CN114860715A - Lanczos space-time network method for predicting flow in real time - Google Patents
Lanczos space-time network method for predicting flow in real time Download PDFInfo
- Publication number
- CN114860715A CN114860715A CN202210273493.0A CN202210273493A CN114860715A CN 114860715 A CN114860715 A CN 114860715A CN 202210273493 A CN202210273493 A CN 202210273493A CN 114860715 A CN114860715 A CN 114860715A
- Authority
- CN
- China
- Prior art keywords
- time
- data
- prediction
- matrix
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/22—Indexing; Data structures therefor; Storage structures
- G06F16/2291—User-Defined Types; Storage management thereof
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2465—Query processing support for facilitating data mining operations in structured databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
- G08G1/0133—Traffic data processing for classifying traffic situation
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/065—Traffic control systems for road vehicles by counting the vehicles in a section of the road or in a parking area, i.e. comparing incoming count with outgoing count
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Fuzzy Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a Lanczos space-time network method for flow real-time prediction, which can be suitable for a highway with large span and long-range fluctuation and can achieve reliable accuracy for real-time flow prediction under the real-time real environment data acquisition. The invention trains a prediction network by data collected in real time; the robustness is strong, and the used graph neural prediction network has strong robustness, stability and practicability; the applicability is strong, and after the training of the neural prediction network of the graph is completed, the neural prediction network of the graph can be directly arranged on an intelligent control and prediction platform without repeated training.
Description
Technical Field
The invention relates to the field of deep learning, in particular to a Lanczos space-time network method for flow real-time prediction in an intelligent prediction management system of a highway.
Background
With continuous innovation and demand increase of intelligent traffic, the future traffic state is predicted by using an intelligent prediction algorithm, and a traffic manager can be prompted to adopt effective management and control means by perceiving the traffic state in advance, and traffic users can be guided to change travel behaviors or demands, so that traffic transport efficiency and travel experience are improved.
The real-time traffic flow prediction refers to the prediction of the real-time traffic flow state by linking the current traffic flow state with the state of the traffic flow for a period of time in the future. At present, some flow real-time prediction methods exist, but the methods have some defects, such as poor prediction accuracy and poor effectiveness, and particularly for a highway with large span and long-range fluctuation, the prediction accuracy is low, and the prediction data is unreliable.
Disclosure of Invention
Aiming at the situation, in order to overcome the defects of the prior art, the invention provides a Lanczos space-time network method for predicting the flow in real time.
In order to achieve the purpose, the invention provides the following technical scheme:
a Lanczos space-time network method for flow real-time prediction comprises the following steps:
(1) constructing multi-dimensional structural feature data and road network data;
acquiring flow and traffic situation data in real time;
the multi-dimensional structured data comprises traffic situation data and image data, and the traffic situation data comprises but is not limited to flow, speed, density and vehicle type ratio data; meteorological data including, but not limited to, weather conditions, temperature; the road network data is node data formed by detectors of a predicted target road network;
(2) determining a network structure;
a data input layer: using a spatiotemporal module as a feature input encoder; space-time convolutional layers: inputting the input characteristics into the space convolution layer and the time convolution layer respectively, and then performing residual operation to obtain a result and returning the result to the output layer; an output layer: using a predictor to output the data of the last layer;
(3) constructing a space convolution layer;
the method comprises the following steps that space convolution is realized as a graph convolution kernel based on a Lanczos method, firstly, a characteristic value and a characteristic vector of a similarity transformation matrix are calculated according to road network data, homotypic transformation between the similarity transformation matrix and the road network matrix is realized through a multilayer perceptron, matrixes of different mining space association rules are obtained, a characteristic set of different association rules is obtained through a graph convolution module with multidimensional data, and space module output is obtained through fusion of linear transformation;
(4) constructing a time convolution layer;
applying GRU as a time convolution module, inputting the empty sequence characteristic data in the step (3) into a gating circulation unit, screening long and short time sequence propagation characteristics by using a gating mechanism, and obtaining time sequence output through weight parameterization;
(5) real-time prediction is realized;
and D, performing two-dimensional convolution operation on the data obtained in the step four to obtain a prediction result, transmitting the prediction result and the prediction target data into a loss function, changing the training parameters until the loss function value reaches an acceptable threshold value through an iterator, and comparing the prediction results under different parameters to obtain the best prediction model.
