CN109118763A - Vehicle flowrate prediction technique based on corrosion denoising deepness belief network - Google Patents
Vehicle flowrate prediction technique based on corrosion denoising deepness belief network Download PDFInfo
- Publication number
- CN109118763A CN109118763A CN201810986737.3A CN201810986737A CN109118763A CN 109118763 A CN109118763 A CN 109118763A CN 201810986737 A CN201810986737 A CN 201810986737A CN 109118763 A CN109118763 A CN 109118763A
- Authority
- CN
- China
- Prior art keywords
- corrosion
- layer
- vehicle flowrate
- boltzmann machine
- prediction
- 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.)
- Granted
Links
- 230000007797 corrosion Effects 0.000 title claims abstract description 165
- 238000005260 corrosion Methods 0.000 title claims abstract description 165
- 238000000034 method Methods 0.000 title claims abstract description 36
- 210000002569 neuron Anatomy 0.000 claims abstract description 16
- 238000012549 training Methods 0.000 claims description 51
- 230000006870 function Effects 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 16
- 230000001771 impaired effect Effects 0.000 claims description 10
- 238000004422 calculation algorithm Methods 0.000 claims description 8
- 238000009826 distribution Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 6
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 claims description 5
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 claims description 5
- 230000006378 damage Effects 0.000 claims description 5
- 230000000694 effects Effects 0.000 claims description 5
- 230000001537 neural effect Effects 0.000 claims description 5
- 238000011144 upstream manufacturing Methods 0.000 claims description 5
- 238000004519 manufacturing process Methods 0.000 claims description 4
- 238000001514 detection method Methods 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000002474 experimental method Methods 0.000 claims description 3
- 230000000717 retained effect Effects 0.000 claims description 3
- 238000007477 logistic regression Methods 0.000 claims description 2
- 230000002123 temporal effect Effects 0.000 abstract description 6
- 238000011161 development Methods 0.000 abstract description 4
- 238000005457 optimization Methods 0.000 abstract description 2
- 239000010410 layer Substances 0.000 description 115
- 238000013135 deep learning Methods 0.000 description 4
- 239000011159 matrix material Substances 0.000 description 4
- 238000003062 neural network model Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000011521 glass Substances 0.000 description 2
- 230000002779 inactivation Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 239000003205 fragrance Substances 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 239000011229 interlayer Substances 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000010534 mechanism of action Effects 0.000 description 1
- 230000000116 mitigating effect Effects 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
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
- G06N3/08—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"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/40—Business processes related to the transportation industry
-
- 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
- G08G1/0137—Measuring and analyzing of parameters relative to traffic conditions for specific applications
- G08G1/0145—Measuring and analyzing of parameters relative to traffic conditions for specific applications for active traffic flow control
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Data Mining & Analysis (AREA)
- Analytical Chemistry (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Chemical & Material Sciences (AREA)
- Marketing (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- General Business, Economics & Management (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Tourism & Hospitality (AREA)
- Development Economics (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Primary Health Care (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A kind of vehicle flowrate prediction technique based on corrosion denoising deepness belief network, vehicle flowrate prediction refers to the vehicle flowrate situation using the historical juncture to predict the vehicle flowrate situation at current and future moment, the present invention proposes random corrosion layer as a kind of regularization means, release the interdependence between a part of neuron, prediction model generalization ability is improved, over-fitting risk is reduced.In combination with concrete application scene, it is contemplated that the spatial coherence of vehicle flowrate and temporal regularity establish vehicle flowrate prediction model, realize that accurate, reliable, real-time vehicle flowrate prediction, effectively optimization traffic scheduling alleviate traffic pressure, improve the operational efficiency of road network.This has great significance to the development of intelligent transportation.
Description
Technical field
The invention belongs to field of artificial intelligence, are related to intelligent transportation and neural network learning systems technology, for one kind
Vehicle flowrate prediction technique based on corrosion denoising deepness belief network.
Background technique
With the development of urbanization, traffic congestion has become a universal phenomenon, this has seriously affected the convenient of trip
Property.Accurately, vehicle flowrate is predicted to be current intelligent transportation urgent problem to be solved reliably, in real time, this can effectively optimize traffic tune
Degree alleviates traffic pressure, improves the operational efficiency of road network, promotes people's lives comfort level and fragrance.In recent years, artificial intelligence
The development of energy has also pushed the flow of research of vehicle flowrate forecasting problem, and the method based on deep learning starts to be applied to intelligent friendship
Logical field.The model robustness established based on such method is good, and learning ability is strong, can cope with traffic conditions complicated and changeable.It is logical
The method using deep learning is crossed, in conjunction with concrete application scene, corresponding vehicle flowrate prediction model is established, to obtain preferable wagon flow
Estimated performance is measured, has great significance to the development of intelligent transportation.
Vehicle flowrate prediction of the invention refers to the vehicle flowrate situation using the historical juncture to predict the current and future moment
Vehicle flowrate situation.Most common vehicle flowrate prediction technique is time series models in early days, and as a kind of linear model, it has knot
Structure is simple, the fast advantage of calculating speed, but it cannot handle traffic conditions complicated and changeable well, and precision of prediction is low.Traffic
The potential rule of flow data be it is nonlinear, can freely learn any from training data using non-parametric machine learning method
The function of form, preferably fitting prediction model.In machine learning, a variety of regression algorithms can be effectively used for forecasting traffic flow, such as
Supporting vector machine model, K arest neighbors regression model, Random Forest model, neural network model etc..Neural network model is a kind of
The model of human brain neural network is simulated, learning ability is powerful, does not need hand-designed feature, it is only necessary to input initial data, mould
Type can learn corresponding Nonlinear Mapping relationship out, therefore be widely used.
With the increase of the prototype network number of plies, the mapping relations that neural network model can learn are based on regard to more complicated
The training method of conventional counter propagation algorithm often makes deep layer network fall into locally optimal solution, reduces model performance, depth letter
Network is read with layer-by-layer greedy pre-training algorithm to solve this problem.
When not taking pre-training strategy and directly deep layer network trained with the Back Propagation Algorithm based on gradient descent method,
If weight parameter is too big when netinit, the awkward situation of local minimum is often fallen into, if weight when netinit
Parameter is too small, will cause the phenomenon that gradient disappears.The successively method of greediness pre-training deep layer network, can make network comparatively fast find
Globe optimum is a kind of appropriate initialization strategy.However, the overfitting problem in network does not still solve very well.
Summary of the invention
The problem to be solved in the present invention is: neural network learning technology is more and more used for vehicle flowrate and predicts, but mesh
Preceding neural net prediction method could be improved, to meet the technical need of vehicle flowrate prediction.
The technical solution of the present invention is as follows: the vehicle flowrate prediction technique based on corrosion denoising deepness belief network, is based on depth
Belief network predicts the vehicle flowrate at current and future moment according to the vehicle flowrate of historical juncture, by the vehicle flowrate number of historical juncture
According to training set and test set is divided into, using training set training deepness belief network model, the prediction of test set test model is used
Performance obtains trained vehicle flowrate prediction model, and the vehicle flowrate for the current and future moment is predicted；Wherein, depth conviction
The middle layer of network is limited Boltzmann machine by the corrosion removal stacked and constitutes, and it is limited that the corrosion removal, which is limited Boltzmann machine,
The input terminal of Boltzmann machine adds a random corrosion layer, using the impaired output of this random corrosion layer as limited Boltzmann
The visible layer of machine, hidden layer, which is not done, to be changed.
Further, random corrosion layer is realized by setting corrosion probability, and corrosion probability is a global hyper parameter, corrosion
Probability is smaller, and more multi-neuron is retained, when corrosion probability is 0, random corrosion layer be degenerated to one it is common identical
Image Planes, output only simply copy input；Corrosion probability is bigger, and more multi-neuron loses activity, the pass between neuron
Connection property is weaker, and feature learning is more difficult, is determined by experiment reasonable corrosion probability value.
It is preferred that when establishing vehicle flowrate prediction model, in conjunction with the information of vehicle flowrate of upstream and downstream and adjacent segments
Relevance and current time information of vehicle flowrate obtain prediction model frame to the dependence of historical juncture information of vehicle flowrate are as follows:
The bottom is prediction model data input layer, is inputted as X1, t-1, X1, t-2..., X1, t-d, X2, t-1, X2, t-2...,
XM, t-d, wherein XI, jI-th of wagon detector is indicated in the vehicle flowrate at j moment, i=1 ..., m, m is total for wagon detector, t
For prediction time, j=t-1 ... t-d, the i.e. input of prediction model be in road network all associated vehicle detectors at current time
All information of vehicle flowrate between to the preceding d moment；
Middle layer is that the corrosion removal stacked is limited Boltzmann machine, and the corrosion removal is limited Boltzmann machine and is based on for one kind
The production stochastic neural net of energy function, whole network are divided into two layers: visible layer and hidden layer, it is seen that layer is limited Bohr
The hereby input layer of graceful machine, hidden layer are the feature extraction layer of limited Boltzmann machine；When pre-training vehicle flowrate prediction model, going
The front end for corroding limited Boltzmann machine is a random corrosion layer, and input data is introduced into random corrosion layer, after corrosion
Impaired output is used as visible layer, and vehicle flowrate prediction model finely tunes and do not have random corrosion layer when testing；
Top is prediction model logistic regression output layer, is exported as Y1, Y2, Y3..., Ym, wherein YiIndicate i-th of vehicle
Vehicle flowrate of the detector in prediction time t.
It is preferred that training deepness belief network model specifically:
Step 1: pre-training corrosion removal is limited Boltzmann machine: setting corrosion probability, inputs as vehicle flowrate data, into heap
Folded first layer corrosion removal is limited the random corrosion layer of Boltzmann machine, is corroded, is obtained impaired with preset corrosion probability
Visible layer of the output as the limited Boltzmann machine of corrosion removal, obtains hidden layer character representation after energy generating function, then
This hidden layer character representation carries out parameter update using log-likelihood function, makes parameter by the reconstruct input of energy generating function
Under the conditions of limited Boltzmann machine probability distribution it is as eligible as possible, the output of hidden layer feature after pre-training is as next
Corrosion removal is limited the input of Boltzmann machine；
Step 2: fixed pre-trained good corrosion removal is limited the weight and offset parameter of Boltzmann machine, starts pre- instruction
Practice next corrosion removal and be limited Boltzmann machine, the input terminal that next corrosion removal is limited Boltzmann machine also closely follows one at random
Corrosion layer, input are corroded with identical default corrosion probability, and for example previous corrosion removal of next training process is limited Bohr
Hereby graceful machine, and so on, the output that each corrosion removal later is limited Boltzmann machine is introduced into next corrosion removal and is limited glass
The random corrosion layer of the graceful machine of Wurz, then the input after destruction is continued to train as visible layer；
Step 3: after all corrosion removals of pre-training are limited Boltzmann machine, at the top of network model plus one layer of prediction is returned
Layer is predicted for vehicle flowrate；
Step 4: thering is supervision to finely tune entire network model with Back Propagation Algorithm, first three period only updates the last layer net
The weight and offset parameter of network, then update all layers of parameter, obtain final trained deepness belief network model.
It is preferred that using test set test model estimated performance when, evaluation criterion use average absolute percentage
Error MAPE:
Wherein YiIt is practical vehicle flowrate,It is prediction vehicle flowrate, N is test sample number.
When training deepness belief network model, the hyper parameter that needs to adjust are as follows: the corrosion removal of stacking is limited Boltzmann machine
Quantity Nlayer, each corrosion removal be limited the hidden node quantity N of Boltzmann machinenode, each corrosion removal be limited Boltzmann machine
Pre-training period Nepoch, prediction current time vehicle flowrate needed for historical time segment number d and corrosion probability Clevel；With net
Lattice search determines that hyper parameter is arranged according to MAPE error function, and to reduce search space, all corrosion removals are limited Bohr hereby
The hidden node quantity N of graceful machinenodeIdentical, pre-training period NepochIt is identical, corrosion probability Clevel is identical.
Corrosion layer of the invention is considered as a kind of regularization means.Its mechanism of action is the nerve for corroding this layer at random
Member, i.e., each node have certain probability to be damaged inactivation.This probability is the corrosion probability pre-set.Before not corroding,
Each neuron can participate in the training of network, mutually coordinated, some neuron is to the extraction of feature by other dependence mind
Influence through member, there are complicated correlations.This complicated correlation is one of the main reason for leading to over-fitting.Random corrosion layer
The interdependence between a part of neuron can be removed, the neuron co-ordination remained is forced, weakens fixed correlation,
Promote network robustness and generalization ability.
This corrosion is random random, so a different visible layer can be obtained each cycle of training, into
One step, obtain the limited Boltzmann machine of different structure.Such operation, which is equivalent to, has trained several heterogeneous networks structures, so
The result of average these types of network structure afterwards.The operation of this average heterogeneous networks structure can improve model generalization ability, to subtracting
Light over-fitting is helpful.
The method that the present invention utilizes deep learning, is based on deepness belief network, and training basic building unit is limited Bohr hereby
When graceful machine, random corrosion layer is added in input terminal, denoising mechanism is merged, improves network generalization, reduce the risk of over-fitting.
In combination with concrete application scene, it is noted that the relevance of the information of vehicle flowrate of upstream and downstream and adjacent segments and current time
Information of vehicle flowrate is to the dependence of historical juncture information of vehicle flowrate, in network structure design by the space correlation of information of vehicle flowrate
Property and temporal regularity take into account, and establish the vehicle flowrate prediction neural network model for having high accuracy.This can be with
Good anticipation is provided for traffic scheduling, alleviates traffic pressure.
Detailed description of the invention
Fig. 1 is that corrosion removal is limited Boltzmann machine training structure figure.
Fig. 2 is the vehicle flowrate prediction model structure chart based on corrosion denoising deepness belief network.
Fig. 3 is the vehicle flowrate prediction model training flow chart based on corrosion denoising deepness belief network.
Fig. 4 is the method for the present invention working day day car volume forecasting effect picture.
Fig. 5 is the method for the present invention working day continuous five overhead traveling cranes volume forecasting effect picture.
Specific embodiment
The method that the present invention utilizes deep learning, is transformed conventional depth belief network, is had more with further
Representative feature improves model generalization ability, effectively mitigation overfitting problem.Conventional depth belief network is limited by what is stacked
Boltzmann machine is built-up, the innovation of the invention consists in that, input terminal when training in each limited Boltzmann machine adds one
A random corrosion layer, using the output of this random corrosion layer as new visible layer, hidden layer, which is not done, to be changed.
Corrosion probability is a global hyper parameter.Corrosion probability is smaller, and more multi-neuron is retained, when corrosion probability is 0
When, random corrosion layer is degenerated to a common identical Image Planes, and output only simply copies input；Corrosion probability is got over
Greatly, more multi-neuron loses activity, and the relevance between neuron is weaker, and feature learning is more difficult.So needing to pass through experiment
Reasonable corrosion probability value is set.
Input after corrosion layer, limited Boltzmann machine obtain a part of nodes inactivation input layer, be equivalent to by
Noise pollution has been arrived, so they not only want the Energy distribution of analog network node, also to have removed the influence of corrosion noise.It incite somebody to action this
The novel Boltzmann machine that invention proposes is referred to as corrosion removal and is limited Boltzmann machine, and this corrosion removal of training is limited Boltzmann
Function forces Hidden unit to acquire more robust feature, obtains the stronger network of generalization.
Based on above-mentioned several points, therefore the vehicle flowrate prediction technique of the invention based on corrosion denoising deepness belief network can be with
Traffic conditions complicated and changeable are coped with, while in view of the spatial coherence of information of vehicle flowrate and time rule when model structure design
Rule property, further increases prediction accuracy.
Concrete model frame of the invention are as follows:
The bottom is model data input layer, is inputted as X1, t-1, X1, t-2..., X1, t-d；X2, t-1, X2, t-2...；
XM, t-1..., XM, t-d, wherein XI, jI-th of wagon detector is indicated in the vehicle flowrate at j moment, i=1 ..., m, m is vehicle inspection
Survey device sum, t is prediction time, j=t-1 ... t-d, the i.e. input of model be in road network all associated vehicle detectors preceding
The information of vehicle flowrate of d period.This fully demonstrates the spatial coherence and temporal regularity that model considers vehicle flowrate, empty
Between correlation show as upstream and downstream section vehicle flowrate and adjacent segments vehicle flowrate and can largely influence currently to be predicted section
Vehicle flowrate, temporal regularity show as having apparent vehicle flowrate tendency information between the continuous period；
Middle layer is that the corrosion removal stacked is limited Boltzmann machine and the most important basic building unit of model.Go corruption
Losing limited Boltzmann machine is a kind of production stochastic neural net based on energy function, and whole network is divided into two layers: visible
Layer and hidden layer, it is seen that layer is the input layer of limited Boltzmann machine, and hidden layer is the feature extraction layer of limited Boltzmann machine,
In the situation known to visible layer state, all hidden nodes are conditional samplings, similarly, in the situation known to hidden layer state
Under, all visible node layers are conditional samplings；
It is a random corrosion in the front end that corrosion removal is limited Boltzmann machine when pre-training vehicle flowrate prediction model
Layer, input data are introduced into random corrosion layer, and the impaired output after corrosion is as visible layer, the fine tuning of vehicle flowrate prediction model and survey
There is no random corrosion layer when examination；
Top is that model logic returns output layer, is exported as Y1, Y2, Y3..., Ym, wherein YiIndicate i-th of vehicle detection
Vehicle flowrate of the device in prediction time t.
Below in conjunction with attached drawing, the present invention is further described.
Corrosion removal is limited Boltzmann machine training process as shown in Figure 1, input first passes around a random corrosion layer, each
Neuron has certain probability to be damaged by corrosion, and releases the complicated correlation between a part of neuron, enhances generalization ability.Institute
With compared to the limited Boltzmann machine of tradition, corrosion removal, which is limited Boltzmann machine, will not only simulate hidden layer and visible layer Energy distribution
State, the corrosion noise that also remove input influence, it has to the useful information for the neuron for going study to remain at random, most
After learn more expressive feature.
Random corrosion process:
maskp~Bernoulli (1-Clevel) (1)
vp=maskp*xp (2)
Wherein Clevel is corrosion probability, maskpIt is the neuron state after random corrosion layer, meets
The distribution of Bernoulli stochastic variable, the probability that the probability that each variable has 1-Clevel is 1, Clever is 0, and state 1 indicates mind
It is not damaged through member is intact, state 0 indicates that neuron is corroded damage.xpIndicate original complete input, vpIndicate by corrosion layer it
Corrosion removal afterwards is limited the visible layer of Boltzmann machine.
When training corrosion removal is limited Boltzmann machine, damaged data enters limited Boltzmann machine, forces Hidden unit sharp
More robust feature is acquired with the neuron connection relationship remained at random, reconstructs and original does not damage data.This step
Rapid network is divided into two layers: visible layer and hidden layer.Visible layer is impaired input, and hidden layer is feature extraction layer.Limited glass ear
The hereby energy function of graceful machine are as follows:
Wherein v is visible layer unit, and h is Hidden unit, and a and b are the biasing of visible layer unit and Hidden unit respectively, under
Marking p and q is unit number, and w is weight matrix.
Corrosion removal, which is limited Boltzmann machine interlayer, connection, and connectionless in layer.Therefore, the feelings known to visible layer state
Under condition, all hidden nodes are conditional samplings, and similarly, in the situation known to hidden layer state, all visible node layers are items
Part is independent.So there is the following conditions new probability formula:
Wherein sigm (x) is a S type logical function, is 1/ [1+exp (x)].
Pre-training corrosion removal is limited the process of Boltzmann machine as shown in Figure 1, specifically describing are as follows:
Step 1: input is introduced into random corrosion layer, and each node has certain probability to be damaged by corrosion, and releases a part mind
Through the interdependence between member, the visible layer list of Boltzmann machine is then limited using the damaged data of output as corrosion removal
Member；
Step 2: random initializtion weight matrix w and bias vector a, b, according to visible layer unit v0State, pass through formula
(4), Hidden unit h is obtained0State；
Step 3: according to Hidden unit h0State, visible layer unit v is reconstructed by formula (5)1State；
Step 4: formula (4) are utilized again, according to visible layer unit v1State, reconstruct Hidden unit h1State；
Step 5: weight matrix w and bias vector a, b are updated using following formula:
Wherein E (vphq)inputIndicate the expectation of input data distribution, E (vphq)reconIndicate the expectation of reconstruct data distribution.
The training process for being limited Boltzmann machine from corrosion removal can see, and whole process does not use the label letter of data
Breath, this is advantageous in the scene for lacking label information.So it is both generation model and an a unsupervised model.
It stacks corrosion removal and is limited Boltzmann machine, customize input layer and output layer, obtain that the present invention is based on depth conviction nets
The vehicle flowrate prediction model of network, as shown in Figure 2.Mainly there are three parts for model of the invention:
The bottom is model data input layer, is inputted as X1, t-1, X1, t-2..., X1, t-d；X2, t-1, X2, t-2...；
XM, t-1..., XM, t-d, wherein XI, jI-th of wagon detector is indicated in the vehicle flowrate at j moment, m indicates wagon detector sum,
I.e. the input of model be road network in all associated vehicle detectors the preceding d period information of vehicle flowrate.This fully demonstrates mould
Type considers the spatial coherence and temporal regularity of vehicle flowrate, and spatial coherence shows as upstream and downstream section vehicle flowrate and phase
Adjacent section vehicle flowrate can largely influence currently to be predicted section vehicle flowrate, and temporal regularity shows as the continuous period
Between have apparent vehicle flowrate tendency information；
Middle layer is that the corrosion removal stacked is limited Boltzmann machine and the most important basic building unit of model, is one
Production stochastic neural net of the kind based on energy function.When pre-training vehicle flowrate prediction model, Bohr is limited hereby in corrosion removal
The front end of graceful machine is a random corrosion layer, and input data is introduced into random corrosion layer, and the impaired output after corrosion is used as can
See that layer, vehicle flowrate prediction model finely tune and do not have random corrosion layer when testing；
Top is that model logic returns output layer, is exported as Y1, Y2, Y3..., Ym, wherein YiIndicate i-th of vehicle detection
Vehicle flowrate of the device in prediction time t.
The whole training process flow chart of vehicle flowrate prediction model based on corrosion denoising deepness belief network of the invention
As shown in figure 3, being summarized as follows:
Step 1: setting corrosion probability inputs as vehicle flowrate data, and the corrosion removal into model first layer is limited Bohr hereby
The random corrosion layer of graceful machine, is corroded with preset corrosion probability, is released the interdependence between a part of neuron, is obtained
Impaired output is limited the visible layer of Boltzmann machine as corrosion removal, obtains hidden layer mark sheet after energy generates model
Show, then this hidden layer character representation is generated model according to energy and obtains reconstruct input, log-likelihood function is utilized according to result
Right value update is carried out, keeps the limited Boltzmann machine probability distribution under Parameter Conditions as eligible as possible, it is hidden after pre-training
Layer feature output is limited the input of Boltzmann machine as next corrosion removal；
Step 2: fixed pre-trained good corrosion removal is limited the weight and offset parameter of Boltzmann machine, starts pre- instruction
Practice second corrosion removal and is limited Boltzmann machine, and so on, each feature extraction output later is all introduced into next
Corrosion removal is limited the random corrosion layer of Boltzmann machine, then the input after destruction is continued to train as visible layer；
Step 3: after all corrosion removals of pre-training are limited Boltzmann machine, at the top of network model plus one layer of prediction is returned
Layer is predicted for vehicle flowrate；
Step 4: thering is supervision to finely tune entire network model with Back Propagation Algorithm, preceding several periods only update the last layer net
The weight and offset parameter of network, then update all layers of parameter.
According to the evaluation criterion MAPE function and test result of model performance, determine the hyper parameter of model: stacking goes corruption
Lose limited Boltzmann machine quantity Nlayer, each corrosion removal be limited the hidden node quantity N of Boltzmann machinenode, each go corruption
Lose the pre-training period N of limited Boltzmann machineepoch, prediction current time vehicle flowrate needed for historical time segment number d and
Corrosion probability Clevel.Then under conditions of determining model specific framework, further training pattern, adjustment weight matrix and partially
Vector is set, model performance is optimal.Wherein, to reduce search space, all corrosion removals are limited the hidden of Boltzmann machine
Node layer quantity NnodeIdentical, pre-training period NepochIt is identical, corrosion probability Clevel is identical.
Corrosion denoising deepness belief network frame of the invention has the following advantages:
First, it is a kind of generative probabilistic model；
Second, it can be used unlabeled data and carrys out unsupervised learning；
Third, first pre-training network, compared to the method for random initializtion, this algorithm can more preferably optimize whole network
Weight, prevent optimization fall into locally optimal solution；
4th, when model pre-training, the front end that each corrosion removal is limited Boltzmann machine is a random corrosion layer, is used
It removes the cross correlation between a part of neuron, improves model generalization ability, study has more representative characteristic, mitigated
Fitting problems.
Fig. 4 is that the vehicle flowrate prediction technique prediction in one day of the invention based on corrosion denoising deepness belief network shows feelings
Condition, Fig. 5 are the prediction techniques of proposition in a Friday workaday prediction performance situation.As can be seen that under the method for the present invention
Prediction curve and practical vehicle flowrate curve co-insides degree are very high, can also have very high prediction quasi- in the case where vehicle flowrate fluctuates biggish situation
True rate.
Claims (6)
1. based on the vehicle flowrate prediction technique of corrosion denoising deepness belief network, it is characterized in that based on deepness belief network according to going through
The vehicle flowrate at history moment predicts the vehicle flowrate at current and future moment, by the vehicle flowrate data of historical juncture be divided into training set and
Test set is trained using training set training deepness belief network model using the estimated performance of test set test model
Vehicle flowrate prediction model, for the current and future moment vehicle flowrate predict；Wherein, the middle layer of deepness belief network is by heap
Folded corrosion removal is limited Boltzmann machine and constitutes, and it is the input in limited Boltzmann machine that the corrosion removal, which is limited Boltzmann machine,
End plus a random corrosion layer, using the impaired output of this random corrosion layer as the visible layer of limited Boltzmann machine, hidden layer
It does not do and changes.
2. the vehicle flowrate prediction technique according to claim 1 based on corrosion denoising deepness belief network, it is characterized in that with
Machine corrosion layer realizes that corrosion probability is a global hyper parameter, and corrosion probability is smaller, more multi-neuron by setting corrosion probability
It is retained, when corrosion probability is 0, random corrosion layer is degenerated to a common identical Image Planes, exports only simple
Ground duplication input；Corrosion probability is bigger, and more multi-neuron loses activity, and the relevance between neuron is weaker, and feature learning is got over
Difficulty is determined by experiment reasonable corrosion probability value.
3. the vehicle flowrate prediction technique according to claim 1 or 2 based on corrosion denoising deepness belief network, it is characterized in that
When establishing vehicle flowrate prediction model, in conjunction with the relevance and current time vehicle of upstream and downstream and the information of vehicle flowrate of adjacent segments
Flow information obtains prediction model frame to the dependence of historical juncture information of vehicle flowrate are as follows:
The bottom is prediction model data input layer, is inputted as X1, t-1, X1, t-2..., X1, t-d, X2, t-1, X2, t-2..., XM, t-d,
Middle XI, jI-th of wagon detector is indicated in the vehicle flowrate at j moment, i=1 ..., m, m is total for wagon detector, and t is prediction
Moment, j=t-1 ... t-d, the i.e. input of prediction model are all associated vehicle detectors in road network at current time to preceding d
All information of vehicle flowrate between a moment；
Middle layer is that the corrosion removal stacked is limited Boltzmann machine, and it is that one kind is based on energy that the corrosion removal, which is limited Boltzmann machine,
The production stochastic neural net of function, whole network are divided into two layers: visible layer and hidden layer, it is seen that layer is limited Boltzmann
The input layer of machine, hidden layer are the feature extraction layer of limited Boltzmann machine；When pre-training vehicle flowrate prediction model, in corrosion removal
The front end of limited Boltzmann machine is a random corrosion layer, and input data is introduced into random corrosion layer, impaired after corrosion
Output is used as visible layer, and vehicle flowrate prediction model finely tunes and do not have random corrosion layer when testing；
Top is prediction model logistic regression output layer, is exported as Y1, Y2, Y3..., Ym, wherein YiIndicate i-th of vehicle detection
Vehicle flowrate of the device in prediction time t.
4. the vehicle flowrate prediction technique according to claim 1 or 2 based on corrosion denoising deepness belief network, it is characterized in that
Training deepness belief network model specifically:
Step 1: pre-training corrosion removal is limited Boltzmann machine: setting corrosion probability, inputs as vehicle flowrate data, into stacking
First layer corrosion removal is limited the random corrosion layer of Boltzmann machine, is corroded with preset corrosion probability, obtains impaired output
It is limited the visible layer of Boltzmann machine as corrosion removal, hidden layer character representation is obtained after energy generating function, then this
Hidden layer character representation carries out parameter update using log-likelihood function, makes Parameter Conditions by the reconstruct input of energy generating function
Under limited Boltzmann machine probability distribution it is as eligible as possible, the output of hidden layer feature after pre-training goes corruption as next
Lose the input of limited Boltzmann machine；
Step 2: fixed pre-trained good corrosion removal is limited the weight and offset parameter of Boltzmann machine, starts under pre-training
One corrosion removal is limited Boltzmann machine, and the input terminal that next corrosion removal is limited Boltzmann machine also closely follows a random corrosion
Layer, input are corroded with identical default corrosion probability, and for example previous corrosion removal of next training process is limited Boltzmann
Machine, and so on, the output that each corrosion removal later is limited Boltzmann machine is introduced into next corrosion removal and is limited Bohr hereby
The random corrosion layer of graceful machine, then the input after destruction is continued to train as visible layer；
Step 3: after all corrosion removals of pre-training are limited Boltzmann machine, at the top of network model plus one layer of prediction returns layer,
It is predicted for vehicle flowrate；
Step 4: thering is supervision to finely tune entire network model with Back Propagation Algorithm, first three period only updates the last layer network
Weight and offset parameter then update all layers of parameter, obtain final trained deepness belief network model.
5. the vehicle flowrate prediction technique according to claim 1 or 2 based on corrosion denoising deepness belief network, it is characterized in that
Using test set test model estimated performance when, evaluation criterion use average absolute percentage error MAPE:
Wherein YiIt is practical vehicle flowrate,It is prediction vehicle flowrate, N is test sample number.
6. the vehicle flowrate prediction technique according to claim 1 or 2 based on corrosion denoising deepness belief network, it is characterized in that
When training deepness belief network model, the hyper parameter that needs to adjust are as follows: the corrosion removal of stacking is limited Boltzmann machine quantity Nlayer、
Each corrosion removal is limited the hidden node quantity N of Boltzmann machinenode, each corrosion removal be limited pre-training week of Boltzmann machine
Phase Nepoch, prediction current time vehicle flowrate needed for historical time segment number d and corrosion probability Clevel；With grid data service,
Determine that hyper parameter is arranged according to MAPE error function, to reduce search space, all corrosion removals are limited Boltzmann machine
Hidden node quantity NnodeIdentical, pre-training period NepochIt is identical, corrosion probability Clevel is identical.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810986737.3A CN109118763B (en) | 2018-08-28 | 2018-08-28 | Vehicle flow prediction method based on corrosion denoising deep belief network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810986737.3A CN109118763B (en) | 2018-08-28 | 2018-08-28 | Vehicle flow prediction method based on corrosion denoising deep belief network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109118763A true CN109118763A (en) | 2019-01-01 |
| CN109118763B CN109118763B (en) | 2021-05-18 |
Family
ID=64860286
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810986737.3A Active CN109118763B (en) | 2018-08-28 | 2018-08-28 | Vehicle flow prediction method based on corrosion denoising deep belief network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109118763B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859477A (en) * | 2019-03-15 | 2019-06-07 | 同盾控股有限公司 | A kind of determination method and apparatus of congestion data |
| CN111523560A (en) * | 2020-03-18 | 2020-08-11 | 第四范式（北京）技术有限公司 | Training method, prediction method, device and system for number prediction model of arriving trucks |
| CN112418504A (en) * | 2020-11-17 | 2021-02-26 | 西安热工研究院有限公司 | Wind speed prediction method based on mixed variable selection optimization deep belief network |
| CN115499467A (en) * | 2022-09-06 | 2022-12-20 | 苏州大学 | Intelligent networking test platform based on digital twin and construction method and system thereof |
| CN117133131A (en) * | 2023-10-26 | 2023-11-28 | 深圳市地铁集团有限公司 | Intelligent traffic control system based on ARM technology system |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101561651B1 (en) * | 2014-05-23 | 2015-11-02 | 서강대학교산학협력단 | Interest detecting method and apparatus based feature data of voice signal using Deep Belief Network, recording medium recording program of the method |
| KR101680241B1 (en) * | 2015-06-23 | 2016-12-06 | 홍익대학교 산학협력단 | Method for predicting human personality based on data regarding human position having specific data type |
| CN106326899A (en) * | 2016-08-18 | 2017-01-11 | 郑州大学 | Tobacco leaf grading method based on hyperspectral image and deep learning algorithm |
| CN106598917A (en) * | 2016-12-07 | 2017-04-26 | 国家海洋局第二海洋研究所 | Upper ocean thermal structure prediction method based on deep belief network |
| CN107451278A (en) * | 2017-08-07 | 2017-12-08 | 北京工业大学 | Chinese Text Categorization based on more hidden layer extreme learning machines |
| CN108175426A (en) * | 2017-12-11 | 2018-06-19 | 东南大学 | A kind of lie detecting method that Boltzmann machine is limited based on depth recursion type condition |
-
2018
- 2018-08-28 CN CN201810986737.3A patent/CN109118763B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101561651B1 (en) * | 2014-05-23 | 2015-11-02 | 서강대학교산학협력단 | Interest detecting method and apparatus based feature data of voice signal using Deep Belief Network, recording medium recording program of the method |
| KR101680241B1 (en) * | 2015-06-23 | 2016-12-06 | 홍익대학교 산학협력단 | Method for predicting human personality based on data regarding human position having specific data type |
| CN106326899A (en) * | 2016-08-18 | 2017-01-11 | 郑州大学 | Tobacco leaf grading method based on hyperspectral image and deep learning algorithm |
| CN106598917A (en) * | 2016-12-07 | 2017-04-26 | 国家海洋局第二海洋研究所 | Upper ocean thermal structure prediction method based on deep belief network |
| CN107451278A (en) * | 2017-08-07 | 2017-12-08 | 北京工业大学 | Chinese Text Categorization based on more hidden layer extreme learning machines |
| CN108175426A (en) * | 2017-12-11 | 2018-06-19 | 东南大学 | A kind of lie detecting method that Boltzmann machine is limited based on depth recursion type condition |
Non-Patent Citations (5)
| Title |
|---|
| CHAIYAPHUM SIRIPANPORNCHANA等: "Travel-time prediction with deep learning", 《2016 IEEE REGION 10 CONFERENCE (TENCON)》 * |
| NITISH SRIVASTAVA等: "Dropout_A simple way to prevent neural networks from overfitting", 《JOURNAL OF MACHINE LEARNING RESEARCH》 * |
| 刘念: "基于深度信念网络的植物叶片识别研究", 《中国优秀硕士学位论文全文数据库·信息科技辑》 * |
| 刘星委 等: "基于深度学习的交通流预测方法可行性研究", 《河北交通教育》 * |
| 王德文 等: "基于贝叶斯正则化深度信念网络的电力变压器故障诊断方法", 《电力自动化设备》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859477A (en) * | 2019-03-15 | 2019-06-07 | 同盾控股有限公司 | A kind of determination method and apparatus of congestion data |
| CN111523560A (en) * | 2020-03-18 | 2020-08-11 | 第四范式（北京）技术有限公司 | Training method, prediction method, device and system for number prediction model of arriving trucks |
| CN111523560B (en) * | 2020-03-18 | 2023-07-25 | 第四范式（北京）技术有限公司 | Method, device and system for training number prediction model of arrival trucks |
| CN112418504A (en) * | 2020-11-17 | 2021-02-26 | 西安热工研究院有限公司 | Wind speed prediction method based on mixed variable selection optimization deep belief network |
| CN112418504B (en) * | 2020-11-17 | 2023-02-28 | 西安热工研究院有限公司 | Wind speed prediction method based on mixed variable selection optimization deep belief network |
| CN115499467A (en) * | 2022-09-06 | 2022-12-20 | 苏州大学 | Intelligent networking test platform based on digital twin and construction method and system thereof |
| CN117133131A (en) * | 2023-10-26 | 2023-11-28 | 深圳市地铁集团有限公司 | Intelligent traffic control system based on ARM technology system |
| CN117133131B (en) * | 2023-10-26 | 2024-02-20 | 深圳市地铁集团有限公司 | Intelligent traffic control system based on ARM technology system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109118763B (en) | 2021-05-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109118763A (en) | Vehicle flowrate prediction technique based on corrosion denoising deepness belief network | |
| Zhou et al. | Explore a deep learning multi-output neural network for regional multi-step-ahead air quality forecasts | |
| Yu et al. | Deep learning: A generic approach for extreme condition traffic forecasting | |
| Pijanowski et al. | Using neural networks and GIS to forecast land use changes: a land transformation model | |
| Jiang et al. | Deepurbanmomentum: An online deep-learning system for short-term urban mobility prediction | |
| CN110164128A (en) | A kind of City-level intelligent transportation analogue system | |
| CN109887282A (en) | A kind of road network traffic flow prediction technique based on level timing diagram convolutional network | |
| Vafakhah | Application of artificial neural networks and adaptive neuro-fuzzy inference system models to short-term streamflow forecasting | |
| Bae et al. | Monthly dam inflow forecasts using weather forecasting information and neuro-fuzzy technique | |
| Sheikh Khozani et al. | Improving Water Quality Index prediction for water resources management plans in Malaysia: application of machine learning techniques | |
| Lin et al. | Air quality forecasting based on cloud model granulation | |
| Oprea et al. | A comparative study of computational intelligence techniques applied to PM2. 5 air pollution forecasting | |
| CN112784479A (en) | Flood flow prediction method | |
| Chen et al. | A short-term traffic flow prediction model based on AutoEncoder and GRU | |
| Rahman et al. | A deep learning approach for network-wide dynamic traffic prediction during hurricane evacuation | |
| CN115860286A (en) | Air quality prediction method and system based on time sequence door mechanism | |
| Zhang et al. | Situational-aware multi-graph convolutional recurrent network (sa-mgcrn) for travel demand forecasting during wildfires | |
| Xiao et al. | Runoff forecasting using machine-learning methods: case study in the middle reaches of Xijiang River | |
| Romlay et al. | Rainfall-runoff model based on ANN with LM, BR and PSO as learning algorithms | |
| CN114582131A (en) | Monitoring method and system based on intelligent ramp flow control algorithm | |
| Wu et al. | Learning spatial–temporal pairwise and high-order relationships for short-term passenger flow prediction in urban rail transit | |
| Vidyarthi et al. | Does ANN really acquire the physics of the system? A study using conceptual components from an established water balance model | |
| Zhao et al. | Exploring the impact of trip patterns on spatially aggregated crashes using floating vehicle trajectory data and graph Convolutional Networks | |
| Yalçın | Weather parameters forecasting with time series using deep hybrid neural networks | |
| Mao et al. | A method of crime rate forecast based on wavelet transform and neural network |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |