CN110516833A - A method of the Bi-LSTM based on feature extraction predicts road traffic state - Google Patents
A method of the Bi-LSTM based on feature extraction predicts road traffic state Download PDFInfo
- Publication number
- CN110516833A CN110516833A CN201910593469.3A CN201910593469A CN110516833A CN 110516833 A CN110516833 A CN 110516833A CN 201910593469 A CN201910593469 A CN 201910593469A CN 110516833 A CN110516833 A CN 110516833A
- Authority
- CN
- China
- Prior art keywords
- lstm
- data
- model
- road
- road traffic
- 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
Classifications
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Business, Economics & Management (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Traffic Control Systems (AREA)
Abstract
A method of the Bi-LSTM based on feature extraction predicts road traffic state, stacking self-encoding encoder (SAE) made of multiple self-encoding encoders stacking extracts the feature of road condition data first, then using the road traffic state space-time characteristic extracted as the input of two-way shot and long term time memory network (Bi-LSTM) model, the building of the Bi-LSTM network model based on feature extraction is completed.The present invention can effectively improve the precision of prediction of road traffic state to the input after the progress feature extraction of road traffic state matrix as Bi-LSTM model.
Description
Technical field
The present invention relates to a kind of, and two-way shot and long term memory network model (Bi-LSTM) the prediction road based on feature extraction is handed over
The method of logical state, the invention belongs to traffic forecast fields.
Background technique
Along with the rapid development of urban economy, city size constantly expands, and people's quality of life is growing, city road
The number of vehicle is constantly increased sharply in road, and thus bring traffic jam issue gradually attracts people's attention, in order to more efficient
Path resource is distributed, Urban Traffic Jam Based is alleviated, traffic flow pattern prediction has vital effect.
Road traffic prediction technique at this stage mainly has: conventional differential integrates rolling average autoregression method, support to
Amount machine, Feedback Neural Network method etc.；The prediction of these conventional methods fails sufficiently to excavate the time in Traffic Net
With space characteristics, therefore precision of prediction is not high.
Summary of the invention
In order to overcome the lower deficiency of precision of prediction of existing road traffic state prediction mode, the present invention proposes one kind
The method of Bi-LSTM prediction road traffic state based on feature extraction, this method are right first with self-encoding encoder (SAE) is stacked
Road traffic state matrix carries out feature extraction, using the feature extracted as the input of two-way LSTM network model, further
The high dimensional feature for extracting road traffic state matrix, through regression forecasting layer output model prediction result, to realize road not
Carry out the prediction of vehicle flowrate in short-term and speed.
The technical solution adopted by the present invention to solve the technical problems is:
A method of the Bi-LSTM based on feature extraction predicts road traffic state, comprising the following steps:
1) traffic state data for extracting a plurality of road carries out data set division and carries out data prediction respectively: determining
The a plurality of section of prediction and time range, the road condition data in all sections within the scope of extraction time, including vehicle flowrate and flat
The status data in all sections is carried out data merging according to same time sequence index by equal vehicle speed data, and construct is empty now
The road traffic state matrix of information；Data set is trained to road traffic state data matrix and test data set divides,
And operation is standardized to training dataset and test data set respectively；
2) it constructs the Bi-LSTM network model based on feature extraction: multiple self-encoding encoders is stacked into building SAE network, benefit
Space-time characteristic extraction is carried out to input data with SAE network, and using the space-time characteristic data after extraction as Bi-LSTM network mould
The Bi-LSTM network model building based on feature extraction is completed in the input of type；
3) training of Bi-LSTM network model and evaluation based on feature extraction: training dataset is inputted and is mentioned based on feature
The Bi-LSTM network model taken, Definition Model loss function, and the call back function for only retaining optimal models is set, to minimize
For the purpose of model loss function, using back-propagation algorithm loop iteration, the optimal Bi- based on feature extraction is finally saved
LSTM network model, and test data set is inputted into the model and carries out model evaluation；
4) it realizes that road traffic state prediction is rapid based on real-time road traffic status information: obtaining real-time road traffic state
Matrix is standardized the input operated and as the Bi-LSTM network model based on feature extraction to data are obtained, by model
Output data carry out anti-normalizing operation, realize real-time road traffic status predication.
Further, the process of the step 1) is as follows:
Step 1.1: determining a plurality of section and the time range of prediction, the road condition in all sections within the scope of extraction time
The status data in all sections is carried out data merging according to same time sequence index by data, and construct shows space time information
Road traffic state matrix；
The vehicle flowrate data and speed data for extracting the n moment within the scope of m section same time of prediction, wherein i-th
The vehicle flowrate data in section are denoted as respectively with speed data: si=[si1,si2,si3…sin]T, vi=[vi1,vi2,vi3…vin]T, i
The vehicle flowrate data in all sections and speed data are carried out data merging according to same time sequence index by=1,2,3 ... m,
Road condition matrix X=[s after merging1,v1,s2,v2…sm,vm], wherein the road traffic state X of t momentt=[s1t,
v1t,s2t,v2t…smt,vmt], t=1,2,3 ... n；
Step 1.2: data set being trained to road traffic state data matrix and test data set divides, and is right respectively
Training dataset and test data set are standardized operation
By road traffic state matrix X according to a:(1-a) ratio cut partition be training set XtrainWith test set Xtest, wherein a
∈ (0,1), Xtrain=[X1,X2,X3…Xna]T, Xtest=[Xna+1,Xna+2,Xna+3…Xn]T, to after division training set and test
Collection carries out deviation normalizing operation respectively, scales the vehicle flowrate in each section proportionally with speed data, is uniformly mapped to
In [0,1] section, deviation standardized calculation formula are as follows:
Wherein,The minimum value of i-th section vehicle flowrate and speed is respectively represented, Respectively represent
The maximum value of i section vehicle flowrate and speed,The wagon flow of i-th article of section t moment after being expressed as deviation standardization
Amount and speed data.It is respectively after training set and the standardization of test set deviation
Further, the process of the step 2) is as follows:
Step 2.1: extracting the space-time characteristic of road traffic state matrix
The SAE that the sparse self-encoding encoder of N layers of building stacks, N >=2, by road traffic state matrixSAE model is inputted,
Extract the space-time characteristic of road traffic state matrix, the calculation expression of first layer self-encoding encoder are as follows:
WhereinThe parameter learnt, X are needed for modelt'rainIt is approximately equal to model
Inputσ is Sigmoid function, function expression are as follows:
Through first layer self-encoding encoder to road traffic state matrixSingle order hidden layer is obtained after carrying out space-time characteristic extraction
Characteristic feature h1, by h1As the input of next self-encoding encoder, it is special that second order hidden layer characterization is obtained after encoded device f encoding operation
Levy h2, exported by model is obtained after decoder r decodingAndIt is approximately equal to h1.Similarly, this layer of sparse self-encoding encoder
Input of the hidden layer characteristic feature practised as next coefficient self-encoding encoder carries out feature learning.By N layers of self-encoding encoder
After coding and decoding operation layer by layer, traffic behavior matrix N rank space-time characteristic feature h is extractedN；
Step 2.2: the Bi-LSTM network model building based on feature extraction
By traffic behavior matrix space-time characteristic feature hNAs the input of Bi-LSTM model, Bi-LSTM network model is used
The double-deck LSTM building, it is respectively preceding to LSTM unit and backward LSTM unit, the calculating process of LSTM unit are as follows:
ft=σ (Wfht-1+WfhN+b1) (6)
it=σ (Wiht-1+WihN+b2) (7)
ot=σ (Woht-1+WohN+b4) (10)
ht=ot⊙tanh(Ct) (11)
Wherein, ⊙ represents dot product, Wf、Wi、Wc、WoIt is expressed as forgeing door, input gate, the power of state cell, out gate
Weight matrix, b1, b2, b3, b4 are expressed as forgeing the biasing size of door, input gate, state cell, out gate, and σ is
Sigmoid activation primitive, tanh activation primitive expression formula are as follows:
T moment, traffic behavior matrix space-time characteristic feature hNIt is through the preceding feature extracted to LSTM unitThrough backward
LSTM unit extract feature beBi-LSTM is exported after carrying out high dimensional feature extraction:
Wherein W*, b*Respectively weight matrix and biasing size, σ are Sigmoid activation primitive.
Input by the output of Bi-LSTM model as regression forecasting layer, completes the Bi-LSTM network based on feature extraction
The building of model, regression forecasting layer functions expression formula are as follows:
Yout=σ (WoutYl+bout) (14)
Wherein Yl=[y1,y2,y3…yl]T, Wout, boutIt is for the weight matrix and biasing size, σ of regression forecasting layer
Sigmoid activation primitive, l are the output unit number of Bi-LSTM network.
Further, the process of the step 3) is as follows:
In t moment, by training datasetBi-LSTM network model is inputted, through feature extraction and returns layer prediction,
I-th road is obtained in the traffic behavior at the following τ moment:
Actual traffic state of i-th road at the following τ moment:
Definition Model loss function:
For the purpose of minimizing model loss function, using back-propagation algorithm loop iteration, obtain optimal based on spy
The Bi-LSTM network model extracted is levied, by test data setInput the optimal Bi-LSTM network mould based on feature extraction
Type is compared with true value according to the output of model, evaluates model.
Further, the process of the step 4) is as follows:
Road real-time traffic states matrix is obtained, is standardized operation and as based on feature extraction to data are obtained
The input of Bi-LSTM network model obtains the traffic status prediction value of i-th road:
Model prediction result is subjected to anti-normalizing operation, obtains predicting true vehicle flowrate and speed state, it is anti-to standardize
Calculation formula are as follows:
Wherein,The minimum value of i-th section vehicle flowrate and speed is respectively represented, Respectively represent
The maximum value of i section vehicle flowrate and speed, vi(t+j), si(t+j)Respectively i-th article of road (t+j) moment vehicle flowrate with
Prediction of speed size realizes the prediction of vehicle flowrate and speed in short time traffic conditions.
Technical concept of the invention are as follows: this method stacks self-encoding encoder made of multiple self-encoding encoders stacking first
(SAE) feature of road condition data is extracted, then using the road traffic state space-time characteristic extracted as two-way
The structure of the Bi-LSTM network model based on feature extraction is completed in the input of shot and long term time memory network (Bi-LSTM) model
It builds.The experimental results showed that the input after the progress feature extraction of road traffic state matrix as Bi-LSTM model, Ke Yiyou
Effect improves the precision of prediction of road traffic state.
Beneficial effects of the present invention: the present invention is by carrying out space-time characteristic extraction to a plurality of road traffic state matrix and inciting somebody to action
Input of the space-time characteristic as Bi-LSTM network model, realize in road traffic state in short-term the vehicle flowrate in a plurality of section and
The prediction of speed has carried out the excavation of depth to the potential feature of road traffic state matrix, has improved traffic status prediction
Accuracy.
Traffic forecast plays the role of key in terms of intelligent transportation induction, and the present invention realizes the traffic in short-term of a plurality of road
The prediction of vehicle flowrate and speed in state.
Detailed description of the invention
Fig. 1 is two-way LSTM network architecture figure.
Fig. 2 is the Bi-LSTM network model prediction result (average speed) based on feature extraction.
Fig. 3 is the Bi-LSTM network model prediction result (vehicle flowrate) based on feature extraction.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.
Referring to Fig.1~Fig. 3, a method of the Bi-LSTM based on feature extraction predicts road traffic state, including following
Step:
1) the traffic state data matrix for extracting a plurality of road carries out data set division and carries out data prediction respectively,
Process is as follows:
Step 1.1: determining a plurality of section and the time range of prediction, the road condition in all sections within the scope of extraction time
The status data in all sections is carried out data merging according to same time sequence index by data, and construct shows space time information
Road traffic state matrix；
The vehicle flowrate data and speed data for extracting the n moment within the scope of m section same time of prediction, wherein i-th
The vehicle flowrate data in section are denoted as respectively with speed data: si=[si1,si2,si3…sin]T, vi=[vi1,vi2,vi3…vin]T, i
The vehicle flowrate data in all sections and speed data are carried out data merging according to same time sequence index by=1,2,3 ... m,
Road condition matrix X=[s after merging1,v1,s2,v2…sm,vm], wherein the road traffic state X of t momentt=[s1t,
v1t,s2t,v2t…smt,vmt], t=1,2,3 ... n；
Step 1.2: data set being trained to road traffic state data matrix and test data set divides, and is right respectively
Training dataset and test data set are standardized operation
By road traffic state matrix X according to a:(1-a) ratio cut partition be training set XtrainWith test set Xtest, wherein a
∈ (0,1), Xtrain=[X1,X2,X3…Xna]T, Xtest=[Xna+1,Xna+2,Xna+3…Xn]T, to after division training set and test
Collection carries out deviation normalizing operation respectively, scales the vehicle flowrate in each section proportionally with speed data, is uniformly mapped to
In [0,1] section, deviation standardized calculation formula are as follows:
Wherein,The minimum value of i-th section vehicle flowrate and speed is respectively represented, Respectively represent
The maximum value of i section vehicle flowrate and speed,The wagon flow of i-th article of section t moment after being expressed as deviation standardization
Amount and speed data.It is respectively after training set and the standardization of test set deviation
2) the Bi-LSTM network model based on feature extraction is constructed, process is as follows:
Step 2.1: extracting the space-time characteristic of road traffic state matrix
The SAE that the sparse self-encoding encoder of N layers of building stacks, N >=2, by road traffic state matrixSAE model is inputted,
Extract the space-time characteristic of road traffic state matrix, the calculation expression of first layer self-encoding encoder are as follows:
WhereinThe parameter learnt, X are needed for modelt'rainIt is approximately equal to model
Inputσ is Sigmoid function, function expression are as follows:
Through first layer self-encoding encoder to road traffic state matrixSingle order hidden layer is obtained after carrying out space-time characteristic extraction
Characteristic feature h1, by h1As the input of next self-encoding encoder, it is special that second order hidden layer characterization is obtained after encoded device f encoding operation
Levy h2, exported by model is obtained after decoder r decodingAndIt is approximately equal to h1.Similarly, this layer of sparse self-encoding encoder
Input of the hidden layer characteristic feature practised as next coefficient self-encoding encoder carries out feature learning.By N layers of self-encoding encoder
After coding and decoding operation layer by layer, traffic behavior matrix N rank space-time characteristic feature h is extractedN；
Step 2.2: the Bi-LSTM network model building based on feature extraction
By traffic behavior matrix space-time characteristic feature hNAs the input of Bi-LSTM model, Bi-LSTM network model is used
The double-deck LSTM building, it is respectively preceding to LSTM unit and backward LSTM unit, the calculating process of LSTM unit are as follows:
ft=σ (Wfht-1+WfhN+b1) (6)
it=σ (Wiht-1+WihN+b2) (7)
ot=σ (Woht-1+WohN+b4) (10)
ht=ot⊙tanh(Ct) (11)
Wherein, ⊙ represents dot product, Wf、Wi、Wc、WoIt is expressed as forgeing door, input gate, the power of state cell, out gate
Weight matrix, b1, b2, b3, b4 are expressed as forgeing the biasing size of door, input gate, state cell, out gate, and σ is
Sigmoid activation primitive, tanh activation primitive expression formula are as follows:
T moment, traffic behavior matrix space-time characteristic feature hNIt is through the preceding feature extracted to LSTM unitThrough backward
LSTM unit extract feature beBi-LSTM is exported after carrying out high dimensional feature extraction:
Wherein W*, b*Respectively weight matrix and biasing size, σ are Sigmoid activation primitive.
Input by the output of Bi-LSTM model as regression forecasting layer, completes the Bi-LSTM network based on feature extraction
The building of model, regression forecasting layer functions expression formula are as follows:
Yout=σ (WoutYl+bout) (14)
Wherein Yl=[y1,y2,y3…yl]T, Wout, boutIt is for the weight matrix and biasing size, σ of regression forecasting layer
Sigmoid activation primitive, l are the output unit number of Bi-LSTM network.
3) training of Bi-LSTM network model and evaluation based on feature extraction；
In t moment, by training datasetBi-LSTM network model is inputted, through feature extraction and returns layer prediction,
I-th road is obtained in the traffic behavior at the following τ moment:
Actual traffic state of i-th road at the following τ moment:
Definition Model loss function:
For the purpose of minimizing model loss function, using back-propagation algorithm loop iteration, obtain optimal based on spy
The Bi-LSTM network model extracted is levied, by test data setInput the optimal Bi-LSTM network mould based on feature extraction
Type is compared with true value according to the output of model, evaluates model；
4) road traffic state prediction is realized based on real-time road traffic status information；
Road real-time traffic states matrix is obtained, is standardized operation and as based on feature extraction to data are obtained
The input of Bi-LSTM network model obtains the traffic status prediction value of i-th road:
Model prediction result is subjected to anti-normalizing operation, obtains predicting true vehicle flowrate and speed state, it is anti-to standardize
Calculation formula are as follows:
Wherein,The minimum value of i-th section vehicle flowrate and speed is respectively represented, Respectively represent
The maximum value of i section vehicle flowrate and speed, vi(t+j), si(t+j)Respectively i-th article of road (t+j) moment vehicle flowrate with
Prediction of speed size realizes the prediction of vehicle flowrate and speed in short time traffic conditions.
In the present embodiment, data in actual experiment, process is as follows:
1) experimental data is chosen
Experimental data amounted to 29 days traffic behavior numbers using 10 sections June 29 1 day to 2011 June in 2011
According to traffic state data includes section vehicle flowrate and average speed data；Data sampling period is 2min, every section vehicle flowrate
Data amount check with average speed is 20880.
It tests using preceding 28 days vehicle flowrates in 10 sections and speed data and is used as training set, last day data are as testing
Collection, test set are used to evaluate model index；When model training, with 2 hours vehicle flowrates of a plurality of section history and number of speed
It is predicted that the vehicle flowrate and speed data in following 30 minutes every sections.
2) parameter determines
Experimental section of the invention is based on realizing under Tensorflow environment, using Keras neural network framework completion net
The building of network model；Design parameter is arranged in model are as follows: constitutes storehouse self-encoding encoder, each self-encoding encoder using 2 layers of self-encoding encoder
Hidden unit number be 128, activation primitive is selected as Sigmoid function；The forward and backward of two-way LSTM (Bi-LSTM)
LSTM unit number is 64, activation primitive tanh；Regression forecasting layer unit number is 300；Model training the number of iterations is
100, minimum batch size is 512 when each repetitive exercise, is all made of Adam optimizer Optimized model parameter.
3) experimental result
The present invention realizes the prediction of the vehicle flowrate of Multiple Sections and speed in short-term, the performance of selection test the set pair analysis model
It is assessed.
In this experiment, evaluated using mean absolute error (MAE), root-mean-square error (RMSE) as model prediction accuracy
Index, calculation formula are respectively as follows:
Wherein ypreFor the vehicle flowrate and speed data matrix of model prediction, ytrueFor true vehicle flowrate and speed data
Matrix chooses the number of data when k is evaluation model performance.
Table 1 is multiple tracks road forecasting traffic flow result:
Table 1.
Claims (5)
1. a kind of method of the Bi-LSTM prediction road traffic state based on feature extraction, which is characterized in that the method includes
Following steps:
1) traffic state data for extracting a plurality of road carries out data set division and carries out data prediction respectively: determining prediction
A plurality of section and time range, the road condition data in all sections within the scope of extraction time, including vehicle flowrate and average vehicle
The status data in all sections is carried out data merging according to same time sequence index by fast data, and construct shows space time information
Road traffic state matrix；Data set is trained to road traffic state data matrix and test data set divides, and point
It is other that operation is standardized to training dataset and test data set；
2) it constructs the Bi-LSTM network model based on feature extraction: multiple self-encoding encoders being stacked into building SAE network, utilize SAE
Network carries out space-time characteristic extraction to input data, and using the space-time characteristic data after extraction as Bi-LSTM network model
The Bi-LSTM network model building based on feature extraction is completed in input；
3) training of Bi-LSTM network model and evaluation based on feature extraction: training dataset is inputted based on feature extraction
Bi-LSTM network model, Definition Model loss function, and the call back function for only retaining optimal models is set, to minimize model
For the purpose of loss function, using back-propagation algorithm loop iteration, the optimal Bi-LSTM net based on feature extraction is finally saved
Network model, and test data set is inputted into the model and carries out model evaluation；
4) it realizes that road traffic state prediction is rapid based on real-time road traffic status information: obtaining real-time road traffic state square
Battle array is standardized the input operated and as the Bi-LSTM network model based on feature extraction to data are obtained, by model
Output data carries out anti-normalizing operation, realizes real-time road traffic status predication.
2. a kind of method of Bi-LSTM prediction road traffic state based on feature extraction as described in claim 1, feature
It is, the process of the step 1) is as follows:
Step 1.1: determining a plurality of section and the time range of prediction, the road condition number in all sections within the scope of extraction time
According to by the status data in all sections according to the progress data merging of same time sequence index, construct shows the road of space time information
Road traffic behavior matrix；
The vehicle flowrate data and speed data for extracting the n moment within the scope of m section same time of prediction, wherein i-th road
The vehicle flowrate data of section are denoted as respectively with speed data: si=[si1,si2,si3…sin]T, vi=[vi1,vi2,vi3…vin]T, will
The vehicle flowrate data and speed data in all sections carry out data merging according to same time sequence index, the road like after merging
State matrix X=[s1,v1,s2,v2…sm,vm], wherein the road traffic state X of t momentt=[s1t,v1t,s2t,v2t…smt,
vmt], t=1,2,3 ... n；
Step 1.2: data set being trained to road traffic state data matrix and test data set divides, and respectively to training
Data set and test data set are standardized operation
By road traffic state matrix X according to a:(1-a) ratio cut partition be training set XtrainWith test set Xtest, wherein a ∈ (0,
1), Xtrain=[X1,X2,X3…Xna]T, Xtest=[Xna+1,Xna+2,Xna+3…Xn]T, to the training set and test set point after division
Not carry out deviation normalizing operation, scale the vehicle flowrate in each section proportionally with speed data, be uniformly mapped to [0,1]
In section, deviation standardized calculation formula are as follows:
Wherein,The minimum value of i-th section vehicle flowrate and speed is respectively represented, Respectively represent i-th
The maximum value of section vehicle flowrate and speed,The vehicle flowrate of i-th article of section t moment after being expressed as deviation standardization
And speed data, it is respectively after training set and the standardization of test set deviation
3. a kind of method of Bi-LSTM prediction road traffic state based on feature extraction as claimed in claim 1 or 2,
It is characterized in that, the process of the step 2) is as follows:
Step 2.1: extracting the space-time characteristic of road traffic state matrix
The SAE that the sparse self-encoding encoder of N layers of building stacks, N >=2, by road traffic state matrixSAE model is inputted, is extracted
The space-time characteristic of road traffic state matrix, the calculation expression of first layer self-encoding encoder are as follows:
WhereinThe parameter learnt, X ' are needed for modeltrainIt is approximately equal to mode inputσ is Sigmoid function, function expression are as follows:
Through first layer self-encoding encoder to road traffic state matrixSingle order hidden layer characterization is obtained after carrying out space-time characteristic extraction
Feature h1, by h1As the input of next self-encoding encoder, second order hidden layer characteristic feature h is obtained after encoded device f encoding operation2,
It is exported by model is obtained after decoder r decodingAndIt is approximately equal to h1；Similarly, this layer of sparse self-encoding encoder study is arrived
Input of the hidden layer characteristic feature as next coefficient self-encoding encoder, carry out feature learning, layer by layer by N layers of self-encoding encoder
After coding and decoding operation, traffic behavior matrix N rank space-time characteristic feature h is extractedN；
Step 2.2: the Bi-LSTM network model building based on feature extraction
By traffic behavior matrix space-time characteristic feature hNAs the input of Bi-LSTM model, Bi-LSTM network model is using double-deck
LSTM building, it is respectively preceding to LSTM unit and backward LSTM unit, the calculating process of LSTM unit are as follows:
ft=σ (Wfht-1+WfhN+b1) (6)
it=σ (Wiht-1+WihN+b2) (7)
ot=σ (Woht-1+WohN+b4) (10)
ht=ot⊙tanh(Ct) (11)
Wherein, ⊙ represents dot product, Wf、Wi、Wc、WoIt is expressed as forgeing door, input gate, the weight square of state cell, out gate
Battle array, b1, b2, b3, b4 are expressed as forgeing the biasing size of door, input gate, state cell, out gate, and σ swashs for Sigmoid
Function living, tanh activation primitive expression formula are as follows:
T moment, traffic behavior matrix space-time characteristic feature hNIt is through the preceding feature extracted to LSTM unitTo LSTM after
Unit extract feature beBi-LSTM is exported after carrying out high dimensional feature extraction:
Wherein W*, b*Respectively weight matrix and biasing size, σ are Sigmoid activation primitive；
Input by the output of Bi-LSTM model as regression forecasting layer, completes the Bi-LSTM network model based on feature extraction
Building, regression forecasting layer functions expression formula are as follows:
Yout=σ (WoutYl+bout) (14)
Wherein Yl=[y1,y2,y3…yl]T, Wout, boutFor the weight matrix and biasing size of regression forecasting layer, σ Sigmoid
Activation primitive, l are the output unit number of Bi-LSTM network.
4. a kind of method of Bi-LSTM prediction road traffic state based on feature extraction as claimed in claim 1 or 2,
It is characterized in that, the process of the step 3) is as follows:
In t moment, by training datasetBi-LSTM network model is inputted, through feature extraction and layer prediction is returned, obtains
Traffic behavior of i-th road at the following τ moment:
Actual traffic state of i-th road at the following τ moment:
Definition Model loss function:
For the purpose of minimizing model loss function, using back-propagation algorithm loop iteration, obtain optimal mentioning based on feature
The Bi-LSTM network model taken, by test data setThe optimal Bi-LSTM network model based on feature extraction is inputted,
It is compared according to the output of model with true value, model is evaluated.
5. a kind of method of Bi-LSTM prediction road traffic state based on feature extraction as claimed in claim 1 or 2,
It is characterized in that, the process of the step 4) is as follows:
Road real-time traffic states matrix is obtained, is standardized operation and as the Bi- based on feature extraction to data are obtained
The input of LSTM network model obtains the traffic status prediction value of i-th road:
Model prediction result is subjected to anti-normalizing operation, obtains predicting true vehicle flowrate and speed state, anti-standardized calculation
Formula are as follows:
Wherein,The minimum value of i-th section vehicle flowrate and speed is respectively represented, Respectively represent i-th
The maximum value of section vehicle flowrate and speed, vi(t+j), si(t+j)Vehicle flowrate and speed of the respectively i-th article of road at (t+j) moment
Degree prediction size, realizes the prediction of vehicle flowrate and speed in short time traffic conditions.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910593469.3A CN110516833A (en) | 2019-07-03 | 2019-07-03 | A method of the Bi-LSTM based on feature extraction predicts road traffic state |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910593469.3A CN110516833A (en) | 2019-07-03 | 2019-07-03 | A method of the Bi-LSTM based on feature extraction predicts road traffic state |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110516833A true CN110516833A (en) | 2019-11-29 |
Family
ID=68623150
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910593469.3A Pending CN110516833A (en) | 2019-07-03 | 2019-07-03 | A method of the Bi-LSTM based on feature extraction predicts road traffic state |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110516833A (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110956808A (en) * | 2019-12-23 | 2020-04-03 | 北京交通大学 | Heavy truck traffic flow prediction method based on non-full-sample positioning data |
| CN111193254A (en) * | 2019-12-06 | 2020-05-22 | 北京国电通网络技术有限公司 | Residential daily electricity load prediction method and device |
| CN111275971A (en) * | 2020-02-18 | 2020-06-12 | 山西交通控股集团有限公司 | Expressway traffic flow prediction method |
| CN111475987A (en) * | 2020-04-02 | 2020-07-31 | 重庆大学 | Gear residual life prediction method based ON SAE and ON-L STM |
| CN111556017A (en) * | 2020-03-25 | 2020-08-18 | 中国科学院信息工程研究所 | Network intrusion detection method based on self-coding machine and electronic device |
| CN111553535A (en) * | 2020-05-06 | 2020-08-18 | 东华大学 | Navigation reference method based on AE-LSTM-BO traffic flow prediction |
| CN111554118A (en) * | 2020-04-24 | 2020-08-18 | 深圳职业技术学院 | Dynamic prediction method and system for bus arrival time |
| CN111862592A (en) * | 2020-05-27 | 2020-10-30 | 浙江工业大学 | Traffic flow prediction method based on RGCN |
| CN112268312A (en) * | 2020-10-23 | 2021-01-26 | 哈尔滨派立仪器仪表有限公司 | Intelligent heat supply management system based on deep learning |
| WO2021109318A1 (en) * | 2019-12-03 | 2021-06-10 | 东南大学 | Method for estimating and predicting short-term traffic circulation state of urban road network |
| CN114358422A (en) * | 2022-01-04 | 2022-04-15 | 中国工商银行股份有限公司 | Research and development progress abnormity prediction method and device, storage medium and electronic equipment |
| CN114387782A (en) * | 2022-01-12 | 2022-04-22 | 智道网联科技(北京)有限公司 | Method and device for predicting traffic state and electronic equipment |
| WO2022241802A1 (en) * | 2021-05-19 | 2022-11-24 | 广州广电运通金融电子股份有限公司 | Short-term traffic flow prediction method under complex road network, storage medium, and system |
| WO2024087281A1 (en) * | 2022-10-25 | 2024-05-02 | 东南大学 | Highway toll station lane allocation method considering short-term vehicle flow |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107230351A (en) * | 2017-07-18 | 2017-10-03 | 福州大学 | A kind of Short-time Traffic Flow Forecasting Methods based on deep learning |
| CN108510741A (en) * | 2018-05-24 | 2018-09-07 | 浙江工业大学 | A kind of traffic flow forecasting method based on Conv1D-LSTM neural network structures |
-
2019
- 2019-07-03 CN CN201910593469.3A patent/CN110516833A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107230351A (en) * | 2017-07-18 | 2017-10-03 | 福州大学 | A kind of Short-time Traffic Flow Forecasting Methods based on deep learning |
| CN108510741A (en) * | 2018-05-24 | 2018-09-07 | 浙江工业大学 | A kind of traffic flow forecasting method based on Conv1D-LSTM neural network structures |
Non-Patent Citations (2)
| Title |
|---|
| BAILINYAN ET AL.: "Traffic Flow Prediction Using LSTM with Feature Enhancement", 《NEUROCOMPUTING》 * |
| 成云: "面向高速公路的交通流预测算法研究", 《中国优秀博硕士学位论文全文数据库(硕士)工程科技Ⅱ辑》 * |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021109318A1 (en) * | 2019-12-03 | 2021-06-10 | 东南大学 | Method for estimating and predicting short-term traffic circulation state of urban road network |
| CN111193254A (en) * | 2019-12-06 | 2020-05-22 | 北京国电通网络技术有限公司 | Residential daily electricity load prediction method and device |
| CN111193254B (en) * | 2019-12-06 | 2021-10-29 | 北京国电通网络技术有限公司 | Residential daily electricity load prediction method and device |
| CN110956808A (en) * | 2019-12-23 | 2020-04-03 | 北京交通大学 | Heavy truck traffic flow prediction method based on non-full-sample positioning data |
| CN111275971A (en) * | 2020-02-18 | 2020-06-12 | 山西交通控股集团有限公司 | Expressway traffic flow prediction method |
| CN111556017A (en) * | 2020-03-25 | 2020-08-18 | 中国科学院信息工程研究所 | Network intrusion detection method based on self-coding machine and electronic device |
| CN111556017B (en) * | 2020-03-25 | 2021-07-27 | 中国科学院信息工程研究所 | Network intrusion detection method based on self-coding machine and electronic device |
| CN111475987A (en) * | 2020-04-02 | 2020-07-31 | 重庆大学 | Gear residual life prediction method based ON SAE and ON-L STM |
| CN111475987B (en) * | 2020-04-02 | 2024-05-24 | 重庆大学 | SAE and ON-LSTM-based gear residual life prediction method |
| CN111554118B (en) * | 2020-04-24 | 2022-01-25 | 深圳职业技术学院 | Dynamic prediction method and system for bus arrival time |
| CN111554118A (en) * | 2020-04-24 | 2020-08-18 | 深圳职业技术学院 | Dynamic prediction method and system for bus arrival time |
| CN111553535B (en) * | 2020-05-06 | 2022-07-29 | 东华大学 | Navigation reference method based on AE-LSTM-BO traffic flow prediction |
| CN111553535A (en) * | 2020-05-06 | 2020-08-18 | 东华大学 | Navigation reference method based on AE-LSTM-BO traffic flow prediction |
| CN111862592A (en) * | 2020-05-27 | 2020-10-30 | 浙江工业大学 | Traffic flow prediction method based on RGCN |
| CN112268312A (en) * | 2020-10-23 | 2021-01-26 | 哈尔滨派立仪器仪表有限公司 | Intelligent heat supply management system based on deep learning |
| WO2022241802A1 (en) * | 2021-05-19 | 2022-11-24 | 广州广电运通金融电子股份有限公司 | Short-term traffic flow prediction method under complex road network, storage medium, and system |
| CN114358422A (en) * | 2022-01-04 | 2022-04-15 | 中国工商银行股份有限公司 | Research and development progress abnormity prediction method and device, storage medium and electronic equipment |
| CN114387782A (en) * | 2022-01-12 | 2022-04-22 | 智道网联科技(北京)有限公司 | Method and device for predicting traffic state and electronic equipment |
| WO2024087281A1 (en) * | 2022-10-25 | 2024-05-02 | 东南大学 | Highway toll station lane allocation method considering short-term vehicle flow |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110516833A (en) | A method of the Bi-LSTM based on feature extraction predicts road traffic state | |
| CN112418547B (en) | Bus stop passenger flow prediction method based on GCN-LSTM combination model | |
| CN110223517B (en) | Short-term traffic flow prediction method based on space-time correlation | |
| CN111210633B (en) | Short-term traffic flow prediction method based on deep learning | |
| CN110021165A (en) | A kind of traffic flow forecasting method based on Autoencoder-LSTM Fusion Model | |
| CN113053115B (en) | Traffic prediction method based on multi-scale graph convolution network model | |
| CN108009674A (en) | Air PM2.5 concentration prediction methods based on CNN and LSTM fused neural networks | |
| CN109816095A (en) | Based on the network flow prediction method for improving gating cycle neural network | |
| CN106971240A (en) | The short-term load forecasting method that a kind of variables choice is returned with Gaussian process | |
| CN111292525B (en) | Traffic flow prediction method based on neural network | |
| CN110070713A (en) | A kind of traffic flow forecasting method based on two-way nested-grid ocean LSTM neural network | |
| CN111815033A (en) | Offshore wind power prediction method based on RCNN and meteorological time sequence characteristics | |
| CN110070715A (en) | A kind of road traffic flow prediction method based on Conv1D-NLSTMs neural network structure | |
| CN109829495A (en) | Timing image prediction method based on LSTM and DCGAN | |
| CN110232461A (en) | More interconnection vector machine water quality prediction methods based on quantum genetic algorithm optimization | |
| CN112949896B (en) | Time sequence prediction method based on fusion sequence decomposition and space-time convolution | |
| CN109086926B (en) | Short-time rail transit passenger flow prediction method based on combined neural network structure | |
| CN113689030B (en) | Short-term wind power prediction method based on bidirectional attention and secondary optimization | |
| CN111626764A (en) | Commodity sales volume prediction method and device based on Transformer + LSTM neural network model | |
| CN110837888A (en) | Traffic missing data completion method based on bidirectional cyclic neural network | |
| Kogekar et al. | A CNN-BiLSTM-SVR based deep hybrid model for water quality forecasting of the river Ganga | |
| CN112580859B (en) | Haze prediction method based on global attention mechanism | |
| Tirumala | Evolving deep neural networks using coevolutionary algorithms with multi-population strategy | |
| CN114565187A (en) | Traffic network data prediction method based on graph space-time self-coding network | |
| CN114781744A (en) | Deep learning multi-step long radiance prediction method based on codec |
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 |
Application publication date: 20191129 |
|
| RJ01 | Rejection of invention patent application after publication |