CN112289034A - Deep neural network robust traffic prediction method based on multi-mode space-time data - Google Patents
Deep neural network robust traffic prediction method based on multi-mode space-time data Download PDFInfo
- Publication number
- CN112289034A CN112289034A CN202011587058.2A CN202011587058A CN112289034A CN 112289034 A CN112289034 A CN 112289034A CN 202011587058 A CN202011587058 A CN 202011587058A CN 112289034 A CN112289034 A CN 112289034A
- Authority
- CN
- China
- Prior art keywords
- matrix
- model
- data
- gcn
- traffic flow
- 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
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 25
- 238000001914 filtration Methods 0.000 claims abstract description 15
- 230000008859 change Effects 0.000 claims abstract description 7
- 238000012546 transfer Methods 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 68
- 230000008569 process Effects 0.000 claims description 12
- 238000012545 processing Methods 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 2
- 230000006870 function Effects 0.000 abstract description 13
- 239000013598 vector Substances 0.000 abstract description 4
- 238000005070 sampling Methods 0.000 abstract description 2
- 238000005065 mining Methods 0.000 abstract 1
- 238000007781 pre-processing Methods 0.000 abstract 1
- 238000013527 convolutional neural network Methods 0.000 description 11
- 238000012549 training Methods 0.000 description 9
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000010276 construction Methods 0.000 description 4
- 238000012795 verification Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 101150073669 NCAN gene Proteins 0.000 description 2
- 229910000831 Steel Inorganic materials 0.000 description 2
- 238000002790 cross-validation Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 239000010959 steel Substances 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 238000000844 transformation Methods 0.000 description 2
- 230000007704 transition Effects 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000011157 data evaluation Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Images
Classifications
-
- 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
- G08G1/0129—Traffic data processing for creating historical data or processing based on historical data
-
- 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
- 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
- G06N3/048—Activation functions
-
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention relates to a deep neural network robust traffic prediction method based on multi-mode space-time data, which comprises the steps of obtaining characteristic data of a traffic flow state at a certain time, constructing a characteristic data set according to a function, taking historical time sequence data as input, and capturing an urban road network topological structure by using a graph convolution network to obtain spatial characteristics; then inputting a gated recursion unit model, acquiring dynamic change through information transfer among units, and capturing time characteristics; the method comprises the steps of obtaining a T-GCN output prediction result in a full connection layer mode, obtaining a final output prediction result by performing Kalman filtering on an obtained output data set, sequentially performing data preprocessing and feature learning on an input vector of traffic flow data by adopting a time chart convolution network model, correcting an output result by using Kalman filtering, mining essential rules in the traffic flow data, and finally sampling the model to obtain high-level feature vectors to predict the state of traffic flow.
Description
Technical Field
The invention relates to the field of traffic control, in particular to a deep neural network robust traffic prediction method based on multi-mode space-time data.
Background
Traffic flow prediction aims at estimating the traffic flow in a certain space-time range, and is a core problem in the research of modern intelligent traffic systems. A reliable, accurate and durable traffic flow prediction model should satisfy the following conditions: (1) the intelligent passenger trip information inquiry system provides economic and time-saving route planning for passengers; (2) the intelligent traffic control system effectively reduces the probability of road congestion and accidents; (3) scientific evaluation and future optimization suggestions can be given to the performances of the two systems. Existing traffic flow prediction models can be roughly classified into three types, namely time series prediction, probability map models and non-parametric strategies. The first two methods have the following problems: (1) the model constructed in the background of lack of specific scenes and manual evaluation does not have specific scene pertinence, but is more generalized in general scenes; (2) the first two strategies cannot capture the non-linear scene in the traffic sequence compared with the non-parameter strategies. However, in terms of non-parametric strategies, represented by artificial neural networks, most artificial neural networks currently have only one single hidden layer in construction, because the application of multiple hidden layers is not necessarily capable of successfully constructing models in corresponding scenes, and in this regard, the linear construction strategy in the time series prediction method is better than the artificial neural network in performance. Therefore, in response to the shortcomings of the artificial neural network, researchers mainly address the shortcomings by mixing two aspects of the neural network and controlling the overfitting scenario in the construction of the neural network. The hybrid neural network integrates different statistical methods and calculation methods, wherein the most representative method is integration with a time sequence prediction strategy, a Takagi-Sugeno fuzzy depth model integrates fuzzy logic and a feedforward neural network to construct a short-term traffic flow model, Stathopoulos et al introduces Kalman filtering to optimize the model on the basis, Srinivasan et al uses fuzzy filtering parameters to perform clustering operation on data to serve as input parameters of the feedforward neural network, and Tan et al uses traditional sliding average and autoregressive sliding average models to serve as input parameters of the neural network. Although the hybrid neural network integrates different computing strategies, high-performance and high-storage computing equipment is required as support, and in addition, hardware conditions often limit the true computing capacity of the hybrid neural network strategy. For the strategy of suppressing the overfitting, the quality of the data sample can be improved by means of noise addition, but this method also needs a lot of calculation time and algorithm processing, and another common method is to use cross validation, namely, a data set is divided into a fitting data set and a validation data set, the former is used for neural network learning, the latter is used for estimating the validity error, and the training is terminated when the error rate is lower than a set threshold value. Liu et al, however, demonstrate that even reasonable cross-validation may not avoid the over-fitting phenomenon, and that validation data sets often do not have sufficient data representation to provide undirected data evaluation capabilities.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a deep neural network robust traffic prediction method based on multi-mode space-time data.