Further, in the step (1), the construction of the multidimensional structural feature data and the road network data is realized by the following substeps:
(1.1) the flow data is historical flow data of a Y time period before T time on N road sections; sampling time intervals of the historical flow data are uniform, the time intervals are delta t, and the sampling flow data volume is Y/delta t;
(1.2) traffic situation data are historical data of Y time period before T time on N road sections and corresponding prediction T + T 1 ,T+t 2 ,...,T+t k The traffic situation data of K time points; sampling time intervals of the historical data are uniform, the time intervals are delta t, and the quantity of the sampled traffic data is Y/delta t;
(1.3) topological relation among road network nodes, applying map theory and describing 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 ) Showing the state of N road sections at the time t, E showing the connectivity among the road sections, A being the adjacent of the road networkConnect matrix, A when i communicates with j road section i;j 1, otherwise 0, wherein a i;j =A j;i I.e. a is a symmetric array.
Further, in the step (2), determining the network structure is realized by the following sub-steps:
(2.1) constructing a network data input-output layer; at a data input layer, a space-time module is used as a characteristic input coder; in the data output layer, a two-dimensional convolution basic unit is used as a predictor of the network output layer.
(2.2) constructing a residual module: and performing residual operation on a result returned by the space-time convolutional layer, namely performing data re-fusion on the input of the network and the input of the space-time convolutional layer every time, and realizing the perception between the hidden layer and the original data by adopting a fusion mode of taking two-dimensional convolution as a residual module.
Further, in the step (3), constructing the spatial convolution layer is realized by the following sub-steps:
(3.1) construction of affinity matrixKnowing the degree of the section i from map theoryN represents the number of the road sections to obtain a degree matrix D of the road network, wherein D ii =d i And further obtaining a Laplace matrix L ═ D-A, and obtaining an affinity matrix after symmetrically normalizing L:
here, I represents an identity matrix of the same type as a;
(3.2) taking an initial non-zero vector s and a positive integer M smaller than N, performing a Lanczos similarity transformation algorithm, and obtaining a matrix V from the set of basis vectors according to the Lanczos algorithm result M =[v 1 ,…,v M ]And a triangular symmetric matrix H M The diagonal and hyper-diagonal elements are { alpha [, respectively 1 ,α 2 ,…,α M And { beta ] 1 ,β 2 ,…,β M H of M The construction rules of (1) are as follows:
(3.3) obtaining H M Then, by making a triangular symmetrical matrix H M Performing characteristic decomposition H M =BRB * ,R、B、B * Are respectively H M The decomposed eigenvalue matrix, eigenvector matrix and transpose thereof; obtaining an approximate eigenvector matrix V ═ V M B, thereforeApproximate eigenvalues and eigenvectors of
(3.4) for the multi-dimensional structured feature data as the graph signal, X is respectively set as the multi-dimensional data feature input and the space convolution layer output t 、S t Then the graph convolution for road network G can be approximated as:
convolving the parameterized graph as:
S t =[VR u V * X t ]W
here, u represents the u-th power, and W represents the learnable parametric weight of the graph signal propagation in the spatial convolution layer;
(3.5) realizing long-distance dependency relationship capture by expanding characteristic values based on And (3) realizing the enhanced convolution filtering:
w is the learnable parameter weight of graph signal propagation in the space convolution layer, E is the length of the scale set l of the associated nodes, and the significance lies in capturing the adjacent node characteristics with the distance li.
Further, building the time convolution layer in step (4) is achieved by the following sub-steps:
the result S of the space convolution layer t As an input of the time module, a gated cyclic unit is used as a forward propagation mode thereof, and the forward propagation model formula is as follows:
wherein σ is an activation function, [ alpha ] is an activation function]Representing tensor connections, representing element multiplications, representing matrix products. W z 、W r 、Which represent the learnable weights of the update gate, reset gate, and hidden layer, respectively, of the time convolution layer.
Further, in the step (5), the real-time prediction is realized by the following sub-steps:
(5.1) designing an optimizer and a loss function of the model; the optimizer is an Adam optimizer, the loss function adopts MAE (mean absolute error) suitable for a regression prediction function, and the specific discrimination formula is as follows:
wherein m represents the total number of samples, y n Representing the true flow value of sample n;predicting the flow of the sample n;
(5.2) obtaining a prediction result; training a prediction model to minimize a model loss function, repeating iteration until the model is completely converged, and simultaneously testing performance optimization model hyperparameters according to a test set;
and (5.3) predicting the traffic flow of the actual road at K time points in the future by using the trained model.
The invention has the beneficial effects that:
(1) according to the Lanczos space-time network method for predicting the flow in real time, disclosed by the invention, in order to better capture the dependency relationship between long distances, a Lanczos algorithm is used as a spectrum domain convolution mode, so that the Lanczos space-time network method is more suitable for large-scale road networks with different spans and sparse distribution; the Lanczos algorithm transforms the spatial incidence matrix in the graph theory into a low-dimensionality incidence matrix according to the similarity transformation, and constructs a prediction model with stronger plasticity and better interpretability while realizing reduction of spatial complexity, thereby remarkably improving the prediction accuracy of the model.
(2) The method can be suitable for the highway with large span and long-range fluctuation, and achieves reliable accuracy for real-time flow prediction under the real-time real environment data acquisition. The invention trains a prediction network by data collected in real time; the robustness is strong, and the used graph neural prediction network has strong robustness, stability and practicability; the applicability is strong, and after the training of the neural prediction network of the graph is completed, the neural prediction network of the graph can be directly arranged on an intelligent control and prediction platform without repeated training.
(3) The method establishes a highway space topological structure by means of the construction of multi-dimensional structural characteristics and based on a map theory, constructs an adjacent matrix of road network space information, develops a Lanczos algorithm to calculate the characteristics and characteristic values of the adjacent matrix to approximate a spectrum domain filter of graph convolution, and captures a long-distance dependency relationship by diagonalizing an approximated triangular symmetric matrix, so that space calculation force can be effectively saved; capturing timing features using a recurrent neural network; then, the stability of the network is ensured by using the staggered network; and finally, obtaining a prediction result by adopting a two-dimensional convolution module, and performing MAE loss function iterative optimization to finally complete the real-time traffic prediction of the road network.
Drawings
Fig. 1 is a schematic diagram of iterative changes of a loss function in a training process in example 1.
FIG. 2 is a schematic diagram of a training process of a network model according to the present invention.
FIG. 3 is a diagram of the Lanczos space-time graph convolution network architecture of the present invention.
Detailed Description
The technical solutions of the present invention are further described in detail below with reference to the accompanying drawings, and it should be noted that the detailed description is only for describing the present invention, and should not be construed as limiting the present invention.
The invention provides a Lanczos space-time network method for predicting flow in real time, which comprises the following steps:
the method comprises the following steps: constructing multi-dimensional structural feature data and road network data;
and acquiring data of flow and traffic situation in real time.
The multi-dimensional structured data comprises traffic situation data and image data, and the traffic situation data comprises data such as but not limited to flow, speed, density, vehicle type ratio and the like; meteorological data including, but not limited to, weather conditions, temperature, etc.;
the road network data is node data formed by detectors of a predicted target road network, and road network information comprises connectivity among nodes, distance and the like.
In the multi-dimensional structured feature data, assuming that a certain current time point is T, the prediction target is T + T 1 ,T+t 2 ,...,T+t k K time points of (a). Δ t ═ t k -t k-1 For example, the predicted time period is 30min when Δ t is 5min and K is 6.
Specifically, the step one of constructing the multidimensional structured feature data and the road network data is realized by the following substeps:
(1.1) the flow data is historical flow data of a Y time period before T time on N road sections; sampling time intervals of the historical flow data are uniform, the time intervals are delta t, and the sampling flow data volume is Y/delta t;
(1.2) traffic situation data are historical data of Y time period before T time on N road sections and corresponding prediction T + T 1 ,T+t 2 ,...,T+t k The traffic situation data of K time points; sampling time intervals of the historical data are uniform, the time intervals are delta t, and the quantity of the sampled traffic data is Y/delta t;
(1.3) topological relation among road network nodes, applying map theory and describing 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 ) Representing the state of N road sections at the time t, E representing the connectivity among the road sections, A being an adjacent matrix of a road network, A being a road section when i is communicated with j i;j 1, otherwise 0, wherein a i;j =A j;i I.e. a is a symmetric array.
Step two: determining a network structure;
a data input layer: using a spatiotemporal module as a feature input encoder; space-time convolutional layers: inputting the input characteristics into the space convolution layer and the time convolution layer respectively, and then performing residual operation to obtain a result and returning the result to the output layer; an output layer: and (4) performing last-layer data output by using a predictor.
Specifically, the determining of the network structure in step two is implemented by the following sub-steps:
(2.1) constructing a network data input-output layer; at a data input layer, a space-time module is used as a characteristic input coder; in the data output layer, a two-dimensional convolution basic unit is used as a predictor of the network output layer.
TABLE 1
Convolutional layer | Number of convolution kernels | Convolution kernel size | Whether to offset compensation |
Conv2D | Prediction length + output dimension | (1, hidden layer dimension) | Is that |
(2.2) constructing a residual module: and performing residual operation on a result returned by the space-time convolutional layer, namely performing data re-fusion on the input of the network and the input of the space-time convolutional layer every time, and realizing the perception between the hidden layer and the original data by adopting a fusion mode of taking two-dimensional convolution as a residual module.
TABLE 2
Convolutional layer | Number of convolution kernels | Convolution kernel size | Whether to offset compensation |
Conv2D | Predicted length | (1, input dimension + hiddenDimension of Tibetan layer) | Is that |
Step three: constructing a space convolution layer;
the spatial convolution is realized by taking a Lanczos method as a graph convolution kernel, firstly calculating characteristic values and characteristic vectors of a similarity transformation matrix according to road network data, realizing homotypic transformation between the similarity transformation matrix and the road network matrix through a multilayer perceptron, obtaining matrixes of different mining spatial association rules, obtaining characteristic sets of different association rules through a graph convolution module together with multidimensional data, and obtaining spatial module output through fusion of linear transformation.
The space map convolutional layer based on the Lanczos method in the third step is realized by the following sub-steps:
(3.1) construction of affinity matrixKnowing the degree of the section i from map theoryN represents the number of the road sections to obtain a degree matrix D of the road network, wherein D ii =d i And further obtaining a Laplace matrix L, wherein L is D-A, and obtaining an affinity matrix after symmetrically normalizing L:
here, I represents an identity matrix of the same type as a.
(3.2) taking the initial non-zero vector s and a positive integer M smaller than N, and performing a Lanczos similarity transformation algorithm, wherein the specific algorithm steps are as follows:
wherein z is an intermediate variable of the pseudo code, and has no practical significance.
From the set of basis vectors, a matrix V is derived based on the Lanczos algorithm results M =[v 1 ,…,v M ]And a triangular symmetric matrix H M The diagonal and hyper-diagonal elements are { alpha [, respectively 1 ,α 2 ,…,α M And { beta ] 1 ,β 2 ,…,β M H of M The construction rules of (1) are as follows:
(3.3) obtaining H M Then, by making a triangular symmetrical matrix H M Performing characteristic decomposition H M =BRB * ,R、B、B * Are respectively H M The decomposed eigenvalue matrix, eigenvector matrix and transpose thereof; obtaining an approximate eigenvector matrix V ═ V M B, thereforeApproximate eigenvalues and eigenvectors of
(3.4) for the multi-dimensional structured feature data as the graph signal, X is respectively set as the multi-dimensional data feature input and the space convolution layer output t 、S t Then the graph convolution for road net G can be approximated as:
convolving the parameterized graph as:
S t =[VR u V * X t ]W
here, u represents the u-th power, and W represents the learnable parametric weight of the graph signal propagation in the spatial convolution layer.
(3.5) realizing long-distance dependency relationship capture by expanding characteristic values based on And (3) realizing the enhanced convolution filtering:
where W is the learnable parameter weight of graph signal propagation in the spatial convolution layer, and E is the length of the associated node scale set l, and if l can be set as: i ═ 0,1,3,5,7,10,20,50,70], meaning that the neighboring node features are captured at a distance li, by which means distance-dependent capture can be used to save space computation efficiently.
Step four: constructing a time convolution layer;
and (3) applying GRU as a time convolution module, inputting the empty sequence characteristic data in the step three into a gating circulation unit, screening long and short time sequence propagation characteristics by using a gating mechanism, and obtaining time sequence output through weight parameterization.
The build-time convolutional layer described in step four is implemented by the following sub-steps:
(4.1) results S of spatial convolution layer t As an input of the time module, a gated cyclic unit is applied as a forward propagation mode, so that the long-term dependency can be well solved, and the forward propagation model formula is as follows:
wherein σ is an activation function [ [ alpha ] ]]Representing tensor connections, representing element multiplications, representing matrix products.W z 、W r 、Which represent the learnable weights of the update gate, reset gate, and hidden layer, respectively, of the time convolution layer.
Step five: real-time prediction is realized;
and D, performing two-dimensional convolution operation on the data obtained in the step four to obtain a prediction result, transmitting the prediction result and the prediction target data into a loss function, passing through an iterator until a loss function value reaches an acceptable threshold value, namely changing trainable parameters, and comparing the prediction results under different parameters to obtain the best prediction model.
The real-time prediction in the step five is realized by the following sub-steps:
(5.1) designing an optimizer and a loss function of the model; the optimizer is an Adam optimizer, the loss function adopts MAE (mean absolute error) suitable for a regression prediction function, and the specific discrimination formula is as follows:
wherein m represents the total number of samples, y n Representing the true flow value of sample n;and predicting the flow of the sample n.
(5.2) obtaining a prediction result; training a prediction model to minimize a model loss function, repeating iteration until the model is completely converged, and simultaneously testing performance optimization model hyperparameters according to a test set;
and (5.3) predicting the traffic flow of the actual road at K time points in the future by using the trained model.
Detailed description of the preferred embodiment 1
The practical application case is a certain highway in Zhejiang, the data set of the practical application case comprises road network data of 229 road sections, the time span is four months, three months are taken as a training data set, and the next month is taken as a testing data set. The whole data set is traffic situation data (flow, speed, maximum speed, minimum speed and temperature) of a five-minute granularity set meter, the traffic situation data of one hour and road network data are used for predicting the flow of the second half hour in the present example, an MAE loss function is adopted, the iteration times are 30 times, the iteration change of the loss function in the training process is shown in figure 1, a train line represents a training set, and a val line represents a verification set; from the change in loss, it can be seen that the loss of the model decreases very rapidly and reaches a plateau by iteration to the 10 th time.
The test set adopts five indexes of MAE (mean absolute error), RMSE (root mean square error), MAPE (mean absolute percentage error), RRSE (relative square root error) and CORR (empirical correlation coefficient) as evaluation indexes, and the test results of the indexes are shown in the following table.
TABLE 3 evaluation index
The prediction represents the prediction of the test set for one month, with 5 minutes for each prediction step and 6 prediction steps for half an hour. It can be seen that the closer the prediction step length is to the measured data, the better each index is represented; since real-time prediction is mostly targeted for short-time predictions such as five minutes, the test results perform best within five minutes (i.e., the first prediction step).
Besides, the prediction error change amplitudes of the MAE, RMSE, MAPE, RRSE and CORR indexes within half an hour of the prediction time are respectively controlled within 12.53%, 13.45%, 8.5%, 13.5% and 1%, and the prediction results are within an acceptable range in terms of stability influenced by the time.
Claims (6)
1. A Lanczos space-time network method for flow real-time prediction is characterized by comprising the following steps:
(1) constructing multi-dimensional structural feature data and road network data;
acquiring flow and traffic situation data in real time;
the multi-dimensional structured data comprises traffic situation data and image data, and the traffic situation data comprises but is not limited to flow, speed, density and vehicle type ratio data; meteorological data including, but not limited to, weather conditions, temperature; the road network data is node data formed by detectors of a predicted target road network;
(2) determining a network structure;
a data input layer: using a spatiotemporal module as a feature input encoder; space-time convolutional layers: inputting the input characteristics into the space convolution layer and the time convolution layer respectively, and then performing residual operation to obtain a result and returning the result to the output layer; an output layer: using a predictor to output the data of the last layer;
(3) constructing a space convolution layer;
the method comprises the following steps that space convolution is realized as a graph convolution kernel based on a Lanczos method, firstly, a characteristic value and a characteristic vector of a similarity transformation matrix are calculated according to road network data, homotypic transformation between the similarity transformation matrix and the road network matrix is realized through a multilayer perceptron, matrixes of different mining space association rules are obtained, a characteristic set of different association rules is obtained through a graph convolution module with multidimensional data, and space module output is obtained through fusion of linear transformation;
(4) constructing a time convolution layer;
applying GRU as a time convolution module, inputting the empty sequence characteristic data in the step (3) into a gating circulation unit, screening long and short time sequence propagation characteristics by using a gating mechanism, and obtaining time sequence output through weight parameterization;
(5) real-time prediction is realized;
and D, performing two-dimensional convolution operation on the data obtained in the step four to obtain a prediction result, transmitting the prediction result and the prediction target data into a loss function, changing the training parameters until the loss function value reaches an acceptable threshold value through an iterator, and comparing the prediction results under different parameters to obtain the best prediction model.
2. The Lanczos spatio-temporal network method for real-time traffic prediction as claimed in claim 1, wherein in the step (1), the construction of the multidimensional structural feature data and the road network data is realized by the following sub-steps:
(1.1) the flow data is historical flow data of a Y time period before T time on N road sections; sampling time intervals of the historical flow data are uniform, the time intervals are delta t, and the sampling flow data volume is Y/delta t;
(1.2) traffic situation data are historical data of Y time period before T time on N road sections and corresponding prediction T + T 1 ,T+t 2 ,...,T+t k The traffic situation data of K time points; sampling time intervals of the historical data are uniform, the time intervals are delta t, and the quantity of the sampled traffic data is Y/delta t;
(1.3) topological relation among road network nodes, applying map theory and describing 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 ) Representing the state of N road sections at the time t, E representing the connectivity among the road sections, A being an adjacent matrix of a road network, A being a road section when i is communicated with j i;j 1, otherwise 0, wherein a i;j =A j;i I.e. a is a symmetric array.
3. The Lanczos spatio-temporal network method for traffic real-time prediction according to claim 1, wherein in the step (2), the determining of the network structure is implemented by the following sub-steps:
(2.1) constructing a network data input-output layer; at a data input layer, a space-time module is used as a characteristic input coder; in the data output layer, a two-dimensional convolution basic unit is used as a predictor of the network output layer.
(2.2) constructing a residual module: and performing residual operation on a result returned by the space-time convolutional layer, namely performing data re-fusion on the input of the network and the input of the space-time convolutional layer every time, and realizing the perception between the hidden layer and the original data by adopting a fusion mode of taking two-dimensional convolution as a residual module.
4. The Lanczos spatio-temporal network method for real-time traffic prediction according to claim 1, wherein in the step (3), the construction of the spatial convolution layer is implemented by the following sub-steps:
(3.1) construction of affinity matrixKnowing the degree of the section i from map theoryN represents the number of the road sections to obtain a degree matrix D of the road network, wherein D ii =d i And further obtaining a Laplace matrix L ═ D-A, and obtaining an affinity matrix after symmetrically normalizing L:
here, I represents an identity matrix of the same type as a;
(3.2) taking an initial non-zero vector s and a positive integer M smaller than N, performing a Lanczos similarity transformation algorithm, and obtaining a matrix V from the set of basis vectors according to the Lanczos algorithm result M =[v 1 ,...,v M ]And a triangular symmetric matrix H M The diagonal and hyper-diagonal elements are { alpha [, respectively 1 ,α 2 ,...,α M And { beta ] 1 ,β 2 ,...,β M H of M The construction rules of (1) are as follows:
(3.3) obtaining H M Then, by making a triangular symmetrical matrix H M Performing characteristic decomposition H M =BRB * ,R、B、B * Are respectively H M The decomposed eigenvalue matrix, eigenvector matrix and transpose thereof; to obtainTo the approximated eigenvector matrix V ═ V M B, thereforeApproximate eigenvalues and eigenvectors of
(3.4) for the multi-dimensional structured feature data as the graph signal, X is respectively set as the multi-dimensional data feature input and the space convolution layer output t 、S t Then the graph convolution for road net G can be approximated as:
convolving the parameterized graph as:
S t =[VR u V*X t ]W
here, u represents the u-th power, and W represents the learnable parametric weight of the graph signal propagation in the spatial convolution layer;
(3.5) realizing long-distance dependency relationship capture by expanding characteristic values based on And (3) realizing the enhanced convolution filtering:
w is the learnable parameter weight of graph signal propagation in the space convolution layer, E is the length of the scale set l of the associated nodes, and the significance lies in capturing the adjacent node characteristics with the distance li.
5. The Lanczos spatio-temporal network method for real-time traffic prediction according to claim 1, wherein the time convolution layer is constructed in the step (4) by the following sub-steps:
the result S of the space convolution layer t As an input of the time module, a gated cyclic unit is used as a forward propagation mode thereof, and the forward propagation model formula is as follows:
wherein σ is an activation function, [ alpha ] is an activation function]Representing tensor connections, representing element multiplications, representing matrix products. W z 、W r 、Which represent the learnable weights of the update gate, reset gate, and hidden layer, respectively, of the time convolution layer.
6. The Lanczos spatio-temporal network method for traffic real-time prediction according to claim 1, wherein in the step (5), the real-time prediction is realized by the following sub-steps:
(5.1) designing an optimizer and a loss function of the model; the optimizer is an Adam optimizer, the loss function adopts MAE (mean absolute error) suitable for a regression prediction function, and the specific discrimination formula is as follows:
wherein m represents the total number of samples, y n Representing the true flow value of sample n;predicting the flow of the sample n;
(5.2) obtaining a prediction result; training a prediction model to minimize a model loss function, repeating iteration until the model is completely converged, and simultaneously testing performance optimization model hyperparameters according to a test set;
and (5.3) predicting the traffic flow of the actual road at K time points in the future by using the trained model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210273493.0A CN114860715A (en) | 2022-03-18 | 2022-03-18 | Lanczos space-time network method for predicting flow in real time |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210273493.0A CN114860715A (en) | 2022-03-18 | 2022-03-18 | Lanczos space-time network method for predicting flow in real time |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114860715A true CN114860715A (en) | 2022-08-05 |
Family
ID=82627232
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210273493.0A Pending CN114860715A (en) | 2022-03-18 | 2022-03-18 | Lanczos space-time network method for predicting flow in real time |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114860715A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116094758A (en) * | 2022-12-02 | 2023-05-09 | 肇庆医学高等专科学校 | Large-scale network flow acquisition method and system |
CN116192669A (en) * | 2023-03-07 | 2023-05-30 | 西安电子科技大学 | Network flow prediction method based on dynamic space-time diagram convolution |
CN117292551A (en) * | 2023-11-27 | 2023-12-26 | 辽宁邮电规划设计院有限公司 | Urban traffic situation adjustment system and method based on Internet of Things |
CN116192669B (en) * | 2023-03-07 | 2024-06-07 | 西安电子科技大学 | Network flow prediction method based on dynamic space-time diagram convolution |
-
2022
- 2022-03-18 CN CN202210273493.0A patent/CN114860715A/en active Pending
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116094758A (en) * | 2022-12-02 | 2023-05-09 | 肇庆医学高等专科学校 | Large-scale network flow acquisition method and system |
CN116094758B (en) * | 2022-12-02 | 2023-07-21 | 肇庆医学高等专科学校 | Large-scale network flow acquisition method and system |
CN116192669A (en) * | 2023-03-07 | 2023-05-30 | 西安电子科技大学 | Network flow prediction method based on dynamic space-time diagram convolution |
CN116192669B (en) * | 2023-03-07 | 2024-06-07 | 西安电子科技大学 | Network flow prediction method based on dynamic space-time diagram convolution |
CN117292551A (en) * | 2023-11-27 | 2023-12-26 | 辽宁邮电规划设计院有限公司 | Urban traffic situation adjustment system and method based on Internet of Things |
CN117292551B (en) * | 2023-11-27 | 2024-02-23 | 辽宁邮电规划设计院有限公司 | Urban traffic situation adjustment system and method based on Internet of things |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109492822B (en) | Air pollutant concentration time-space domain correlation prediction method | |
CN111612243B (en) | Traffic speed prediction method, system and storage medium | |
CN113053115B (en) | Traffic prediction method based on multi-scale graph convolution network model | |
CN113313947B (en) | Road condition evaluation method of short-term traffic prediction graph convolution network | |
CN112071065A (en) | Traffic flow prediction method based on global diffusion convolution residual error network | |
CN114860715A (en) | Lanczos space-time network method for predicting flow in real time | |
Dawson et al. | A comparison of artificial neural networks used for river forecasting | |
CN114299723B (en) | Traffic flow prediction method | |
CN112330951B (en) | Method for realizing road network traffic data restoration based on generation of countermeasure network | |
CN113379107A (en) | Regional ionized layer TEC forecasting method based on LSTM and GCN | |
CN115935796A (en) | Time-space heterogeneous and synchronous graph convolution network traffic flow prediction method | |
CN115759461A (en) | Internet of things-oriented multivariate time sequence prediction method and system | |
CN113808396A (en) | Traffic speed prediction method and system based on traffic flow data fusion | |
CN113947182A (en) | Traffic flow prediction model construction method based on double-stage stack graph convolution network | |
CN115862324A (en) | Space-time synchronization graph convolution neural network for intelligent traffic and traffic prediction method | |
CN114997506A (en) | Atmospheric pollution propagation path prediction method based on link prediction | |
CN115545334A (en) | Land use type prediction method, land use type prediction device, electronic device, and storage medium | |
CN112115754B (en) | Short-time traffic flow prediction model based on firework differential evolution hybrid algorithm-extreme learning machine | |
CN115327504B (en) | Sea clutter amplitude distribution non-typed prediction method based on measurement condition parameters | |
CN110839253A (en) | Method for determining wireless grid network flow | |
CN115953902A (en) | Traffic flow prediction method based on multi-view space-time diagram convolution network | |
CN114267170A (en) | Traffic flow prediction method based on graph space-time transform model considering human mobility | |
CN112446550B (en) | Short-term building load probability density prediction method | |
CN114004421B (en) | Traffic data missing value interpolation method based on space-time integrated learning | |
CN113469331B (en) | Vehicle tail gas prediction method and system based on global and local space-time diagram convolution |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |