spatial convolution layer; Spatial convolution is implemented as the graph convolution kernel based on the Lanczos method. First, the eigenvalues and eigenvectors of the similarity transformation matrix are calculated according to the road network data. The isomorphic transformation between the similarity transformation matrix and the road network matrix is realized through a multi-layer perceptron to obtain different matrices for mining spatial association rules. The matrices are then combined with multi-dimensional data through a graph convolution module to obtain feature sets of different association rules. The spatial module output is obtained through a fusion of linear transformations. (4) Construct a temporal convolution layer; Use GRU as the temporal convolution module, input the empty sequence feature data of step (3) into the gated recurrent unit, use the gating mechanism to filter the propagation features of long and short time series, and obtain the time series output through weight parameterization; (5) Achieve real-time prediction; Perform a two-dimensional convolution operation on the data obtained in step 4 to obtain the prediction result, and pass it into the loss function together with the predicted target data. After passing through the iterator, the loss function value reaches an acceptable threshold. That is, the training parameters are changed, and the prediction results under different parameters are compared to obtain the best prediction model.
2. According to claim 1, a Lanczos spatiotemporal network method for real-time traffic prediction, Its characteristics are: In step (1), constructing multi-dimensional structured feature data and road network data is achieved by the following sub-steps: (1.1) The traffic data is the historical traffic data of the Y period before time T on N road sections; the sampling time interval of the historical traffic data is uniform, the time interval is Δt, and the amount of sampled traffic data is Y / Δt; (1.2) Traffic situation data is the historical data of Y time periods before time T on N road sections and the corresponding predictions T+t 1 , T+t 2 , ..., T+t k The traffic situation data of K time points; the sampling time interval of historical data is uniform, the time interval is Δt, and the amount of sampled traffic data is Y / Δt; (1.3) The topological relationship between the nodes in the road network is applicable to graph theory and is described as G t =(V t , E, A), represents the state of the road network at time t, where V t =(v t:1 ;v t:2 ;...;v t:N ) represents the status of N road segments at time t, E represents the connectivity between these road segments, A is the adjacency matrix of the road network, and when the i and j road segments are connected, A i;j =1, otherwise 0, where A i;j =A j;i , that is, A is a symmetric matrix.
3. According to claim 1, a Lanczos spatiotemporal network method for real-time traffic prediction, Its characteristics are: In step (2), determining the network structure is achieved through the following sub-steps: (2.1) Construct the network data input-output layer; in the data input layer, use the spatiotemporal module as the feature input encoder; in the data output layer, use the two-dimensional convolutional basic unit as the predictor of the network output layer; (2.2) Constructing the residual module: Perform residual operation on the results returned by the spatiotemporal convolution layer, that is, the input of the network and the input of the spatiotemporal convolution layer are re-fused each time, and two-dimensional convolution is used as the fusion method of the residual module to achieve perception between the hidden layer and the original data.
4. The Lanczos spatiotemporal network method for real-time traffic prediction according to claim 1, Its characteristics are: In step (3), constructing the spatial convolutional layer is achieved through the following sub-steps: (3.1) Constructing affinity matrix From graph theory, we know that the degree of road section i is N represents the number of road sections, and the degree matrix D of the road network is obtained, where D ii =d i , and then we get the Laplace matrix L = DA, and after symmetric normalization of L, we get the affinity matrix: Here, I represents the identity matrix of the same type as A; (3.2) Take the initial non-zero vector s and a positive integer M less than N, and perform the Lanczos similarity transformation algorithm. According to the results of the Lanczos algorithm, the matrix V can be obtained from this set of basis vectors M =[v 1 , ..., v M ], and a triangular symmetric matrix H M , whose diagonal and superdiagonal elements are {α 1 , α 2 , ..., α M } and {β 1 , β 2 , ..., β M }, its H M The construction rules are as follows: According to the nature of the algorithm, V M * V M The transpose of (3.3) to get H M Then, by its triangular symmetric matrix H M Perform eigendecomposition H M =BRB * , R, B, B * H M The decomposed eigenvalue matrix, eigenvector matrix and their transpose; the approximate eigenvector matrix V = V M B, therefore The approximate eigenvalues and eigenvectors of are (3.4) For the graph signal, i.e., multidimensional structured feature data, let the multidimensional data feature input and the spatial convolution layer output be X t , S t , then the graph convolution of the road network G can be approximated as: The parameterized graph convolution is: S t =[VR u V*X t ]W Here, u represents the uth power, and W represents the learnable parameter weights of the graph signal propagation in the spatial convolutional layer; (3.5) Then, by expanding the eigenvalues, we can capture long-distance dependencies. Implement enhanced convolution filtering: Where W is the learnable parameter weight of the graph signal propagation in the spatial convolution layer, and E is the length of the associated node size set l, which is intended to capture the features of adjacent nodes with a distance of li. This way of capturing long-distance dependencies can effectively save space computing power.
5. The Lanczos spatiotemporal network method for real-time traffic prediction according to claim 1, Its characteristics are: The construction of the temporal convolutional layer in step (4) is achieved through the following sub-steps: The result of the spatial convolution layer S t As the input of the time module, a gated recurrent unit is used as its forward propagation method, and the forward propagation model formula is: Where σ is the activation function, [] represents tensor connection, * represents element multiplication, and · represents matrix product; W z , W r , Represent the update gate, reset gate and learnable weights of the hidden layer of the temporal convolutional layer respectively.
6. A Lanczos spatiotemporal network method for real-time traffic prediction according to claim 1, Its characteristics are: In step (5), the real-time prediction is achieved through the following sub-steps: (5.1) Design the optimizer and loss function of the model; the optimizer is the Adam optimizer, and the loss function uses the MAE (mean absolute error) suitable for the regression prediction function. The specific discriminant formula is: Where m represents the total number of samples, y n Represents the actual flow value of sample n; is the flow prediction value of sample n; (5.2) Obtain prediction results; train the prediction model to minimize the model loss function, iterate repeatedly until the model is fully converged, and optimize the model hyperparameters based on the test set performance; (5.3) Use the trained model to predict the traffic flow on the actual road at K future time points.
Citation Information
Patent Citations
Short-term traffic flow control method based on graph convolution recurrent neural network
CN110827544A
Traffic flow prediction method and system for urban road network
CN112183862A
Deep neural network robust traffic prediction method based on multi-modal spatio-temporal data
CN112289034A
Traffic prediction method based on multi-scale graph convolutional network model
CN113053115A
Traffic speed prediction method based on space-time diagram convolution-generative adversarial network
CN113450561A