The purpose of the invention is realized by the following technical scheme:
a deep neural network robust traffic prediction method based on multi-modal spatiotemporal data comprises the following steps:
s100: obtaining the characteristic data of traffic flow state at a certain time according to the functionH (l) A feature data set is constructed which is,H (l) for mapping neural networkslThe output of layer and the input result of original data set can be expressed asx(t i ), i = 1,2, ... , n}；
S200: taking time series data corresponding to a certain time in S100 as input, and capturing an urban road network topological structure by using a graph convolution network to obtain spatial characteristics;
s300: inputting the time sequence with the space characteristics obtained by output into a gated recursive unit model, obtaining dynamic change through information transfer among units, and capturing time characteristics;
s400: T-GCN output prediction result obtained in full connection layer modey(t i ), i = 1,2, ... , n}; and the obtained output data set is processed by Kalman filtering to obtain a final output prediction result.
Further, the graph convolution network constructs a filter in the Fourier domain through the adjacency matrix and the feature matrix, the filter is applied to each node of the graph, the spatial features between the nodes are captured through the first-order neighborhood of the filter, and then a graph convolution network model is constructed by superposing a plurality of convolution layers:
wherein,
a self-join or adjacency matrix is introduced into the representation matrix,Awhich represents the adjacency matrix, is,I N representing an identity matrix;
the degree of representation matrix, i.e. ofjA hidden layer andia matrix under the network of visible layers is
，
Is as followsjA hidden layer andian adjacency matrix under the network of visible layers;H (l) is composed oflThe output of the layer(s) is,θ (l) the parameters related to the layer are included; the σ () function represents a Sigmoid function for constructing a nonlinear model.
Further, the step of obtaining the spatial features is as follows:
setting a certain node as a central road in an urban road, obtaining a topological relation between the central road and surrounding roads by a GCN model, coding a road network topological structure and road attributes, and capturing a spatial dependency relation by adopting a two-layer GCN model, wherein spatial characteristics are as follows:
wherein,xa matrix of the representative features is then generated,Arepresents a adjacency matrix;w 0 for the first layer of the GCN spatial weight matrix model,w 1 for the second layer GCN spatial weight matrix model,Relu（）i.e. a linear rectification function.
Further, in the process of acquiring the spatial characteristics, for the whole T-GCN model, the output of a GCN part at a certain moment enters the GRU model as data input, and finally an output result is generated through data processing of an update gate and a reset gate of the GRU part.
Further, in the process of acquiring the spatial characteristics, for a given road section, a characteristic matrix is definedxI.e. the length of the historical time series, the matrix being determined by the number of nodes of the matrix for the road sectionnAnd the node attribute characteristicspThe temperature of the molten steel is controlled by the temperature control device,
i.e. representing the traffic flow speed at the time t on the road section; for a contiguous matrixAIn other words, it reflects the traffic flow connection parameter between roads becauseRelu（）Normalizing the function to obtain a result ranging from 0 to 1, wherein 0 indicates no link between two roads, and 1 indicates a link between two roads, and the adjacency matrixAI.e. number of different road nodesnThereby forming the structure.
Further, in the Kalman filtering processing process, the initial predicted value of the traffic flow isy(t i ), i = 1,2, ... , nCan be converted intoy(t + T) And the predicted value istThe traffic flow at a time is related to the traffic flow before and after the time, and therefore:
wherein∂ 0 ，∂ 1 And∂ 2 respectively representtThe parameters of the matrix before and after the matrix,Vrepresenting the traffic flow at that moment in time,δ 0 for artificially introduced parametric noise, defined here as a covariance matrix, andtfront and rear traffic flowsV( )The following transformations are made:
according to kalman filtering, the above equation can be integrated as follows:
wherein,
s (t) is a state vector,ϕ(t) is
The corresponding state transition matrix is then used to determine,n(t) is an artificially designed noise processing function, and the covariance matrix is defined asδ 0 (t)；
Therefore, the traffic flow prediction formula subjected to the kalman filtering process is as follows, whereinβ(t) isϕ(t) corresponding observation matrix:
the invention has the beneficial effects that: compared with the traditional traffic flow prediction model, the prediction data value of the traffic flow prediction model constructed by the scheme has smaller deviation with the actually measured real data value and can better match with the actual traffic flow level, and the model is proved to have higher prediction precision in the aspect of traffic flow prediction.
Drawings
FIG. 1 is a flowchart of a TK-GCN model;
FIG. 2 is a comparison of traffic flow vehicle speed prediction data and actual traffic flow vehicle speed observation data for two models.
Detailed Description
The technical solution of the present invention is further described in detail with reference to the following specific examples, but the scope of the present invention is not limited to the following.
The traditional Convolutional Neural Network (CNN) cannot overcome the ability of graph strategies to construct urban roads, because urban road graphs are not two-dimensional spatial grids, and the traditional CNN cannot represent the complex topological structure of road networks and capture the dependency relationship of various parameters in the space. Therefore, aiming at the defects of the traditional CNN in expressing urban road traffic flow, the Lihaifeng team of the university of China and south in 2019 proposes a time graph convolution network (T-GCN) model which integrates a Graph Convolution Network (GCN) and a Gated Recursion Unit (GRU), wherein the graph convolution network captures the spatial dependency among the nodes of the traffic flow graph by learning a complex topological structure, and the gated recursion unit captures the temporal dependency by learning the dynamic change of traffic data.
In describing the spatial dependency relationship, the construction of the GCN model constructs a filter in the fourier domain through the adjacency matrix and the feature matrix, the filter is applied to each node of the graph, the spatial features between the nodes are captured through the first-order neighborhood of the filter, and then the GCN model is constructed by stacking a plurality of convolution layers:
wherein
The representation matrix is introduced with self-join, i.e. adjacency, matrices,I N representing an identity matrix;
the degree of representation matrix, i.e. ofjA hidden layer andia matrix under the network of visible layers is
，
Is as followsjA hidden layer andian adjacency matrix under the network of visible layers;H(l)is composed oflThe output of the layer(s) is,θ(l)the parameters related to the layer are included; the σ () function represents a Sigmoid function for constructing a nonlinear model. Therefore, a certain node is set as a central road in the urban road, and the central road can be obtained by the GCN modelAnd the topological relation between the road network and the surrounding roads encodes the topological structure of the road network and the road attributes to obtain the spatial dependence relation.
The spatial dependency is captured here using a two-layer GCN model, as follows:
wherein,xa matrix of the representative features is then generated,Arepresents a adjacency matrix;w 0 for the first layer of the GCN spatial weight matrix model,w 1 for the second layer GCN spatial weight matrix model,Relu（）i.e. a linear rectification function.
In the description of the time dependency relationship, because the long short-term memory network (LSTM) requires a relatively long training time for the data, the GRU model with a relatively simple structure and a relatively fast overall training speed is selected to obtain the time dependency from the traffic data. Thus, for the entire T-GCN model, the output of the GCN portion at a certain time goes into the GRU model as a data input, and the data processing through the update gate and the reset gate of the GRU portion finally produces an output result.
Due to the fact that the space-time evolution of the actual scene traffic flow has nonlinear and unstable characteristics, and good compatibility and complementation can not be achieved under the complex scene conditions and the multi-parameter reference among algorithms in a mixed model constructed by different deep confidence networks, based on the T-GCN model, Kalman filtering is introduced in the output process of a T-GCN output result to improve the original model, the TK-GCN model is obtained, and the deviation caused by the introduction of the multi-parameter is reduced in a simulation mode.
For a given road sectionRoadWe define its feature matrixxI.e. the length of the historical time series, the matrix being determined by the number of nodes of the matrix for the road sectionnAnd the node attribute characteristicspThe temperature of the molten steel is controlled by the temperature control device,
i.e. at time tA traffic flow speed down the road segment; for a contiguous matrixAIn other words, it reflects the traffic flow connection parameter between roads becauseRelu（）And (4) normalizing the function, wherein the calculation result ranges from 0 to 1, 0 indicates that no connection exists between two roads, and 1 indicates that the two roads are connected. Adjacency matrixAI.e. number of different road nodesnThereby forming the structure.
Therefore, the specific flow of the whole traffic flow prediction algorithm is as follows:
1. obtaining characteristic data of traffic flow state at a certain time, constructing characteristic data set according to function, and expressing the input result of original data set asx(t i ), i = 1,2, ... , n}；
2. Taking the historical time sequence data as input, capturing the topological structure of the urban road network by using a graph convolution network, and obtaining spatial characteristics;
3. inputting the time sequence with the space characteristics obtained by output into a gated recursive unit model, obtaining dynamic change through information transfer among units, and capturing time characteristics;
4. T-GCN output prediction result obtained in full connection layer modey(t i ), i = 1,2, ... , n}; and the obtained output data set is processed by Kalman filtering to obtain a final output prediction result, and the whole process is shown in FIG. 1.
In the filtering process of Kalman, the initial predicted value of traffic flowy(t i ), i = 1,2, ... , nCan be converted intoy(t + T) And the predicted value istThe traffic flow at a time is related to the traffic flow before and after the time, and therefore:
wherein∂ 0 ，∂ 1 And∂ 2 respectively representtThe parameters of the matrix before and after the matrix,Vrepresenting the traffic flow at that moment in time,δ 0 which is artificially introduced parametric noise, here defined as a covariance matrix. WhiletFront and rear traffic flowsV() The following transformations are made:
according to kalman filtering, the above equation can be integrated as follows:
wherein,
s (t) is a state vector,ϕ(t) is
The corresponding state transition matrix is then used to determine,n(t) is an artificially designed noise processing function, and the covariance matrix is defined asδ 0 (t)。
Therefore, the traffic flow prediction formula subjected to the kalman filtering process is as follows, whereinβ() Namely an observation matrix:
the experimental data and the simulation and verification work of the model of the research are based on that No. 5 month 1 to No. 10 month 30 adults in 2019 recorded by a Chuanhao system video intelligent analysis platform in real time encircle the eastern section of the urban expressway, the real-time speed is acquired once every 5 minutes, and the total number of sampling points in the whole data acquisition section is 397. Because the traffic flow data has certain regularity, and the double holidays and the working days respectively present different data characteristics, in order to fully utilize the regularity between the data, the scheme excludes the data of the double holidays and the national legal holidays from the training data for predicting the traffic flow of the highway and uses actual data for verification. The data is composed of an adjacency matrix and a feature matrix, and the adjacency matrix is calculated through the distance between camera points of the intelligent eye platform. In the experiment, the input data were normalized into interval [0,1] uniformly by the activation function. In addition, 80% of the data is used as a training set, and the remaining 20% is used as a test set, which is used as an input of the model. We predict traffic speeds of 10, 20, 30, 40, 50 and 60 minutes into the future.
The model of the invention is based on TK-GCN, and hyper-parameters involved in the model training process comprise: learning rate, training set capacity, and number of hidden layers. In the study, we set the learning rate to 0.001, and select 32 training sets, with the number of training iterations being 6000. In the deep learning framework, the sizes of different hidden layer values have great influence on the prediction accuracy, and the hidden layer values need a large amount of tests to be selected. Therefore, in order to select the optimal value of the hidden layer, attempts are made to select the optimal value by comparing the prediction results for different numbers of hidden layers. In the process of repeated verification, it is found that when the value of the number of hidden layers is greater than 60, the value of the mean square error no longer changes significantly, so the upper limit value of the number of hidden layers is set to 60, the number of nodes of the final output layer is {5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60}, and the value of the number of hidden layers is designed to be 40. The entire model is trained by an Adaptive Moment Estimation (Adam) optimizer.
In order to be able to better analyze and evaluate the prediction results, the invention uses the absolute mean error (MAE), the relative mean error (MRE) and the coefficient of determination (R)2) Three criteria are defined as follows:
absolute mean error (MAE)
Relative mean error (MRE)
Determining the coefficient (R)2)
WhereinY P A predicted value representing a traffic flow is displayed,Y R a true observation value representing the traffic flow,Y A is the average value within the statistical interval. The smaller the values of MAE and MRE are, the smaller the error of the predicted value is;R 2 and calculating a correlation coefficient, wherein the larger the numerical value is, the closer the prediction result is to the true level. The improved TK-GCN model provided by the invention is compared with a T-GCN model, a CRBM-DBN model, an LGCN-NGCN model and an FL-GCN-CNN model, and as can be known from Table 1, the error performance of the prediction model provided by the invention is obviously superior to that of mixed models of the T-GCN, the CRBM-DBN, the LGCN-NGCN and the FL-GCN-CNN based on a deep confidence network after the same data set is used for verification aiming at each algorithm, so that the effectiveness of the improved deep learning strategy in traffic flow prediction is shown.
TABLE 1 comparison of Performance indicators
And the FL-GCN-CNN model is higher in prediction accuracy compared with other models based on the deep confidence network, so that only the TK-GCN model and the FL-GCN-CNN model are compared with the actual traffic flow speed in terms of the prediction effect of the real traffic flow. Here, the prediction of the actual traffic flow is shown in fig. 2, taking the traffic flow speed tracking from No. 5/month 13 in 2019 to No. 5/month 17 in 2019 for 60 minutes as an example.
Referring to fig. 2, it can be seen that the true observed value, that is, the solid line part in fig. 2, has more accurate data, the data fluctuation of which is a nonlinear curve, and a plurality of peaks and troughs appear in the data, compared to the TK-GCN model and the FL-GCN-CNN model, the data of which has larger data error, especially larger data error in the peak and trough parts, compared to the true observed value, according to the FL-GCN-CNN model, the prediction by using the TK-GCN model, although there is still a certain error with the true observed value, obviously has smaller error, especially the peak and trough sections fit the true data better, and the overall curve trend fits the true observed value better, compared to the FL-GCN-CNN model.
Because a certain traffic flow level always exists in the expressway within 24 hours, the difference between the prediction of the traffic flow by the two models and the actual traffic flow speed change trend is not large, but the prediction performance of the TK-GCN algorithm on wave crests and wave troughs, namely the prediction of the traffic flow speed in the up-and-down peak period, and the prediction performance of the rising and falling trend of the traffic flow speed are realized, and the calculated prediction result is more consistent with the real traffic flow data change.
The method is applied to short-time traffic flow prediction based on an improved time chart convolution network model (TK-GCN). The method introduced by the scheme utilizes a Graph Convolution Network (GCN) to capture a road network topological structure and models the spatial dependence of the road network; meanwhile, a Gated Recursion Unit (GRU) is used for capturing dynamic changes of road traffic data, modeling is carried out on the time dependence of the dynamic changes, and then Kalman filtering processing is carried out on predicted data, so that essential characteristics of traffic flow data are mined; and finally, predicting the future traffic flow by using the model. The experimental result shows that compared with the traditional traffic flow prediction model, the prediction data value of the traffic flow prediction model constructed by the scheme has smaller deviation with the actual measured real data value and can better match with the actual traffic flow level, and the model is proved to have higher prediction precision in the aspect of traffic flow prediction.
The foregoing is illustrative of the preferred embodiments of this invention, and it is to be understood that this invention is not limited to the disclosed forms, but is intended to cover other embodiments, as may be used in various other combinations, modifications, and environments and is capable of changes within the scope of the invention as set forth, either as indicated by the above teachings or as may be learned by the practice of the invention. And that modifications and variations may be effected by those skilled in the art without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (5)
1. A deep neural network robust traffic prediction method based on multi-modal spatiotemporal data is characterized by comprising the following steps:
s100: obtaining the characteristic data of traffic flow state at a certain time according to the functionH (l) A feature data set is constructed which is,H (l) for mapping neural networkslThe output of layer and the input result of original data set can be expressed asx(t i ), i = 1,2, ... , n}；
S200: taking time series data corresponding to a certain time in S100 as input, and capturing an urban road network topological structure by using a graph convolution network to obtain spatial characteristics;
s300: inputting the time sequence with the space characteristics obtained by output into a gated recursive unit model, obtaining dynamic change through information transfer among units, and capturing time characteristics;
s400: T-GCN output prediction result obtained in full connection layer modey(t i ), i = 1,2, ... , n}; and the obtained output data set is processed by Kalman filtering to obtain a final output prediction result.
2. The method for deep neural network robust traffic prediction based on multi-modal spatiotemporal data as claimed in claim 1, wherein the graph convolution network constructs a filter in the fourier domain through the adjacency matrix and the feature matrix, the filter is applied to each node of the graph, the spatial features between the nodes are captured through the first-order neighborhood of the filter, and then a graph convolution network model is constructed by superposing a plurality of convolution layers:
wherein,
a self-join or adjacency matrix is introduced into the representation matrix,Awhich represents the adjacency matrix, is,I N representing an identity matrix;
the degree of representation matrix, i.e. ofjA hidden layer andia matrix under the network of visible layers is
，
Is as followsjA hidden layer andian adjacency matrix under the network of visible layers;H (l) is composed oflThe output of the layer(s) is,θ (l) the parameters related to the layer are included; the σ () function represents a Sigmoid function for constructing a nonlinear model.
3. The deep neural network robust traffic prediction method based on multi-modal spatiotemporal data according to claim 2, characterized in that the spatial features are obtained by the steps of:
setting a certain node as a central road in an urban road, obtaining a topological relation between the central road and surrounding roads by a GCN model, coding a road network topological structure and road attributes, and capturing a spatial dependency relation by adopting a two-layer GCN model, wherein spatial characteristics are as follows:
wherein,xa matrix of the representative features is then generated,Arepresents a adjacency matrix;w 0 for the first layer GCN spatial weightThe matrix model is a model of the matrix,w 1 for the second layer GCN spatial weight matrix model,Relu（）i.e. a linear rectification function.
4. The method for deep neural network robust traffic prediction based on multi-modal spatio-temporal data as claimed in claim 3, wherein in the process of obtaining the spatial features, for the whole T-GCN model, the output of a GCN part at a certain time enters the GRU model as data input, and the output result is finally generated through data processing of an update gate and a reset gate of the GRU part.
5. The method of claim 4, wherein in the process of obtaining the spatial features, for a given road segment, we define its feature matrixxI.e. the length of the historical time series, the matrix being determined by the number of nodes of the matrix for the road sectionnAnd the node attribute characteristicspIs formed for the adjacent matrixAIn other words, it reflects the traffic flow connection parameter between roads becauseRelu（）Normalizing the function to obtain a result ranging from 0 to 1, wherein 0 indicates no link between two roads, and 1 indicates a link between two roads, and the adjacency matrixAI.e. number of different road nodesnThereby forming the structure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011587058.2A CN112289034A (en) | 2020-12-29 | 2020-12-29 | Deep neural network robust traffic prediction method based on multi-mode space-time data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011587058.2A CN112289034A (en) | 2020-12-29 | 2020-12-29 | Deep neural network robust traffic prediction method based on multi-mode space-time data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112289034A true CN112289034A (en) | 2021-01-29 |
Family
ID=74426538
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011587058.2A Pending CN112289034A (en) | 2020-12-29 | 2020-12-29 | Deep neural network robust traffic prediction method based on multi-mode space-time data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112289034A (en) |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112801386A (en) * | 2021-02-04 | 2021-05-14 | 北京工业大学 | Short-term traffic flow prediction method based on three-stage model |
| CN113053115A (en) * | 2021-03-17 | 2021-06-29 | 北京工商大学 | Traffic prediction method based on multi-scale graph convolution network model |
| CN113065690A (en) * | 2021-03-19 | 2021-07-02 | 浙江财经大学 | Traffic prediction method and device |
| CN113268916A (en) * | 2021-04-07 | 2021-08-17 | 浙江工业大学 | Traffic accident prediction method based on space-time graph convolutional network |
| CN113487856A (en) * | 2021-06-04 | 2021-10-08 | 兰州理工大学 | Traffic flow combination prediction model based on graph convolution network and attention mechanism |
| CN113610286A (en) * | 2021-07-27 | 2021-11-05 | 中国地质大学（武汉） | PM accounting for spatio-temporal correlations and meteorological factors2.5Concentration prediction method and device |
| CN113823090A (en) * | 2021-09-22 | 2021-12-21 | 南通大学 | Traffic flow prediction method based on multi-space-time graph convolution network |
| CN114205251A (en) * | 2021-12-09 | 2022-03-18 | 西安电子科技大学 | Switch link resource prediction method based on space-time characteristics |
| CN114333327A (en) * | 2022-01-05 | 2022-04-12 | 北京航空航天大学合肥创新研究院(北京航空航天大学合肥研究生院) | Traffic state prediction method based on multi-stream neural network |
| CN114511154A (en) * | 2022-02-22 | 2022-05-17 | 南通大学 | Traffic flow prediction method based on space-time complex graph convolution network |
| CN114860715A (en) * | 2022-03-18 | 2022-08-05 | 浙江大学 | Lanczos space-time network method for predicting flow in real time |
| CN115116225A (en) * | 2022-06-23 | 2022-09-27 | 上海交通大学 | Hybrid traffic flow-oriented data-driven random model prediction control method |
| CN115116212A (en) * | 2022-05-06 | 2022-09-27 | 浙江科技学院 | Traffic prediction method for road network, computer device, storage medium and program product |
| CN115482666A (en) * | 2022-09-13 | 2022-12-16 | 杭州电子科技大学 | Multi-graph convolution neural network traffic prediction method based on data fusion |
| CN115809747A (en) * | 2023-02-06 | 2023-03-17 | 东南大学 | Pyramid cause-and-effect network-based coupling information flow long-term prediction method |
| CN115938103A (en) * | 2022-11-01 | 2023-04-07 | 北京航空航天大学 | Space-time streaming data prediction method based on incremental learning |
| CN116153069A (en) * | 2023-02-09 | 2023-05-23 | 东南大学 | Traffic flow model and data fusion driven traffic state estimation method and device |
| CN116824838A (en) * | 2022-11-07 | 2023-09-29 | 苏州规划设计研究院股份有限公司 | Road network level traffic flow integrated prediction method based on self-adaptive time sequence decomposition |
| CN117275215A (en) * | 2023-05-04 | 2023-12-22 | 长江空间信息技术工程有限公司(武汉) | Urban road congestion space-time prediction method based on graph process neural network |
| CN117292584A (en) * | 2023-11-24 | 2023-12-26 | 南京航空航天大学 | Multi-machine-field terminal area flow prediction method and system based on GCN-LSTM model |
| CN117351718A (en) * | 2023-10-19 | 2024-01-05 | 东南大学 | Urban road network mixed traffic flow daily evolution prediction method based on intelligent networking environment |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105321345A (en) * | 2015-09-18 | 2016-02-10 | 浙江工业大学 | Road traffic flow prediction method based on ARIMA model and kalman filtering |
| CN109754605A (en) * | 2019-02-27 | 2019-05-14 | 中南大学 | A kind of traffic forecast method based on attention temporal diagram convolutional network |
| CN110969854A (en) * | 2019-12-13 | 2020-04-07 | 深圳先进技术研究院 | Traffic flow prediction method, system and terminal equipment |
| CN111079975A (en) * | 2019-11-14 | 2020-04-28 | 青岛海信网络科技股份有限公司 | Traffic data prediction method and device and vehicle control method |
| CN111210089A (en) * | 2020-01-17 | 2020-05-29 | 大连大学 | Stock price prediction method of gated cyclic unit neural network based on Kalman filtering |
| CN111696355A (en) * | 2020-06-29 | 2020-09-22 | 中南大学 | Dynamic graph convolution traffic speed prediction method |
-
2020
- 2020-12-29 CN CN202011587058.2A patent/CN112289034A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105321345A (en) * | 2015-09-18 | 2016-02-10 | 浙江工业大学 | Road traffic flow prediction method based on ARIMA model and kalman filtering |
| CN109754605A (en) * | 2019-02-27 | 2019-05-14 | 中南大学 | A kind of traffic forecast method based on attention temporal diagram convolutional network |
| CN111079975A (en) * | 2019-11-14 | 2020-04-28 | 青岛海信网络科技股份有限公司 | Traffic data prediction method and device and vehicle control method |
| CN110969854A (en) * | 2019-12-13 | 2020-04-07 | 深圳先进技术研究院 | Traffic flow prediction method, system and terminal equipment |
| CN111210089A (en) * | 2020-01-17 | 2020-05-29 | 大连大学 | Stock price prediction method of gated cyclic unit neural network based on Kalman filtering |
| CN111696355A (en) * | 2020-06-29 | 2020-09-22 | 中南大学 | Dynamic graph convolution traffic speed prediction method |
Non-Patent Citations (4)
| Title |
|---|
| LING ZHAO ET AL.: "T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction", 《IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS》 * |
| THOMAS N. KIPF ET AL.: "Semi-supervised Classification with Graph Convolutional Networks", 《ICLR 2017》 * |
| 朱凯利等: "基于图卷积神经网络的交通流量预测", 《智能计算机与应用》 * |
| 杨兆升: "《城市交通流诱导系统理论与模型》", 31 January 2000, 人民交通出版社 * |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112801386A (en) * | 2021-02-04 | 2021-05-14 | 北京工业大学 | Short-term traffic flow prediction method based on three-stage model |
| CN113053115A (en) * | 2021-03-17 | 2021-06-29 | 北京工商大学 | Traffic prediction method based on multi-scale graph convolution network model |
| CN113065690A (en) * | 2021-03-19 | 2021-07-02 | 浙江财经大学 | Traffic prediction method and device |
| CN113065690B (en) * | 2021-03-19 | 2023-06-30 | 浙江财经大学 | Traffic prediction method and device |
| CN113268916A (en) * | 2021-04-07 | 2021-08-17 | 浙江工业大学 | Traffic accident prediction method based on space-time graph convolutional network |
| CN113487856A (en) * | 2021-06-04 | 2021-10-08 | 兰州理工大学 | Traffic flow combination prediction model based on graph convolution network and attention mechanism |
| CN113610286A (en) * | 2021-07-27 | 2021-11-05 | 中国地质大学（武汉） | PM accounting for spatio-temporal correlations and meteorological factors2.5Concentration prediction method and device |
| CN113610286B (en) * | 2021-07-27 | 2024-03-29 | 中国地质大学（武汉） | PM taking into account space-time correlation and meteorological factors 2.5 Concentration prediction method and device |
| CN113823090A (en) * | 2021-09-22 | 2021-12-21 | 南通大学 | Traffic flow prediction method based on multi-space-time graph convolution network |
| CN114205251B (en) * | 2021-12-09 | 2022-12-02 | 西安电子科技大学 | Switch link resource prediction method based on space-time characteristics |
| CN114205251A (en) * | 2021-12-09 | 2022-03-18 | 西安电子科技大学 | Switch link resource prediction method based on space-time characteristics |
| CN114333327A (en) * | 2022-01-05 | 2022-04-12 | 北京航空航天大学合肥创新研究院(北京航空航天大学合肥研究生院) | Traffic state prediction method based on multi-stream neural network |
| CN114511154A (en) * | 2022-02-22 | 2022-05-17 | 南通大学 | Traffic flow prediction method based on space-time complex graph convolution network |
| CN114860715A (en) * | 2022-03-18 | 2022-08-05 | 浙江大学 | Lanczos space-time network method for predicting flow in real time |
| CN115116212A (en) * | 2022-05-06 | 2022-09-27 | 浙江科技学院 | Traffic prediction method for road network, computer device, storage medium and program product |
| CN115116225A (en) * | 2022-06-23 | 2022-09-27 | 上海交通大学 | Hybrid traffic flow-oriented data-driven random model prediction control method |
| CN115116225B (en) * | 2022-06-23 | 2023-08-04 | 上海交通大学 | Data-driven random model predictive control method for mixed traffic flow |
| CN115482666A (en) * | 2022-09-13 | 2022-12-16 | 杭州电子科技大学 | Multi-graph convolution neural network traffic prediction method based on data fusion |
| CN115482666B (en) * | 2022-09-13 | 2024-05-07 | 杭州电子科技大学 | Multi-graph convolution neural network traffic prediction method based on data fusion |
| CN115938103A (en) * | 2022-11-01 | 2023-04-07 | 北京航空航天大学 | Space-time streaming data prediction method based on incremental learning |
| CN116824838A (en) * | 2022-11-07 | 2023-09-29 | 苏州规划设计研究院股份有限公司 | Road network level traffic flow integrated prediction method based on self-adaptive time sequence decomposition |
| CN115809747B (en) * | 2023-02-06 | 2023-05-09 | 东南大学 | Pyramid causal network-based coupling information flow long-term prediction method |
| CN115809747A (en) * | 2023-02-06 | 2023-03-17 | 东南大学 | Pyramid cause-and-effect network-based coupling information flow long-term prediction method |
| CN116153069A (en) * | 2023-02-09 | 2023-05-23 | 东南大学 | Traffic flow model and data fusion driven traffic state estimation method and device |
| CN116153069B (en) * | 2023-02-09 | 2024-01-30 | 东南大学 | Traffic flow model and data fusion driven traffic state estimation method and device |
| CN117275215A (en) * | 2023-05-04 | 2023-12-22 | 长江空间信息技术工程有限公司(武汉) | Urban road congestion space-time prediction method based on graph process neural network |
| CN117351718A (en) * | 2023-10-19 | 2024-01-05 | 东南大学 | Urban road network mixed traffic flow daily evolution prediction method based on intelligent networking environment |
| CN117351718B (en) * | 2023-10-19 | 2024-08-06 | 东南大学 | Urban road network mixed traffic flow daily evolution prediction method based on intelligent networking environment |
| CN117292584A (en) * | 2023-11-24 | 2023-12-26 | 南京航空航天大学 | Multi-machine-field terminal area flow prediction method and system based on GCN-LSTM model |
| CN117292584B (en) * | 2023-11-24 | 2024-02-02 | 南京航空航天大学 | Multi-machine-field terminal area flow prediction method and system based on GCN-LSTM model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112289034A (en) | Deep neural network robust traffic prediction method based on multi-mode space-time data | |
| CN107862864B (en) | Driving condition intelligent prediction estimation method based on driving habits and traffic road conditions | |
| CN110164128B (en) | City-level intelligent traffic simulation system | |
| CN109117987B (en) | Personalized traffic accident risk prediction recommendation method based on deep learning | |
| CN108648457B (en) | Method, device and computer readable storage medium for speed prediction | |
| CN106951976B (en) | Bus passenger flow prediction method based on mode classification | |
| CN110648014B (en) | Regional wind power prediction method and system based on space-time quantile regression | |
| CN111563706A (en) | Multivariable logistics freight volume prediction method based on LSTM network | |
| CN111612243A (en) | Traffic speed prediction method, system and storage medium | |
| CN113672865A (en) | Traffic flow prediction method based on depth map Gaussian process | |
| CN112071062B (en) | Driving time estimation method based on graph convolution network and graph attention network | |
| CN113591380B (en) | Traffic flow prediction method, medium and equipment based on graph Gaussian process | |
| CN112542049A (en) | Comprehensive management and control platform for intelligent traffic | |
| CN107967803A (en) | Traffic congestion Forecasting Methodology based on multi-source data and variable-weight combined forecasting model | |
| CN110689183B (en) | Cluster photovoltaic power probability prediction method, system, medium and electronic device | |
| CN111047078B (en) | Traffic characteristic prediction method, system and storage medium | |
| CN113420414A (en) | Short-term traffic flow prediction model based on dynamic space-time analysis | |
| CN112699941B (en) | Plant disease severity image classification method, device, equipment and storage medium | |
| CN112966871A (en) | Traffic jam prediction method and system based on convolution long-short term memory neural network | |
| CN111967308A (en) | Online road surface unevenness identification method and system | |
| CN113435658A (en) | Traffic flow prediction method based on space-time fusion correlation and attention mechanism | |
| CN116311921A (en) | Traffic speed prediction method based on multi-spatial scale space-time converter | |
| CN107481523A (en) | A kind of traffic flow speed Forecasting Methodology and system | |
| CN115762147B (en) | Traffic flow prediction method based on self-adaptive graph meaning neural network | |
| CN109583741B (en) | Driving track risk assessment method and system |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210129 |