CN116258258A - Multi-information fusion space-time diagram convolution traffic flow prediction method - Google Patents
Multi-information fusion space-time diagram convolution traffic flow prediction method Download PDFInfo
- Publication number
- CN116258258A CN116258258A CN202310137926.4A CN202310137926A CN116258258A CN 116258258 A CN116258258 A CN 116258258A CN 202310137926 A CN202310137926 A CN 202310137926A CN 116258258 A CN116258258 A CN 116258258A
- Authority
- CN
- China
- Prior art keywords
- matrix
- time
- traffic flow
- space
- adjacent
- 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
- 238000010586 diagram Methods 0.000 title claims abstract description 43
- 230000004927 fusion Effects 0.000 title claims abstract description 41
- 238000000034 method Methods 0.000 title claims abstract description 33
- 239000011159 matrix material Substances 0.000 claims abstract description 111
- 238000003062 neural network model Methods 0.000 claims abstract description 21
- 238000012360 testing method Methods 0.000 claims abstract description 9
- 238000012549 training Methods 0.000 claims abstract description 9
- 238000012545 processing Methods 0.000 claims abstract description 6
- 230000000737 periodic effect Effects 0.000 claims description 22
- 238000013528 artificial neural network Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 12
- 230000007246 mechanism Effects 0.000 claims description 9
- 238000005096 rolling process Methods 0.000 claims description 9
- 230000009466 transformation Effects 0.000 claims description 9
- 238000010276 construction Methods 0.000 claims description 8
- 230000003442 weekly effect Effects 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 230000003213 activating effect Effects 0.000 claims description 3
- 230000002776 aggregation Effects 0.000 claims description 3
- 238000004220 aggregation Methods 0.000 claims description 3
- 238000004880 explosion Methods 0.000 claims description 3
- 230000015654 memory Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 230000008447 perception Effects 0.000 claims description 3
- 238000012546 transfer Methods 0.000 claims description 3
- BUCXEFZXWKUCCY-UHFFFAOYSA-N 4-methyl-3-(2-phenylethyl)-1,2,4-oxadiazol-5-one Chemical compound O1C(=O)N(C)C(CCC=2C=CC=CC=2)=N1 BUCXEFZXWKUCCY-UHFFFAOYSA-N 0.000 abstract description 3
- 238000002229 photoelectron microspectroscopy Methods 0.000 abstract description 3
- 238000002474 experimental method Methods 0.000 description 7
- 238000013527 convolutional neural network Methods 0.000 description 5
- 238000013461 design Methods 0.000 description 4
- 238000005065 mining Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000000306 recurrent effect Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000009792 diffusion process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000008676 import Effects 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000000547 structure data Methods 0.000 description 2
- 238000011144 upstream manufacturing Methods 0.000 description 2
- 241000288105 Grus Species 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000014509 gene expression Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
-
- 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
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Tourism & Hospitality (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Development Economics (AREA)
- Health & Medical Sciences (AREA)
- Educational Administration (AREA)
- Primary Health Care (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Quality & Reliability (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Operations Research (AREA)
- Mathematical Physics (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a multi-information fusion space-time diagram convolution traffic flow prediction method, which comprises the following steps: firstly, constructing a graph adjacency matrix by considering distance factors, then constructing a fusion time sequence by considering periodicity factors, designing a depth space-time graph convolution gating neural network model, inputting the depth space-time graph convolution gating neural network model, wherein the input of the depth space-time graph convolution gating neural network model comprises two parts of historical traffic flow data and a space road network adjacency relation, obtaining the fusion time sequence and the graph adjacency matrix through data processing, and then transmitting the fusion time sequence and the graph adjacency matrix into the depth space-time graph convolution gating neural network model for training and testing; and finally, predicting and outputting the traffic flow through the trained model to complete the prediction of the traffic flow. The invention tests on a real road traffic flow PEMS data set; the result shows that the prediction error of the designed model is lower than that of the current prediction method based on graph convolution, and the prediction performance is better.
Description
Technical Field
The invention belongs to the technical field of forming dies, and particularly relates to a multi-information fusion space-time diagram convolution traffic flow prediction method.
Background
The urban road network is complicated, and scientific and accurate traffic flow prediction plays an important role in the fields of road congestion judgment, path planning and the like. The traffic flow has stronger periodicity and is closely connected with the state at the last moment. Some researchers pay attention to the logic relation of the sequence in the time dimension, and the traffic flow is predicted through a historical average method, a time sequence model, a KNN algorithm, an SVM algorithm, an RNN network and other machine learning models.
In practice, purely considering the time-dimensional relationship is more limited. A difficulty with accurate traffic flow predictions is that the current state of one road segment can affect the future state of neighboring road segments at the same time. By adding space dimension information, reasonable analysis of space-time characteristic information of traffic flow has great influence on improving prediction accuracy of the model. ConvLSTM concepts have been proposed to replace part of the Hadamard product operation in LSTM networks by convolution operations, thereby mining spatial information and reducing prediction errors. The other method is to combine the CNN network which extracts the features in the two-dimensional space with the RNN network related to the sequence prediction, mine the space feature information through CNN and combine the RNN network to predict. Both methods are applied to the traditional two-dimensional space-time matrix, and in fact, only the influence of the spatial relationship between the upstream road section and the downstream road section can be considered, so that the method is not well applicable to the complicated road environment.
In order to characterize complex adjacencies between road segments, a complex topology can be stored in the form of a graph. The appearance of the graph convolutional neural network (Graph Convolutional Network, GCN) lays a foundation for the deep learning model to mine the characteristic information of the graph structure. The graph convolution recursive network GCRN is proposed, and the graph convolution network is combined with the cyclic neural network, so that a theoretical basis is provided for mining space-time sequence characteristics. DCRNN (Difusion Convolutional Recurrent Neural Network, DCRNN) networks have also been proposed to model the trend of traffic flow as a diffuse process on a directed graph, with spatial features acquired by bi-directional random mobility of the graph. The traffic flow has a strong cycle rule in the daily and weekly cycle, a researcher designs a multi-component space-time diagram convolution network in consideration of the periodic characteristic of the traffic flow, and training and prediction are respectively carried out on the sub-components of the daily, weekly and the weekly during passing, and finally the prediction results of the three components are weighted and combined to obtain final output.
The space-time characteristics of mining traffic flow based on graph structures have become research hotspots, but the current model constructed based on graph neural networks has some problems:
1) In the storage of the graph adjacency matrix, the traditional mode only considers whether the road sections are adjacent or not, and the problem of different spatial correlation caused by the distance between the nodes is ignored.
2) On a network architecture, a space diagram convolution module and a time sequence module are designed to respectively extract characteristics of traffic flow in a space dimension and a time dimension; and extracting the traffic flow space characteristic information through a space diagram convolution module, and mining the time-dimensional correlation of the traffic flow through a time sequence prediction module.
3) In consideration of the problem of periodicity factors, the conventional model design, day and Zhou Duli components are respectively trained and tested, and finally a plurality of sub-component results are weighted and output, so that the time and space overhead cost is increased.
Based on the method, a space-time diagram convolution traffic flow prediction method with multiple information fusion is provided, a network architecture is improved, correlation operation is carried out on characteristic information extracted by a space convolution module and a fusion period time sequence, and final output can be obtained by a single component prediction module.
Disclosure of Invention
The technical problem to be solved by the invention is to provide a multi-information fusion space-time diagram convolution traffic flow prediction method for solving the problems in the background art.
In order to solve the technical problems, the invention adopts the following technical scheme: a multi-information fusion space-time diagram convolution traffic flow prediction method comprises the following steps:
firstly, constructing a graph adjacency matrix by considering distance factors, then constructing a fusion time sequence by considering periodicity factors, designing a depth space-time graph convolution gating neural network model, inputting the depth space-time graph convolution gating neural network model, wherein the input of the depth space-time graph convolution gating neural network model comprises two parts of historical traffic flow data and a space road network adjacency relation, obtaining the fusion time sequence and the graph adjacency matrix through data processing, and then transmitting the fusion time sequence and the graph adjacency matrix into the depth space-time graph convolution gating neural network model for training and testing; and finally, predicting and outputting the traffic flow through the trained model to complete the prediction of the traffic flow.
Further, the distance factor is considered to construct a graph adjacency matrix to carry out traffic flow modeling based on a graph structure, namely a space road network G is defined, and the space road network G comprises a point set and an edge set:
G＝＜V,E＞ (1)
v is a point set, and each road section in the road network is taken as a node;
e is an edge set and represents the adjacent relation between road sections, and the data is stored in the form of an adjacent matrix; if two road sections are adjacent, the corresponding adjacent matrix value is set to be 1, otherwise, the corresponding adjacent matrix value is set to be 0, and if A, B is two adjacent road sections, the larger the distance between A, B is, the smaller the mutual influence degree is, and a graph adjacent matrix fusing distance factors is designed:
defining the weight ratio of the adjacent node relative to the self node by using alpha as an adjustment coefficient, wherein the range is (0, 1);
setting the main diagonal of the adjacent matrix as 1 to ensure that the self node has the maximum weightHeavy values, obtaining updated adjacency matrix
And based on the number of adjacent nodes of each node, an updated adjacent matrix is obtained>
A degree matrix D of (2);
the distance between road segments is considered to give different values to the adjacent matrix, so that the network model can differentially analyze the learning weight parameters and finally fuse the adjacent matrix constructed by the distance factors.
Further, the construction of the fusion time sequence taking into consideration the periodicity factor is specifically as follows: considering the daily and weekly periodicity of traffic flow, and combining the traffic flow of adjacent time periods to construct a fusion time sequence, and constructing inter-fusion sequence information T history ：
T H Representing a time series of adjacent time periods,
h is the length of the sequence of adjacent time periods,
T D is a daily cycle sequence;
dn is the number of the daily cycle sequences and represents the daily cycle sequence of the previous dn;
pre is the length of a daily cycle sequence, and is consistent with the length of a time sequence to be predicted;
T W for the periodic sequence, wn is the number of periodic sequences, and represents the periodic sequence of the previous wn, T is taken H 、T D 、T W Fused into a historical time series T history 。
Further, the depth space-time diagram convolutional gated neural network model is divided into two parts in total: the space map rolling module and the time sequence predicting module extract spatial characteristics among roads and networks through the space map rolling module, and the extracted spatial characteristic information is combined with time sequence related information to be transmitted into the time sequence predicting module for sequenceThe method comprises the steps of predicting, firstly, inputting a traffic flow sequence X as a first layer of the model by the model, adding an adjacent matrix A fused with distance factors into the GCN network together, and extracting road network space characteristics to obtain a space characteristic vector G' so as to analyze the space characteristics of traffic flow; the periodic time sequence T { x is added in the hierarchy h ,x d ,x w Matrix operation is carried out on the output h of the previous hidden layer to obtain the input x of the time sequence prediction module t And (3) inputting the predicted output Y into a GRU network model for time sequence prediction, and finally obtaining the final predicted output Y through a full connection layer.
Further, the spatial map convolution module combines the adjacency matrix a with the deep neural network to parse the spatiotemporal nature of traffic flow, and for each hidden layer, combines the adjacency matrix with a transfer function:
H (l+1) ＝f(H (l) ,A) (4)
by adjacency matrices A and H (l) Matrix multiplication is performed, and a weight matrix W is combined (l) Performing linear transformation of the weight multiplication, and performing nonlinear variation by activating function sigma to obtain input H of the next hidden layer (l+1) ：
f(H (l) ,A)＝σ(AH (l) W (l) ) (5)
The information aggregated in the nodes does not contain the characteristics of the self-loop, the self-loop is added in the graph, the influence of the self-loop is considered, and the matrix A is added with the identity matrix on matrix transformation to obtain an updated adjacent matrix
The graph convolution module acquires the adjacent node characteristics through continuous aggregation operation, when the number of adjacent nodes of a certain node is large, the numerical value is large in the characteristic characteristics, the model convergence is influenced, the symmetric normalization is carried out through the addition degree matrix D, the problem of data explosion caused by the multiplication of parameters among a plurality of hidden layers is prevented, and the graph convolution module enables
On the adjacency matrix representation, there are
A ij Matrix value d for row j column i i 、d j Representing the degree of the matrix in row i and column j; then there are:
the one-layer graph convolution can acquire the spatial characteristic information of the neighbor nodes in the graph, acquire the information of the secondary neighbor nodes for expanding the receptive field of the model, improve the spatial perception capability of the model, and adopt a two-layer graph convolution mechanism, namely
Finally, let G l ' +1 ＝f(H (l) And A), transmitting the data to a time sequence prediction module.
Further, the time series prediction module input includes a current input x t Hidden state h transferred from the previous node t-1 ，h t-1 Containing information about previous nodes, combined with x t And h t-1 Obtaining the output y of the current hidden node t And hidden state h passed to the next node t The method comprises the steps of carrying out a first treatment on the surface of the State h transmitted by the last t-1 And input x of the current node t To obtain two gating states, as shown in the following formula; where r controls the reset gating and z is the gating that controls the update.
The model can acquire the spatial characteristic information G in the road network through the upper layer spatial map convolution module, and the spatial characteristic information G and the upper hidden layer output h t-1 As the input of the time sequence prediction module, in order to make the model pay more attention to the periodic characteristic information, and simultaneously fuse the time sequence t, the multi-information fusion is used for characteristic extraction, and finally the characteristic extraction is input into a network unit;
redefining the gate control input in the module to achieve the aim of multi-information fusion input, wherein the input is shown in formulas 13 and 14;
r t ＝σ(w gr *G t +w hr *h t-1 +w tr *t t +b r ) (13)
z t ＝σ(w gz *G t +w hz *h t-1 +w tz *t t +b z ) (14)
w b is a bias term for the weight parameter to be learned;
respectively divide the space characteristic information G t Information h of hidden layer t Periodic information t t Multiplying the corresponding weight matrix w respectively, adding the bias term b, and performing nonlinear transformation through an activation function;
after the gating signal is obtained, the data after the gating reset is reset, and the data is scaled to the range of-1 to 1 through a tanh activation function, so that the data is obtained
Selective memory and forgetting of hidden states by update gating, (1-z) t )*h t-1 The hidden state is left behind selectively,
selectively memorizing the hidden state to finally obtain the unit output h t ；
The constructed depth space-time diagram convolution neural network captures complex space information in an urban road topological structure through a diagram convolution module, combines a gate control circulation unit to acquire time correlation of traffic flow, and finally obtains a final output sequence Y through a full connection layer to complete a traffic flow prediction task.
Compared with the prior art, the invention has the following advantages:
in the method provided by the invention, the distance factors between road segments are considered to give different weight values to the graph adjacent matrix, and a space graph convolution module and a time sequence prediction module are designed in model construction; extracting space characteristic information through a space diagram convolution module, and fusing periodic time sequence information to be transmitted to a time sequence prediction module; and (3) redesigning a gating mechanism of the GRU unit to achieve the aim of inputting multi-characteristic information, and finally obtaining a predicted output result. The experiment is carried out on a real road traffic flow PEMS data set; the result shows that the prediction error of the designed model is lower than that of the current prediction method based on graph convolution, and the prediction performance is better.
Drawings
FIG. 1 is a traffic flow modeling diagram based on a graph structure in an embodiment of the invention;
FIG. 2 is a diagram of the construction of an adjacency matrix in an embodiment of the present invention;
FIG. 3 is a time series diagram of a fusion cycle in an embodiment of the invention;
FIG. 4 is a schematic diagram of the overall architecture of a model in an embodiment of the invention;
FIG. 5 is a diagram of a network model architecture in an embodiment of the invention;
FIG. 6 is a graph showing the results of different models at different time steps in experimental examples of the present invention;
FIG. 7 shows an experimental example of the present invention
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Embodiment, as shown in fig. 1-5, the present invention provides a technical solution: a multi-information fusion space-time diagram convolution traffic flow prediction method comprises the following steps:
firstly, constructing a graph adjacency matrix by considering distance factors, then constructing a fusion time sequence by considering periodicity factors, designing a depth space-time graph convolution gating neural network model, inputting the depth space-time graph convolution gating neural network model, wherein the input of the depth space-time graph convolution gating neural network model comprises two parts of historical traffic flow data and a space road network adjacency relation, obtaining the fusion time sequence and the graph adjacency matrix through data processing, and then transmitting the fusion time sequence and the graph adjacency matrix into the depth space-time graph convolution gating neural network model for training and testing; and finally, predicting and outputting the traffic flow through the trained model to complete the prediction of the traffic flow.
The traditional traffic flow space-time two-dimensional matrix has the transverse dimension and the longitudinal dimension respectively representing a specific dimension of the time dimension and the space dimension. Although the local spatial features of the data can be extracted by convolution operations in the convolutional neural network, the convolutional neural network can only act on standard euclidean two-dimensional structure data. The storage mode of the two-dimensional matrix determines that the adjacent relation between road sections can only be a linear relation, and at most, the position relation between two adjacent road sections at the upstream and downstream of the road sections can only be recorded. Road network structure is increasingly complex, and because of the existence of road intersections and branches, a plurality of adjacent road sections of a certain road section exist. The traditional two-dimensional matrix storage mode cannot represent the complex topological relation in the current road network. The form of the graph structure can store the adjacency relation of each road section between road networks in a most visual way, so that traffic flow modeling is performed based on the graph structure, as shown in fig. 1:
constructing a graph adjacency matrix by considering distance factors, and modeling traffic flow based on a graph structure, namely defining a space road network G, wherein the space road network G comprises a point set and an edge set:
G＝＜V,E＞ (1)
v is a point set, and each road section in the road network is taken as a node;
e is an edge set and represents the adjacent relation between road sections, and the data is stored in the form of an adjacent matrix; if two road sections are adjacent, the corresponding adjacent matrix value is set to be 1, otherwise, the corresponding adjacent matrix value is set to be 0, and if A, B is two adjacent road sections, the larger the distance between A, B is, the smaller the mutual influence degree is, and a graph adjacent matrix fusing distance factors is designed:
defining the weight ratio of the adjacent node relative to the self node by using alpha as an adjustment coefficient, wherein the range is (0, 1);
Setting the main diagonal of the adjacent matrix as 1 to ensure that the self node has the maximum weight value to obtain an updated adjacent matrix
And based on the number of adjacent nodes of each node, an updated adjacent matrix is obtained>
A degree matrix D of (2);
the network model can differentially analyze the learning weight parameters by giving different values to the adjacency matrix by considering the distance between road segments, and finally fuses the adjacency matrix constructed by distance factors, and an adjacency matrix constructed by the distance factors is shown in fig. 2, wherein alpha is set to 0.9 in the example.
The traffic flow has strong periodicity, the daily periodic variation of regularity is presented in different time periods, and the weekly weekdays and weekends have slight differences. Taking the traffic flow day and week periodicity into consideration, and combining the traffic flows of adjacent time periods to construct a fusion time sequence, as shown in fig. 3, taking the traffic flow day and week periodicity into consideration, and combining the traffic flows of adjacent time periods to construct a fusion time sequence, and constructing inter-fusion sequence information T history ：
T H Representing a time series of adjacent time periods,
h is the length of the sequence of adjacent time periods,
T D is a daily cycle sequence;
dn is the number of the daily cycle sequences and represents the daily cycle sequence of the previous dn;
pre is the length of a daily cycle sequence, and is consistent with the length of a time sequence to be predicted;
T W for the periodic sequence, wn is the number of periodic sequences, and represents the periodic sequence of the previous wn, T is taken H 、T D 、T W Fused into a historical time series T history 。
The input of the designed depth space-time diagram convolution gate control neural network model comprises two parts of historical traffic flow data and a space road network adjacent relation, and a fusion time sequence T is obtained through relevant data processing history Adjoining the graph to the matrix a. Training and testing in the afferent depth space-time diagram convolution gating neural network modelTesting; finally, the prediction output of traffic flow is carried out through a trained model, and as shown in fig. 4, the depth space-time diagram convolution gating neural network model is divided into two parts in total: the space map rolling module and the time sequence prediction module are used for extracting spatial characteristics among roads and networks through the space map rolling module, the extracted spatial characteristic information is combined with time sequence related information and is transmitted to the time sequence prediction module for sequence prediction, the model firstly takes a traffic flow sequence X as a first layer of input of the model, and an adjacent matrix A fused with distance factors is added and is input to the GCN network for extracting the spatial characteristics of the roads and networks, so that a spatial characteristic vector G' is obtained, and the spatial characteristics of traffic flow are planing; the periodic time sequence T { x is added in the hierarchy h ,x d ,x w Matrix operation is carried out on the output h of the previous hidden layer to obtain the input x of the time sequence prediction module t And finally, obtaining a final prediction output Y through a full connection layer, wherein the network model structure diagram is shown in figure 5.
The complex logical relationships between road networks are preserved by the irregular graph structure, and the proposal of the graph convolution network makes it possible to extract features from the graph structure data. In the deep network model, an adjacency matrix a is combined with the deep neural network to parse the spatiotemporal nature of traffic flow. For each hidden layer, in combination with the adjacency matrix, there is a transfer function:
H (l+1) ＝f(H (l) ,A) (4)
by adjacency matrices A and H (l) Matrix multiplication is performed, and a weight matrix W is combined (l) Performing linear transformation of the weight multiplication, and performing nonlinear variation by activating function sigma to obtain input H of the next hidden layer (l+1) ：
f(H (l) ,A)＝σ(AH (l) W (l) ) (5)
The information aggregated in the nodes does not contain the characteristics of the self-loop, the self-loop is added in the graph, the influence of the self-loop is considered, and the matrix A is added with the identity matrix on matrix transformation to obtain an updated adjacent matrix
/>
The graph convolution module acquires the adjacent node characteristics through continuous aggregation operation, when the number of adjacent nodes of a certain node is large, the numerical value is large in the characteristic characteristics, the model convergence is influenced, the symmetric normalization is carried out through the addition degree matrix D, the problem of data explosion caused by the multiplication of parameters among a plurality of hidden layers is prevented, and the graph convolution module enables
On the adjacency matrix representation, there are
A ij Matrix value d for row j column i i 、d j Representing the degree of the matrix in row i and column j; then there are:
the one-layer graph convolution can acquire the spatial characteristic information of the neighbor nodes in the graph, acquire the information of the secondary neighbor nodes for expanding the receptive field of the model, improve the spatial perception capability of the model, and adopt a two-layer graph convolution mechanism, namely
Finally, let G l ' +1 ＝f(H (l) And A), transmitting the data to a time sequence prediction module.
TimeThe sequence module is designed based on the GRU, gate Recurrent Unit, GRU gating loop unit. The GRU input output architecture is similar to RNN. The input includes the current input x t Hidden state h transferred from the previous node t-1 ，h t-1 Including information about previous nodes. Binding x t And h t-1 The GRU obtains the output y of the current hidden node t And hidden state h passed to the next node t 。
The GRU contains two important gating mechanisms: reset gate, update gate. State h transmitted by the last t-1 And input x of the current node t To obtain two gating states as shown in the following equation.
Where r controls the reset gating and z is the gating that controls the update.
And the model can acquire the space characteristic information G in the road network through the upper layer of space diagram convolution module. Outputting the spatial characteristic information G and the previous hidden layer to h t-1 As input to the time series prediction module. In order to make the model pay more attention to the periodic characteristic information, and simultaneously fuse the time period sequence t, the multi-information fusion is used for characteristic extraction, and finally the characteristic extraction is input into the GRU network unit. Here, the gating inputs in the GRU modules are redefined to achieve the goal of multi-information fusion inputs, see equations 13, 14.
r t ＝σ(w gr *G t +w hr *h t-1 +w tr *t t +b r ) (13)
z t ＝σ(w gz *G t +w hz *h t-1 +w tz *t t +b z ) (14)
w is the weight parameter to be learnedB is the bias term. Respectively divide the space characteristic information G t Information h of hidden layer t Periodic information t t And multiplying the corresponding weight matrix w, adding the bias term b, and performing nonlinear transformation through an activation function.
After the gating signal is obtained, the data after the gating reset is reset, and the data is scaled to the range of-1 to 1 through a tanh activation function, so that the data is obtained
And selectively memorizing and forgetting the hidden state through updating the gate control. (1-z) t )*h t-1 The hidden state is left behind selectively,
selectively memorizing the hidden state to finally obtain the unit output h t 。
The constructed depth space-time diagram convolution neural network captures complex space information in an urban road topological structure through a diagram convolution module, combines a gate control circulation unit to acquire time correlation of traffic flow, and finally obtains a final output sequence Y through a full connection layer to complete a traffic flow prediction task.
Experimental example: the environment where the experiment is carried out is an InterCore i7-9700 x 8 processor, a 16G running memory, an RTX2060 display card, a CDUA 10.0 deep learning framework, a Cudnn deep learning network acceleration library, the network building is built based on a TensorFlow framework, and the programming environment is Python3.5.
In order to verify the performance of the model, the data set adopts a PEMS data set truly disclosed by a highway network in California, and each key value pair of the data set comprises three attributes { flow, current, speed }, which are three indexes of traffic flow, lane occupancy and vehicle speed respectively, and the sampling interval is 5min. We intercept the flow traffic flow values and access them as corresponding csv files. And in order to test the comprehensive performance of the model on different data sets, selecting PEMS03, PEMS04 and PEMS08 data sets for experiments. And meanwhile, intercepting the data sets of 33 partial detectors on the PEMS03 data set, namely PEMS03-33, and verifying the performance of the model under a small sample. The ratio of the experimental training set to the test set was divided into 8:2, the data set overview is shown in table 1.
Table 1 overview of experimental data set
And (3) for the data which is partially missing in the data set, utilizing the data of the adjacent sequence of the default values, and adopting a linear interpolation method to fill the default values.
Because of the large trend of traffic flow, the data input into the model is firstly subjected to z-score standardization processing, so that the convergence efficiency of the model is accelerated, and the final prediction result is subjected to inverse standardization to obtain output.
In the figure adjacency matrix construction, a corresponding adjacency matrix is constructed based on the detector adjacency positional relationship table as shown in table 2. "from" and "to" record the detector numbers of the start and stop of a certain road section, respectively, and "distance" represents the distance between road sections.
TABLE 2 detector position adjacency position relation table
Setting evaluation indexes and experimental parameters;
the evaluation index adopts average absolute error MAE (Mean Absolute Error, MAE) and root mean square error RMSE (Root Mean Squard Error, RMSE) commonly used in regression problem, and the calculation formula is shown as formula x.
The smaller the value is, the smaller the error between the representative model predicted value and the actual value is, and the higher the model prediction precision is.
Where n represents the length of the sequence to be predicted,
representing the true value of traffic flow at point in time i, y i Representing the traffic flow predictions of the model at point i.
The super-parameter setting is carried out, and the model super-parameter mainly comprises parameters such as GRU hidden units, input history sequence length, daily cycle sequence number, weekly cycle sequence number, learning rate, batch zie and the like. The influence of the number of hidden units of different GRUs on the model precision is set in an experiment, and the numerical value is finally selected to be 64, so that the model has better performance. And the influence of different learning rates on the model convergence speed and accuracy is verified, and finally the learning rate alpha=0.002 is selected, so that the convergence speed is higher on the premise of ensuring the model accuracy. The experiment verifies the influence of the number d of the daily cycle sequences and the number w of the weekly cycle sequences on the model precision, the experimental setting d and w are respectively valued in {1,2 and 3}, and finally, the model error is minimum when the result d=1 and w=3 is verified. The selection of the optimizers was based on comparing RMSPropOptimizer, adamOptimizer, adaGradOptimizer the convergence curves of the different optimizers, and selecting an adamoptimer optimizer combining RMSProp with gradient descent.
Experimental results and analysis:
in order to comprehensively judge the prediction precision of the model, selecting different time steps, taking 5min as a time interval, gradually adding the prediction time range from 5min to 1h, recording 12 groups of error expressions of the model on PEMS04 and PEMS08 data sets along with the increase of the prediction time, judging through RMSE and MAE indexes, and the experimental results are shown in table 3;
table 3 results of model on PEMS04, PEMS08 at different time steps
As can be seen from table 3, the RMSE and MAE indexes of the model gradually increase with increasing prediction time range, and the prediction performance of the model decreases. The reason for the result is that on the time dimension, the influence degree of the data to be predicted by the history adjacent data gradually becomes smaller and the correlation gradually decreases along with the increase of the time range of the same road section; secondly, in the space dimension, the data of the neighbor nodes continuously change along with the increase of the time range, so that the space correlation is more difficult to acquire; the two are overlapped to increase the training difficulty of the model, and long-term traffic flow prediction is more challenging.
The multi-model comparison experiment selects LSTM and GRU networks which are representative of time sequence prediction as reference models, and selects a graph rolling network model which is currently applied to traffic flow prediction, wherein the graph rolling network model comprises DCRNN, STGCN, MCSTGCN, STSGCN, ASTGCN, and the prediction performances of the graph rolling network model under PEMS04 and PEMS08 data sets are respectively compared.
In order to comprehensively judge the prediction performance of the models, 12 groups of data of each model within 5-60 min are recorded, the time step is 5min, the average calculation is carried out on the 12 groups of results, and finally the comprehensive prediction error result of the model is obtained.
LSTM: long and short term memory networks, the concept of an import gate mechanism (import gate, export gate, forget gate) was proposed to solve the long term dependency problem of RNN
GRU: the gating circulation unit is improved on the basis of LSTM, an input gate and a forget gate of LSTM are combined into an updating gate, a gating mechanism is achieved, and compared with LSTM parameters, the model is easier to converge.
DCRNN: and (3) a diffusion convolution recurrent neural network, and establishing diffusion convolution to replace matrix multiplication in the GRU network.
STGCN: and establishing a prediction model of pure convolution, designing a space convolution layer and a time convolution layer to perform feature extraction, and setting fewer parameters to achieve higher convergence speed in the past.
MCSTGCN: and (3) respectively predicting three independent convolution modules of the multi-component space-time diagram convolution network during design, day and week, and finally carrying out weighting operation on the prediction results of the three components to obtain final output.
STSGCN: and constructing a local space-time diagram, wherein the local space-time diagram consists of 3 adjacent time slices, and a plurality of STSGCN models are deployed in different time periods so as to solve the heterogeneity of long-time space-time network data.
ASTGCN: and introducing an attention mechanism into the time convolution and space convolution module, and finally obtaining final output through graph convolution operation.
Fig. 6 plots the corresponding error curves for different models over time steps. It can be seen from the graph that the prediction error of the model gradually increases with the time step.
Table 4 records the overall predicted performance of the 8 models at 5min-60min, averaged by summing the errors over all time steps, as the final error value for the model.
Table 4 errors of different models on PEMS04, PEMS08 datasets
Because spatial information is not considered, the LSTM and GRU two circulating neural networks can only capture the correlation of the time dimension, and compared with the network model constructed by other graph convolution, the error is larger; and the variability is more pronounced as the predicted time range increases. Meanwhile, the prediction errors of DCRNN and STGCN models considering the characteristic information of the space dimension and the time dimension are obviously reduced compared with LSTM and GRU; the MCSTGCN, ASTGCN taking into account the periodicity factor further reduces the prediction error in the traffic flow prediction problem by setting up the time, day, zhou Duli prediction component. The STSGCN provides a space-time synchronous modeling mechanism through the improvement of a network architecture, and shows great advantages. The model constructs a novel adjacency matrix by respectively setting a space diagram convolution module and a time sequence prediction module and simultaneously taking distance factors into consideration, and adds and fuses periodic information sequences at the network level, so that the prediction performance of the model has optimal performance. As can be seen from the graph, the advantages of the model are obvious within the time range of 5-15 min, but the differences between the model and other models are gradually reduced along with the increase of the time range. The time-space correlation of traffic flow in a long time range is more difficult to capture, and the traffic flow is one of the directions for continuous research and improvement in the future.
Distance factor influence experiments to test the influence of the improved adjacency matrix on model accuracy, the error performance of the adjacency matrix constructed by fusing distance factors and the adjacency matrix constructed by traditional unaccounted distance factors on four data sets of PEMS03, PEMS03-33, PEMS04 and PEMS08 is compared, and the experimental results are shown in table 5.
TABLE 5 influence of fusion distance factor construction adjacency matrix on model prediction error
The adjacent matrix constructed by fusing the distance factors is used for constructing the differential matrix by considering the distance factors, and the spatial characteristic information is extracted in a deeper level by learning different weight parameters during model training. As can be seen from Table 5, the prediction error of the model was reduced by 1.0% -2.0% for the RMSE indicator and 1.6% -3.3% for the MAE indicator over four different data sets compared to the adjacency matrix constructed without distance factors.
FIG. 7 is an exemplary graph of model predictions over different data sets. The blue dotted line represents the actual traffic flow value, and the red solid line represents the traffic flow value predicted by the model. It can be seen that the traffic flow value predicted by the model has a higher degree of fit with the true value.
According to the prediction method of the multi-information fusion space-time diagram convolution traffic flow prediction model, in the construction mode of the diagram adjacency matrix, the data definition form of the incoming model is improved by considering the road section node distance factors, so that the model can obtain adjacency node weight coefficients in a differentiated mode, and the prediction precision of the model is improved to a certain extent. In the design of a model network architecture, a single-component network architecture model is designed, and an improved depth space-time diagram convolution gating neural network model integrating distance and periodicity factors is provided. In comparison with other graph convolution network models, the comprehensive prediction error is minimum, and the rationality of the network architecture of the model is verified; the error of short-time traffic flow prediction within 15min can be further reduced, and the method is convenient and practical.
It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.
Claims (6)
1. A multi-information fusion space-time diagram convolution traffic flow prediction method is characterized by comprising the following steps:
firstly, constructing a graph adjacency matrix by considering distance factors, then constructing a fusion time sequence by considering periodicity factors, designing a depth space-time graph convolution gating neural network model, inputting the depth space-time graph convolution gating neural network model, wherein the input of the depth space-time graph convolution gating neural network model comprises two parts of historical traffic flow data and a space road network adjacency relation, obtaining the fusion time sequence and the graph adjacency matrix through data processing, and then transmitting the fusion time sequence and the graph adjacency matrix into the depth space-time graph convolution gating neural network model for training and testing; and finally, predicting and outputting the traffic flow through the trained model to complete the prediction of the traffic flow.
2. The method for predicting traffic flow by using space-time diagram convolution with multiple information fusion according to claim 1, wherein the distance factor construction diagram adjacency matrix is considered to perform traffic flow modeling based on a diagram structure, namely a space road network G is defined, and the space road network G comprises a point set and an edge set:
G＝＜V,E＞ (1)
v is a point set, and each road section in the road network is taken as a node;
e is an edge set and represents the adjacent relation between road sections, and the data is stored in the form of an adjacent matrix; if two road sections are adjacent, the corresponding adjacent matrix value is set to be 1, otherwise, the corresponding adjacent matrix value is set to be 0, and if A, B is two adjacent road sections, the larger the distance between A, B is, the smaller the mutual influence degree is, and a graph adjacent matrix fusing distance factors is designed:
defining the weight ratio of the adjacent node relative to the self node by using alpha as an adjustment coefficient, wherein the range is (0, 1);
setting the main diagonal of the adjacent matrix as 1 to ensure that the self node has the maximum weight value to obtain an updated adjacent matrix
And based on the number of adjacent nodes of each node, an updated adjacent matrix is obtained>
A degree matrix D of (2);
the distance between road segments is considered to give different values to the adjacent matrix, so that the network model can differentially analyze the learning weight parameters and finally fuse the adjacent matrix constructed by the distance factors.
3. The method for predicting the traffic flow by the time-space diagram convolution with multiple information fusion according to claim 2, wherein the construction of the fusion time sequence taking into consideration the periodicity factor is specifically as follows: considering the daily and weekly periodicity of traffic flow, and combining the traffic flow of adjacent time periods to construct a fusion time sequence, and constructing inter-fusion sequence information T history ：
T H Representing a time series of adjacent time periods,
h is the length of the sequence of adjacent time periods,
T D is a daily cycle sequence;
dn is the number of the daily cycle sequences and represents the daily cycle sequence of the previous dn;
pre is the length of a daily cycle sequence, and is consistent with the length of a time sequence to be predicted;
T W for the periodic sequence, wn is the number of periodic sequences, and represents the periodic sequence of the previous wn, T is taken H 、T D 、T W Fused into a historical time series T history 。
4. A multi-information fusion space-time diagram convolutional traffic flow prediction method according to claim 3, wherein the depth space-time diagram convolutional gating neural network model is divided into two parts in total: the space map rolling module and the time sequence prediction module are used for extracting spatial characteristics among roads and networks through the space map rolling module, the extracted spatial characteristic information is combined with time sequence related information and is transmitted to the time sequence prediction module for sequence prediction, the model firstly takes a traffic flow sequence X as a first layer of input of the model, and an adjacent matrix A fused with distance factors is added and is input to the GCN network for extracting the spatial characteristics of the roads and networks, so that a spatial characteristic vector G' is obtained, and the spatial characteristics of traffic flow are planing; the periodic time sequence T { x is added in the hierarchy h ,x d ,x w Matrix operation is carried out on the output h of the previous hidden layer to obtain the input x of the time sequence prediction module t And (3) inputting the predicted output Y into a GRU network model for time sequence prediction, and finally obtaining the final predicted output Y through a full connection layer.
5. The method of claim 4, wherein the spatial map convolution module combines the adjacency matrix a with the deep neural network to parse the spatiotemporal properties of the traffic flow, and for each hidden layer, combines the adjacency matrix with a transfer function:
H (l+1) ＝f(H (l) ,A) (4)
by adjacency matrices A and H (l) Matrix multiplication is performed, and a weight matrix W is combined (l) Performing linear transformation of the weight multiplication, and performing nonlinear variation by activating function sigma to obtain input H of the next hidden layer (l+1) ：
f(H (l) ,A)＝σ(AH (l) W (l) ) (5)
The information aggregated in the nodes does not contain the characteristics of the self-loop, the self-loop is added in the graph, the influence of the self-loop is considered, and the matrix A is added with the identity matrix on matrix transformation to obtain an updated adjacent matrix
The graph convolution module acquires the adjacent node characteristics through continuous aggregation operation, when the number of adjacent nodes of a certain node is large, the numerical value is large in the characteristic characteristics, the model convergence is influenced, the symmetric normalization is carried out through the addition degree matrix D, the problem of data explosion caused by the multiplication of parameters among a plurality of hidden layers is prevented, and the graph convolution module enables
On the adjacency matrix representation, there are
A ij Matrix value d for row j column i i 、d j Representing the degree of the matrix in row i and column j; then there are:
the one-layer graph convolution can acquire the spatial characteristic information of the neighbor nodes in the graph, acquire the information of the secondary neighbor nodes for expanding the receptive field of the model, improve the spatial perception capability of the model, and adopt a two-layer graph convolution mechanism, namely
Finally, let G' l+1 ＝f(H (l) And A), transmitting the data to a time sequence prediction module.
6. The method of claim 5, wherein the time series prediction module input comprises a current input x t Hidden state h transferred from the previous node t-1 ，h t-1 Containing information about previous nodes, combined with x t And h t-1 Obtaining the output y of the current hidden node t And hidden state h passed to the next node t The method comprises the steps of carrying out a first treatment on the surface of the State h transmitted by the last t-1 And input x of the current node t To obtain two gating states, as shown in the following formula; wherein r controls reset gating, z is gating to control updating;
the model can acquire the spatial characteristic information G in the road network through the upper layer spatial map convolution module, and the spatial characteristic information G and the upper hidden layer output h t-1 As the input of the time sequence prediction module, in order to make the model pay more attention to the periodic characteristic information, and simultaneously fuse the time sequence t, the multi-information fusion is used for characteristic extraction, and finally the characteristic extraction is input into a network unit;
redefining the gate control input in the module to achieve the aim of multi-information fusion input, wherein the input is shown in formulas 13 and 14;
r t ＝σ(w gr *G t +w hr *h t-1 +w tr *t t +b r ) (13)
z t ＝σ(w gz *G t +w hz *h t-1 +w tz *t t +b z ) (14)
w is a weight parameter to be learned, and b is a bias term;
respectively divide the space characteristic information G t Information h of hidden layer t Periodic information t t Multiplying the corresponding weight matrix w respectively, adding the bias term b, and performing nonlinear transformation through an activation function;
after the gating signal is obtained, the data after the gating reset is reset, and the data is scaled to the range of-1 to 1 through a tanh activation function, so that the data is obtained
Selective memory and forgetting of hidden states by update gating, (1-z) t )*h t-1 The hidden state is left behind selectively,
selectively memorizing the hidden state to finally obtain the unit output h t ；
The constructed depth space-time diagram convolution neural network captures complex space information in an urban road topological structure through a diagram convolution module, combines a gate control circulation unit to acquire time correlation of traffic flow, and finally obtains a final output sequence Y through a full connection layer to complete a traffic flow prediction task.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310137926.4A CN116258258A (en) | 2023-02-20 | 2023-02-20 | Multi-information fusion space-time diagram convolution traffic flow prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310137926.4A CN116258258A (en) | 2023-02-20 | 2023-02-20 | Multi-information fusion space-time diagram convolution traffic flow prediction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116258258A true CN116258258A (en) | 2023-06-13 |
Family
ID=86683889
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310137926.4A Pending CN116258258A (en) | 2023-02-20 | 2023-02-20 | Multi-information fusion space-time diagram convolution traffic flow prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116258258A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117116051A (en) * | 2023-10-25 | 2023-11-24 | 深圳市交投科技有限公司 | Intelligent traffic management system and method based on artificial intelligence |
| CN117744950A (en) * | 2024-01-24 | 2024-03-22 | 深圳宇翊技术股份有限公司 | Travel demand analysis method, device, equipment and storage medium |
| CN117935561A (en) * | 2024-03-20 | 2024-04-26 | 山东万博科技股份有限公司 | Intelligent traffic flow analysis method based on Beidou data |
| CN118094439A (en) * | 2024-04-19 | 2024-05-28 | 江苏苏商银行股份有限公司 | Bank abnormal transaction detection method, system and device |
| CN118193991A (en) * | 2024-05-16 | 2024-06-14 | 北京理工大学 | Method and system for predicting migration organisms based on meteorological factors embedded in neural network |
-
2023
- 2023-02-20 CN CN202310137926.4A patent/CN116258258A/en active Pending
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117116051A (en) * | 2023-10-25 | 2023-11-24 | 深圳市交投科技有限公司 | Intelligent traffic management system and method based on artificial intelligence |
| CN117116051B (en) * | 2023-10-25 | 2023-12-22 | 深圳市交投科技有限公司 | Intelligent traffic management system and method based on artificial intelligence |
| CN117744950A (en) * | 2024-01-24 | 2024-03-22 | 深圳宇翊技术股份有限公司 | Travel demand analysis method, device, equipment and storage medium |
| CN117744950B (en) * | 2024-01-24 | 2024-07-02 | 深圳宇翊技术股份有限公司 | Travel demand analysis method, device, equipment and storage medium |
| CN117935561A (en) * | 2024-03-20 | 2024-04-26 | 山东万博科技股份有限公司 | Intelligent traffic flow analysis method based on Beidou data |
| CN117935561B (en) * | 2024-03-20 | 2024-05-31 | 山东万博科技股份有限公司 | Intelligent traffic flow analysis method based on Beidou data |
| CN118094439A (en) * | 2024-04-19 | 2024-05-28 | 江苏苏商银行股份有限公司 | Bank abnormal transaction detection method, system and device |
| CN118193991A (en) * | 2024-05-16 | 2024-06-14 | 北京理工大学 | Method and system for predicting migration organisms based on meteorological factors embedded in neural network |
| CN118193991B (en) * | 2024-05-16 | 2024-07-23 | 北京理工大学 | Method and system for predicting migration organisms based on meteorological factors embedded in neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116258258A (en) | Multi-information fusion space-time diagram convolution traffic flow prediction method | |
| CN112801404B (en) | Traffic prediction method based on self-adaptive space self-attention force diagram convolution | |
| CN110223517B (en) | Short-term traffic flow prediction method based on space-time correlation | |
| CN113053115B (en) | Traffic prediction method based on multi-scale graph convolution network model | |
| Zhang et al. | Taxi demand prediction using parallel multi-task learning model | |
| CN107563122B (en) | Crime prediction method based on interleaving time sequence local connection cyclic neural network | |
| CN111489013A (en) | Traffic station flow prediction method based on space-time multi-graph convolution network | |
| CN113762595B (en) | Traffic time prediction model training method, traffic time prediction method and equipment | |
| CN112766597B (en) | Bus passenger flow prediction method and system | |
| CN114495507B (en) | Traffic flow prediction method integrating space-time attention neural network and traffic model | |
| CN115578851A (en) | Traffic prediction method based on MGCN | |
| CN112419711B (en) | Closed parking lot parking demand prediction method based on improved GMDH algorithm | |
| CN115148019A (en) | Early warning method and system based on holiday congestion prediction algorithm | |
| CN114944053A (en) | Traffic flow prediction method based on spatio-temporal hypergraph neural network | |
| CN110310479A (en) | A kind of Forecast of Urban Traffic Flow forecasting system and method | |
| CN115392554A (en) | Track passenger flow prediction method based on depth map neural network and environment fusion | |
| CN108446798A (en) | Urban population flow prediction method based on dual path space-time residual error network | |
| CN116259172A (en) | Urban road speed prediction method considering space-time characteristics of traffic network | |
| CN115828990A (en) | Time-space diagram node attribute prediction method for fused adaptive graph diffusion convolution network | |
| CN113516304A (en) | Space-time joint prediction method and device for regional pollutants based on space-time graph network | |
| Zhang et al. | Spatiotemporal residual graph attention network for traffic flow forecasting | |
| CN111524349B (en) | Context feature injected multi-scale traffic flow prediction model establishing method and using method | |
| CN117131979A (en) | Traffic flow speed prediction method and system based on directed hypergraph and attention mechanism | |
| CN116630113A (en) | SCC-RCE prediction unit-based prediction network and gridding PM2.5 concentration prediction method | |
| CN115063972B (en) | Traffic speed prediction method and system based on graph convolution and gating circulation unit |
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 |