CN109598939B - A kind of prediction of short-term traffic volume method based on multitask multiple view learning model - Google Patents
A kind of prediction of short-term traffic volume method based on multitask multiple view learning model Download PDFInfo
- Publication number
- CN109598939B CN109598939B CN201811578615.7A CN201811578615A CN109598939B CN 109598939 B CN109598939 B CN 109598939B CN 201811578615 A CN201811578615 A CN 201811578615A CN 109598939 B CN109598939 B CN 109598939B
- Authority
- CN
- China
- Prior art keywords
- time
- space
- road segment
- prediction
- indicate
- 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.)
- Active
Links
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
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
- G08G1/0133—Traffic data processing for classifying traffic situation
Landscapes
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a kind of prediction of short-term traffic volume methods based on multitask multiple view learning model, include the following steps: step 1, individually construct Spatio-Temporal Data Model for Spatial for each road segment；Step 2 constructs Multiple Kernel Learning model；Step 3 constructs objective function using multitask multiple view feature learning model；Step 4 introduces particle swarm algorithm, and the objective function obtained to step 3 optimizes；Step 5 repeats step 1 and step 2 for any road segment to obtain input feature vector, input feature vector is brought into the objective function after optimization, realizes and carries out prediction of short-term traffic volume to any road segment.This method realizes the efficient prediction of traffic in short-term, solve the problems, such as that the global prediction ability of temporal-spatial heterogeneity and model is unable to reach equilibrium, it solves the Parametric optimization problem of model, can be widely applied to urban planning, flow of personnel investigation, auto navigation, emergency response, space-time approachability analysis and traffic pollution modeling.
Description
Technical field
The present invention relates to a kind of prediction of short-term traffic volume methods more particularly to a kind of based on multitask multiple view learning model
Prediction of short-term traffic volume method belongs to information technology service field.
Background technique
Not with sensor network, running fix, wireless telecommunications, mobile Internet, high-performance calculation and memory technology
It is disconnected to develop and popularize, there are a series of time series datas with location tags, referred to as space-time data.These space-time datas
In contain abundant and useful information and knowledge is needed by automatic mining, to expedite the emergence of the continuous of spatiotemporal data structure technology
Development.Traffic becomes the test site of space-time modeling technology, many space-time modeling Technology applications as a kind of typical space-time data
The relevant application of traffic.And space-time prediction of short-term traffic volume is as the major class in spatiotemporal data structure family, recent years by
To extensive concern.
However, under urban transportation space-time auto-correlation and the background of temporal-spatial heterogeneity, existing space-time prediction of short-term traffic volume
Model still remains shortcomings: 1) in modeling process, existing method is by each geographical unit (road segment or subregion)
As individual prediction task, and ignore holistic correlation between geographical unit, so that the reasonability of existing method has shortcoming.
2) by the constraint of temporal-spatial heterogeneity, existing space-time modeling method needs individually to model for each geographical unit, so that model exists
After having trained, the set of one group of prediction model is obtained.Also, for unbred geographical unit, need to be fitted ginseng again
Number, this allows for temporal-spatial heterogeneity and the global prediction ability of model is unable to reach equilibrium.3) existing spatio-temporal segmentation is being instructed
During white silk, the method for grid search is generallyd use to determine the hyper parameter of model.The training time of model is usually with parameter
Number is exponentially increased, and is difficult to determine optimal model structure.More than solving the problems, such as, the present invention is based on population calculations
Method integrates the efficient prediction of traffic in short-term of multitask multiple view feature learning model realization.
Summary of the invention
In order to solve shortcoming present in above-mentioned technology, the present invention provides one kind to be learnt based on multitask multiple view
The prediction of short-term traffic volume method of model.
In order to solve the above technical problems, the technical solution adopted by the present invention is that: one kind is learnt based on multitask multiple view
The prediction of short-term traffic volume method of model, includes the following steps:
Step 1 individually constructs Spatio-Temporal Data Model for Spatial for each road segment；
Spatio-Temporal Data Model for Spatial includes Spatial Dimension and time dimension, wherein Spatial Dimension indicates to influence target road segment
Spatial neighbors number, time dimension indicate influence current time historical traffic condition time window length；
Spatial neighbors are obtained using cross-correlation function to portray Spatial Dimension, meanwhile, when stacking each sequentially in time
Between the spatio-temporal state matrix that is spaced, then can be obtained three-dimensional space-time tensor, including space-time adjacent to tensor, when tensor sum null cycle when
Empty trend tensor, then to three space-time tensors of building respectively according to the history number of days of setting, training number of days and test number of days
It is divided, then history space-time tensor, training space-time tensor, test space-time tensor can be obtained for prediction of short-term traffic volume model
Building, to characterize the transportation condition of each road segment all historical junctures, and finally obtain space-time adjacent to view, when null cycle
View, space-time trend view；
Step 2 constructs multitask multiple view learning model；
Using STKNN model as kernel function, respectively to step 1 obtain space-time adjacent to view, space-time period view,
The result of space-time trend view predicted, then the prediction result of each view is carried out to high-rise Semantic mapping, as when
The input feature vector of empty multitask multiple view learning model；
Step 3 constructs objective function using multitask multiple view feature learning model；
The input feature vector obtained according to step 2, while the prediction of each road segment transportation condition being appointed as one
Business, so that each task has consistent characteristic dimension, using multitask multiple view feature learning model come between learning tasks
Correlation and view between consistency, while one group of sharing feature of all road Piece Selections is limited, to realize target
Function can predict the synchronous of entire road network transportation condition；
Step 4 introduces particle swarm algorithm, and the objective function obtained to step 3 optimizes；
Step 5 repeats step 1 and step 2 to obtain input feature vector, by input feature vector band for any road segment
Enter the objective function to after optimization, realizes and prediction of short-term traffic volume is carried out to any road segment.
Further, a kind of cross-correlation function ccf of stepU, l(φ) is expressed as formula I:
Wherein, φ indicates time delay, and time delay E indicates desired value operator,It is road segment u between the time
Every the transportation condition of t,Indicate road segment l time interval t transportation condition,Indicate road segment l when
Between be spaced t+ φ transportation condition, μuAnd μlRespectivelyWithMean value,For road segment u historical traffic condition when
Between sequence,For the historical traffic condition time sequence of road segment l.
Further, objective function Π (W) is expressed as formula VII in step 3:
Wherein,It indicates so that the weight matrix W, y that objective function Π (W) is minimizedlIndicate road segment l
The feature vector of the true value of predicted value, XlwlIndicate the global feature matrix of road segment l and the weight vectors of road segment l
Between product,Indicate feature vector of the space-time adjacent to view of road segment l,Indicate the space-time week of road segment l
The feature vector of phase view,Indicate the feature vector of the space-time trend view of road segment l,WithTable respectively
Show weight of the space-time adjacent to view of task l, the weight of space-time period view, the weight of space-time trend view,It indicates
2- norm, SL, mFor the cross-correlation coefficient between road segment l and road segment m, SL, mValue it is bigger, show two road segments
Transportation condition is more related, and Laplce's regular terms of figure makes wlAnd wmIt is more like, thus can with autocoding spatial coherence,
wlThe weight vectors of expression task l, wmIndicate the weight vectors of expression task m；For enhancing the robustness of model；||W
||2,1Indicate the L of W2,1Norm, so that all tasks automatically select one group of shared determinant attribute, it can be every in W by calculating
The L of a line2The sum of norm obtains；α, β,For coupling parameter, for adjusting the inconsistent intensity of different views, γ is for punishing
The inconsistency of mapping function between different task is penalized, θ is used for the sparsity of controlling feature, μ L2Norm regularization parameter.
Further, particle swarm algorithm mainly comprises the steps that in step 4
1) the number of iterations is set, population scale initializes the position and speed of all particles；
2) training multi task model, calculates the fitness value of each particle；
3) individual is determined according to the appropriateness value of particle using the MAPE error extension of training data as fitness function
Extreme value place and group's extreme value place；
4) according to the speed and position of formula VIII and formula IX more new particle；
Formula VIII and formula IX are as follows:
Wherein,Indicate the speed of i-th of particle in kth time iterative process,It indicates in+1 iterative process of kth
The speed of i-th of particle,Indicate the individual extreme value place of i-th of particle in kth time iterative process,Indicate kth time repeatedly
I-th of particle be in the position of search space during generation,Indicate that i-th of particle is empty in search in+1 iterative process of kth
Between position, GkIndicate the population extreme value place of kth time iteration, ω is inertia weight, c1And c2For acceleration factor, r1And r2For
Random number between [0,1]；
5) judge whether the termination condition for reaching particle swarm algorithm, i.e., whether reach maximum number of iterations, if not up to
Then return step 2), the fitness value of particle, more new individual extreme value and kind are recalculated according to the particle rapidity of update and position
Group's extreme value；If reached, using the position of population extreme value as the optimized parameter set of multi-task learning model.
The invention discloses a kind of prediction of short-term traffic volume methods based on multitask multiple view learning model, can help to solve
Certainly many city management problems, including urban planning, flow of personnel investigation, auto navigation, emergency response, space-time approachability analysis
It is modeled with traffic pollution, there is very extensive answering property.Method of the invention is more reasonable, overcomes existing space-time and hands in short-term
Many defects present in logical prediction model make the global prediction ability of temporal-spatial heterogeneity and model reach balanced, obtain optimal
Model structure.
Detailed description of the invention
Fig. 1 is overall step flow chart of the invention.
Fig. 2 is the flow chart of particle swarm algorithm of the present invention.
Specific embodiment
The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
As shown in Figure 1, presenting overall step flow chart of the invention, specifically are as follows:
Step 1 individually constructs Spatio-Temporal Data Model for Spatial for each road segment；
Spatio-Temporal Data Model for Spatial includes Spatial Dimension and time dimension, wherein Spatial Dimension indicates to influence target road segment
Spatial neighbors number, time dimension indicate influence current time historical traffic condition time window length.Therefore, really
The key for determining Spatio-Temporal Data Model for Spatial is how to choose suitable spatial neighbors and time window length.For each road segment
The transportation condition of any time can be indicated using space-time adjacent to matrix, space-time period matrix, space-time trend matrix simultaneously,
To indicate the space-time dependence and temporal-spatial heterogeneity of transportation condition.
In terms of spatial neighbors selection, it is contemplated that road network has significant heterogeneity, and different road segments have
Different spatial neighbors, therefore, the present invention construct Spatio-Temporal Data Model for Spatial for each road segment.Assuming that there are M for road network
Road segment, TN are total historical time intervals length, and transportation condition time series is represented byWherein,Indicate that dimension is the set of real numbers of M × TN,
Indicate the historical traffic condition time sequence of m-th road segment,Indicate that dimension is the one-dimensional vector of TN.Road segment
Spatial neighbors can be chosen by cross-correlation function, it is assumed thatFor the historical traffic condition time sequence of road segment u, definition
Cross-correlation function ccf of the road segment u and l in time delay φU, l(φ) is shown in formula I:
Wherein, E indicates desired value operator,For road segment u time interval t transportation condition,It indicates
Road segment l time interval t transportation condition,Indicate transportation condition of the road segment l in time interval t+ φ, μu
And μlRespectivelyWithMean value,For the historical traffic condition time sequence of road segment u,For road segment l's
Historical traffic condition time sequence.
By definition, it can be seen that, cross-correlation function can regard the function about time delay as, so that cross-correlation function takes
Being worth maximum time-delay value is the average delay time ψ that section l (surrounding section) influences section u (prediction section)l, ψl
Meet formula II:
Wherein,For as dependent variable ccfU, l(φ) seeks the function of variable φ when maximum.
And if only if ψlIn predicted time range delta t, prediction road segment could be had an impact, mesh can be chosen as
The spatial neighbors of road segment are marked, formal definitions are as shown in formula III:
Wherein,It indicates to meet 0≤ψ of condition for alllThe road segment of≤Δ t,For the space of road segment u
The set of neighbours,Indicate the spatial neighbors number of road segment u.
In terms of time window selection, it is assumed that the collection of the time adjacent partition of time interval t is combined into TC, the collection of period distances
It is combined into TP, the collection at trend interval is combined into TQ, then TC={ t-lc, t- (lc-1) ..., t-1 }, TP={ t-lpp, t- (lp-
1) p ..., t-p }, TQ={ t-lqq, t- (lq-1) q ..., t-q }, wherein lc is time adjacent partition number, that is, is taken
Lc time interval t adjacent time interval, lc=| TC |；Lp be time cycle gap length, lp=| TP |, p be the time week
Phase length, p=1 indicate to take the transportation condition of the previous day time interval t；Lq be time trend gap length, lq=| TQ |, q is
Time trend length, q=1 indicate to take the transportation condition of the last week time interval t.
Spatial neighbors and time window are being obtained, road segment u can be obtained in the space-time of time interval t adjacent to matrixSpace-time period matrixSpace-time trend matrixIt is i.e. each
The transportation condition of road segment any time utilizes three spatio-temporal state matrix synchronization representations, whereinExpression dimension is lc
The real number matrix of × ln,WithIt indicates by a similar method.It can be taken simultaneously in each time interval of guarantee
To history under the premise of time series, history cycle time series, the historical trending time sequence, stack sequentially in time
The spatio-temporal state matrix of each time interval, then can be obtained three-dimensional space-time tensor, including space-time adjacent to tensor, when null cycle
Amount and space-time trend tensor.Finally, we are to three space-time tensors of building respectively according to the history number of days of setting, training number of days
It is divided with test number of days, then history space-time tensor, training space-time tensor, test space-time tensor can be obtained for traffic in short-term
The building of prediction model to characterize the transportation condition of each road segment all historical junctures, and finally obtains the neighbouring view of space-time
Figure, space-time period view, space-time trend view.
Step 2 constructs multitask multiple view learning model；
On the basis of step 1, the present invention respectively models three views as kernel function using STKNN model, then
The prediction result of each view is carried out to high-rise Semantic mapping, the input as space-time multitask multiple view learning model is special
Sign.Assuming that road segment u time interval t space-time adjacent to matrix be MCU, t, by with the space-time shape in history space-time tensor
State matrix compares distance, chooses K apart from nearest spatio-temporal state matrix, to lower a period of time of the spatio-temporal state matrix of each selection
The value at quarter is integrated, and the predicted value of t+1 moment road segment u can be obtained.Therefore, key herein be how to determine away from
From function and the integration function of neighbours.
By taking space-time is adjacent to matrix as an example, since the transportation condition of road segment is indicated using spatio-temporal state matrix,
The present invention constructs weight matrix in time and Spatial Dimension respectively, calculates different moments by weight euclidean distance function
The distance between spatio-temporal state matrix.Its basic thought is: closer with predicted time in time dimension, the weight of distribution is bigger；
In Spatial Dimension, more related to the prediction space of road segment, weight is bigger.Temporal Weight matrix form such as formula IV institute
Show:
dp=| | WTu·MCU, t·WSu-WTu·MCU, p·WSu||2
Wherein, WTuAnd WSuRespectively indicate time and the Spatial weight matrix of road segment u, wtU, iAnd wsU, iIt respectively indicates
I-th of diagonal entry of weight matrix, ti indicate the ti time interval of the neighbouring time series set of time interval t.vsi
Indicate the si spatial neighbors in the set of the spatial neighbors of road segment u.MCU, tIt is road segment u time interval t's
Space-time is adjacent to matrix, MCU, pIndicate p-th of the history of road segment u adjacent to spatio-temporal state matrix, dpFor p-th of history adjacent to when
The distance between dummy status matrix and the spatio-temporal state matrix of t time interval.
Finally, distributing weight, distribution using the transportation condition for the subsequent time that Gauss weighting function is each candidate neighbor
Strategy be it is closer with the spatio-temporal state matrix distance of current time interval, then distribute bigger weight.With space-time adjacent to matrix
For, weight distribution function is defined as formula V:
Wherein,Indicate predicted value of the road segment u at the t+1 moment, subscript C indicates space-time propinquity.For the transportation condition of space-time k-th of candidate neighbor subsequent time of road segment u in tensor, ωU, kFor road
The weight of k-th of candidate neighbor of segment u, dkIt is Gauss weight for the distance between k-th of candidate neighbor and road segment u, a
Parameter.Using with space-time adjacent to tensor it is similar by the way of, when can respectively obtain null cycle tensor, road piece in space-time trend tensor
Prediction result of the section u in time interval t+1WithSubscript p and q respectively indicate time-space periodicity and space-time trend
Property.
Step 3 constructs objective function using multitask multiple view feature learning model；
Space-time is obtained adjacent to view, space-time period view, all training of space-time trend view and testing time by step 2
The prediction result at interval.The prediction result of three views is mapped as high-rise semantic feature by the present invention, utilizes heterogeneous space-time
Semantic feature constructs the input of objective function, while by the prediction of each road segment transportation condition as a task, so that
Each task has consistent characteristic dimension.Using multitask multiple view feature learning model come the correlation between learning tasks
Consistency between view, while one group of sharing feature of all road Piece Selections is limited, to realize entire road network
The synchronous prediction of transportation condition.Specific embodiment is as follows:
Assuming that the space-time of road segment l is adjacent to the feature vector of viewIts
Middle NlFor the sample number of road segment l,Expression dimension is NlReal vector,Corresponding i-th of sample of road segment l
Space-time is adjacent to predicted valueI.e.Similar,Indicate road
The feature vector of the space-time period view of segment l,Indicate the space-time of road segment l
The feature vector of trend view,Respectively indicate the space-time of i-th of sample of road segment l
Period forecasting valueWith trend prediction valueThe feature vector of the true value of road segment l predicted value isThe feature vector of single view is integrated together, the whole of road segment l can be obtained
Body characteristics matrix is
Different views features the transportation condition of road segment from different levels, therefore each view is with different
Weight.Assuming that space-time is adjacent to the weight of viewThe weight of space-time period view isThe weight of space-time trend view isIn order to simplify processing, by the way of Linear Mapping, the weight contribution value of each view is obtained, i.e.,Multiple single-views are integrated together, so that being complementary to one another between view to obtain knowing for enhancing
Know, to obtain the predicted value f of each task integrationl(Xl), i.e., shown in formula VI:
Wherein,The weight vectors of expression task l,WithRespectively indicate task l's
Space-time is adjacent to the weight of view, the weight of space-time period view, the weight of space-time trend view.Indicate the weight matrix of M task, XlwlIndicate the global feature of road segment l
Product between matrix and the weight vectors of road segment l.Since different views can not be got herein to whole prediction result
The priori knowledge of contribution margin, therefore the predicted value by the way of average to be integrated.
In view of (such as propinquity, periodicity, tendency) describes same link piece to different views in terms of different
The internal characteristics of section, the prediction result of each view should be close as far as possible, i.e. consistency between holding view.Therefore,
Enhance the learning ability of single view using regularization term, introducesSo that neighbouring
The prediction result of view and the prediction result of period view are as close possible to improve the precision of prediction model, wherein α is
Coupling parameter, for adjusting the inconsistent intensity of different views, M indicates the number of road segment existing for road network,Indicate 2- norm.On the other hand, it is contemplated that road segment is organically linked together by road network, the friendship of road segment
Gating condition is influenced by its direct or indirect upstream and downstream road segment transportation condition, and the transportation condition between two road segments is got over
Correlation, travel pattern more have similitude, and therefore, we introduce Laplce's regular terms based on figure to grab road segment
The global temporal correlation of l, so that the transportation condition of the road segment with similar travel pattern has lesser deviation, such asIt is shown.In addition, wlIn i-th of element representation, first of task ith feature weight
The property wanted, as soon as we limit the correlation that the common characteristic set of all road Piece Selections group comes between characterization task,
It is to say, all tasks can reach this target by introducing Lasso penalty term based on one group of common character subset.It adopts
With least square loss function, the objective function Π (W) of multitask multiple view feature learning model can form turn to a supervision
Learning framework, as shown in formula VII:
Wherein,It indicates so that the weight matrix W, y that objective function Π (W) is minimizedlIndicate road segment l
The feature vector of the true value of predicted value, XlwlIndicate the global feature matrix of road segment l and the weight vectors of road segment l
Between product,Indicate feature vector of the space-time adjacent to view of road segment l,Indicate the space-time week of road segment l
The feature vector of phase view,Indicate the feature vector of the space-time trend view of road segment l,WithTable respectively
Show weight of the space-time adjacent to view of task l, the weight of space-time period view, the weight of space-time trend view,It indicates
2- norm, SL, mFor the cross-correlation coefficient between road segment l and road segment m, SL, mValue it is bigger, show two road segments
Transportation condition is more related, and Laplce's regular terms of figure makes wlAnd wmIt is more like, thus can with autocoding spatial coherence,
wlThe weight vectors of expression task l, wmIndicate the weight vectors of expression task m；For enhancing the robustness of model；||W
||2,1Indicate the L of W2,1Norm, so that all tasks automatically select one group of shared determinant attribute, it can be every in W by calculating
The L of a line2The sum of norm obtains；α, β,For coupling parameter, for adjusting the inconsistent intensity of different views, γ is for punishing
The inconsistency of mapping function between different task is penalized, θ is used for the sparsity of controlling feature, μ L2Norm regularization parameter.
Since there are non-smooth L2,1Norm, therefore multitask multiple view feature learning model can be regarded as one
Non- smooth convex optimization problem, can be used FISTA, AGD scheduling algorithm to solve the problems, such as this.
Step 4 introduces particle swarm algorithm, optimizes to objective function；
The regularization parameters of space-time multitask multiple view learning model adjusted including multiple needs, such as α, β,γ,
θ, μ, reasonable parameter setting will affect the precision of prediction model to a certain extent, and therefore, present invention introduces particle swarm algorithms
(PSO) objective function of Lai Zidong optimal prediction model obtains the optimized parameter of space-time multitask multiple view feature learning model
Set, while accelerating the training speed of model.
Assuming that there are the molecular population Z=(Z of np grain can have in solution space1, Z2..., Znp), each particle utilizes
Position, speed and fitness value characterize.It is i-th
Position of the son in search space represents a potential optimal solution of extremal optimization problem, respectively corresponds multitask multiple view
The regularization parameter of habit model objective function, β,γ, θ, μ, such as locI, αIndicate that the regularization parameter α of i-th of particle exists
The position of search space, locI, β,locI, γ, locI, θ, locI, μIt is all made of similar representation method.Speed for i-th of particle in search space, vI, αIndicate i-th of particle
Speed of the regularization parameter α in search space, vI, β,vI, γ, vI, θ, vI, μIt is all made of similar representation method.For individual extreme value place, pI, αIndicate the regularization parameter α of i-th of particle
Extreme value place, pI, β,pI, γ, pI, θ, pI, μIt is all made of similar representation method.Population extreme value place isgαIndicate extreme value place of the regularization parameter α in population, gβ,gγ, gθ, gμ?
Using similar representation method.Using the MAPE error extension of training data as fitness function, for judging the excellent of particle
A fitness value is once just recalculated in bad degree, the every update of particle.Each particle is with certain speed in solution room
Movement, by more new individual extreme value Pbest and group extreme value Gbest come the position and speed of more new individual.In iteration each time
Process, the position and speed of particle update respectively as shown in formula VIII and formula IX:
Wherein,Indicate the speed of i-th of particle in kth time iterative process,It indicates in+1 iterative process of kth
The speed of i-th of particle,Indicate the individual extreme value place of i-th of particle in kth time iterative process,Indicate kth time repeatedly
I-th of particle be in the position of search space during generation,Indicate that i-th of particle is empty in search in+1 iterative process of kth
Between position, GkIndicate the population extreme value place of kth time iteration, ω is inertia weight, c1And c2For acceleration factor, r1And r2For
Random number between [0,1].The flow chart of particle swarm algorithm (PSO algorithm) is as shown in Fig. 2, mainly include following 5 steps:
1) the number of iterations is set, population scale initializes the position and speed of all particles；
2) training multi task model, calculates the fitness value of each particle；
3) individual is determined according to the appropriateness value of particle using the MAPE error extension of training data as fitness function
Extreme value place and group's extreme value place；
4) according to the speed and position of formula VIII and formula IX more new particle；
5) judge whether the termination condition for reaching PSO algorithm, i.e., whether reach maximum number of iterations, returned if not up to
Step 2) is returned, the fitness value of particle, more new individual extreme value and population pole are recalculated according to the particle rapidity of update and position
Value；If reached, using the position of population extreme value as the optimized parameter set of multi-task learning model.
Step 5 is optimized objective function by step 4, therefore.It can be applied to carry out any road segment
Prediction of short-term traffic volume repeats step 1 and step 2 to obtain input feature vector, by input feature vector band for any road segment
Enter the objective function to after optimization, realizes and prediction of short-term traffic volume is carried out to any road segment.
The invention discloses a kind of prediction of short-term traffic volume methods based on multitask multiple view learning model, for each road
Road segment indicates the transportation condition of any time using space-time adjacent to matrix, space-time period matrix, space-time trend matrix, from
And indicate the space-time dependence and temporal-spatial heterogeneity of transportation condition；Using one group of kernel function respectively obtain space-time adjacent to view, when
Null cycle view, space-time trend view prediction result, it is further that the prediction result of each view is special as high-rise semanteme
Sign；By the prediction of each road segment as a task, pass through the correlation and view between constraint learning tasks holding task
Consistency between figure allows prediction model to grab the global temporal correlation of road network and enhance it and predicts energy
Power.Meanwhile one group of all road Piece Selections shared feature is limited to realize the synchronous prediction of all road segments；It considers
Include multiple hyper parameters in space-time multitask multiple view learning model, be further introduced into particle swarm algorithm and carry out solving optimization problem,
The optimized parameter set of space-time multi-task learning model is obtained, while accelerating the training speed of model.The present invention is synchronous to consider road
Space-time dependence, temporal-spatial heterogeneity existing for road network, the global temporal correlation between prediction task and multiple views it
Between consistency, the efficient prediction of traffic in short-term is realized using multitask multiple view learning model；It is built by unified space-time
Mold framework solves temporal-spatial heterogeneity and model using the input feature vector of heterogeneous Spatio-Temporal Data Model for Spatial building prediction model
Global prediction ability is unable to reach balanced problem；Particle swarm algorithm is applied to space-time multitask multiple view model, is solved
The Parametric optimization problem of model.By the above strategy, to realize the efficient prediction of traffic in short-term.
The invention has the following advantages:
Space-time is constructed using cross-correlation function adjacent to view, space-time period view, space-time trend view to portray respectively
The transportation condition of each road segment, to portray the space-time dependence and temporal-spatial heterogeneity of transportation condition；Based on Multiple Kernel Learning
Thought, each view respectively corresponds a kernel function, to obtain the prediction result of each view.This group of prediction result is worked as
Make high-level semantic feature mapping to construct the input feature vector matrix of space-time multitask multiple view learning model；Building unification
Space-time multitask multiple view model, by increase in objective function one group of regular terms come between guarantee task correlation and
Consistency between view, so that prediction model has global prediction ability and can grab the global temporal and spatial correlations of road network
Property；Introduce particle swarm algorithm and carry out the parameter selection of optimization object function so that prediction model have optimal model structure and
Accelerate the training speed of model.The present invention can help to solve the problems, such as many city management, including urban planning, flow of personnel tune
It looks into, auto navigation, emergency response, space-time approachability analysis and traffic pollution modeling.
Above embodiment is not limitation of the present invention, and the present invention is also not limited to the example above, this technology neck
The variations, modifications, additions or substitutions that the technical staff in domain is made within the scope of technical solution of the present invention, also belong to this hair
Bright protection scope.
Claims (4)
1. a kind of prediction of short-term traffic volume method based on multitask multiple view learning model, it is characterised in that: the traffic in short-term
Prediction technique includes the following steps:
Step 1 individually constructs Spatio-Temporal Data Model for Spatial for each road segment；
Spatio-Temporal Data Model for Spatial includes Spatial Dimension and time dimension, wherein Spatial Dimension indicates to influence the sky of target road segment
Between neighbours number, time dimension indicate influence current time historical traffic condition time window length；
Spatial neighbors are obtained using cross-correlation function to portray Spatial Dimension, meanwhile, it is stacked between each time sequentially in time
Every spatio-temporal state matrix, then can be obtained three-dimensional space-time tensor, including space-time adjacent to tensor, when null cycle tensor sum space-time become
Then gesture tensor carries out three space-time tensors of building according to the history number of days of setting, training number of days and test number of days respectively
It divides, then the structure that history space-time tensor, training space-time tensor, test space-time tensor are used for prediction of short-term traffic volume model can be obtained
Build, to characterize the transportation condition of each road segment all historical junctures, and finally obtain space-time adjacent to view, when null cycle regard
Figure, space-time trend view；
Step 2 constructs multitask multiple view learning model；
Using STKNN model as kernel function, respectively to the space-time of step 1 acquisition adjacent to view, space-time period view, space-time
The result of trend view is predicted, then the prediction result of each view is carried out to high-rise Semantic mapping, more as space-time
The input feature vector of task multiple view learning model；
Step 3 constructs objective function using multitask multiple view feature learning model；
The input feature vector obtained according to step 2, while the prediction of each road segment transportation condition being made as a task
Obtaining each task has consistent characteristic dimension, using multitask multiple view feature learning model come the correlation between learning tasks
Consistency between property and view, while one group of sharing feature of all road Piece Selections is limited, thus function to achieve the objective energy
Synchronous prediction to entire road network transportation condition；
Step 4 introduces particle swarm algorithm, and the objective function obtained to step 3 optimizes；
Step 5 repeats step 1 and step 2 for any road segment to obtain input feature vector, input feature vector is brought into
Objective function after optimization is realized and carries out prediction of short-term traffic volume to any road segment.
2. the prediction of short-term traffic volume method according to claim 1 based on multitask multiple view learning model, feature exist
In: a kind of cross-correlation function ccf of stepU, l(φ) is expressed as formula I:
Wherein, φ indicates time delay, and time delay E indicates desired value operator,It is road segment u in time interval t
Transportation condition,Indicate road segment l time interval t transportation condition,Indicate road segment l between the time
Every the transportation condition of t+ φ, μuAnd μlRespectivelyWithMean value,For the historical traffic condition time sequence of road segment u
Column,For the historical traffic condition time sequence of road segment l.
3. the prediction of short-term traffic volume method according to claim 1 based on multitask multiple view learning model, feature exist
In: objective function П (W) is expressed as formula VII in step 3:
Wherein,It indicates so that the weight matrix W, y that objective function ∏ (W) is minimizedlIndicate road segment l prediction
The feature vector of the true value of value, XlwlIt indicates between the global feature matrix of road segment l and the weight vectors of road segment l
Product,Indicate feature vector of the space-time adjacent to view of road segment l,Indicate road segment l when null cycle regard
The feature vector of figure,Indicate the feature vector of the space-time trend view of road segment l,WithIt respectively indicates and appoints
Be engaged in l space-time adjacent to the weight of view, the weight of space-time period view, the weight of space-time trend view,Indicate 2- model
Number, SL, mFor the cross-correlation coefficient between road segment l and road segment m, SL, mValue it is bigger, show two road segment traffic
Condition is more related, and Laplce's regular terms of figure makes wlAnd wmIt is more like, therefore can be with autocoding spatial coherence, wlTable
Show the weight vectors of task l, wmIndicate the weight vectors of expression task m；For enhancing the robustness of model；||W|
|2,1Indicate the L of W2,1Norm, so that all tasks automatically select one group of shared determinant attribute, it can be every in W by calculating
The L of a line2The sum of norm obtains；α, β,For coupling parameter, for adjusting the inconsistent intensity of different views, γ is for punishing
The inconsistency of mapping function between different task is penalized, θ is used for the sparsity of controlling feature, μ L2Norm regularization parameter.
4. the prediction of short-term traffic volume method according to claim 1 based on multitask multiple view learning model, feature exist
In: particle swarm algorithm mainly comprises the steps that in step 4
1) the number of iterations is set, population scale initializes the position and speed of all particles；
2) training multi task model, calculates the fitness value of each particle；
3) individual extreme value is determined according to the appropriateness value of particle using the MAPE error extension of training data as fitness function
Position and group's extreme value place；
4) according to the speed and position of formula VIII and the more new particle of formula Ⅸ；
The formula VIII and formula Ⅸ are as follows:
Wherein,Indicate the speed of i-th of particle in kth time iterative process,It indicates i-th in+1 iterative process of kth
The speed of a particle,Indicate the individual extreme value place of i-th of particle in kth time iterative process,Indicate kth time iteration mistake
I-th of particle be in the position of search space in journey,Indicate that i-th of particle is in search space in+1 iterative process of kth
Position, GkIndicate the population extreme value place of kth time iteration, ω is inertia weight, c1And c2For acceleration factor, r1And r2For [0,
1] random number between；
5) judge whether the termination condition for reaching particle swarm algorithm, i.e., whether reach maximum number of iterations, returned if not up to
Step 2) is returned, the fitness value of particle, more new individual extreme value and population pole are recalculated according to the particle rapidity of update and position
Value；If reached, using the position of population extreme value as the optimized parameter set of multi-task learning model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811578615.7A CN109598939B (en) | 2018-12-24 | 2018-12-24 | A kind of prediction of short-term traffic volume method based on multitask multiple view learning model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811578615.7A CN109598939B (en) | 2018-12-24 | 2018-12-24 | A kind of prediction of short-term traffic volume method based on multitask multiple view learning model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109598939A CN109598939A (en) | 2019-04-09 |
| CN109598939B true CN109598939B (en) | 2019-08-27 |
Family
ID=65964378
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811578615.7A Active CN109598939B (en) | 2018-12-24 | 2018-12-24 | A kind of prediction of short-term traffic volume method based on multitask multiple view learning model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109598939B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110633871B (en) * | 2019-09-25 | 2022-09-16 | 大连理工大学 | Regional traffic demand prediction method based on convolution long-term and short-term memory network |
| CN111311904B (en) * | 2020-01-17 | 2021-06-11 | 东南大学 | Traffic state estimation method based on floating car data weighted tensor reconstruction |
| CN111862590A (en) * | 2020-05-13 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | Road condition prediction method, road condition prediction device and storage medium |
| CN112215408A (en) * | 2020-09-24 | 2021-01-12 | 交控科技股份有限公司 | Rail transit passenger flow volume prediction method and device |
| CN112562339B (en) * | 2020-12-09 | 2022-01-04 | 福州大学 | Robust traffic flow prediction method based on multitask graph convolutional network |
| CN112599246B (en) * | 2021-03-03 | 2021-06-29 | 四川华迪信息技术有限公司 | Vital sign data processing method, system, device and computer readable medium |
| CN116050640B (en) * | 2023-02-01 | 2023-10-13 | 北京交通大学 | Short-time passenger flow prediction method of multi-mode traffic system based on self-adaptive multi-graph convolution |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104933862A (en) * | 2015-05-26 | 2015-09-23 | 大连理工大学 | Urban traffic jam intelligent combination prediction method based on track of floating vehicle |
| CN106384521A (en) * | 2016-09-18 | 2017-02-08 | 广东工业大学 | Single-intersection traffic signal optimization control method based on public transport priority |
| CN106971548A (en) * | 2017-05-18 | 2017-07-21 | 福州大学 | The Short-time Traffic Flow Forecasting Methods of optimizable multi-core adaptive SVMs |
| EP3035314B1 (en) * | 2014-12-18 | 2017-11-15 | Be-Mobile Tech NV | A traffic data fusion system and the related method for providing a traffic state for a network of roads |
| CN108986470A (en) * | 2018-08-20 | 2018-12-11 | 华南理工大学 | The Travel Time Estimation Method of particle swarm algorithm optimization LSTM neural network |
-
2018
- 2018-12-24 CN CN201811578615.7A patent/CN109598939B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3035314B1 (en) * | 2014-12-18 | 2017-11-15 | Be-Mobile Tech NV | A traffic data fusion system and the related method for providing a traffic state for a network of roads |
| CN104933862A (en) * | 2015-05-26 | 2015-09-23 | 大连理工大学 | Urban traffic jam intelligent combination prediction method based on track of floating vehicle |
| CN106384521A (en) * | 2016-09-18 | 2017-02-08 | 广东工业大学 | Single-intersection traffic signal optimization control method based on public transport priority |
| CN106971548A (en) * | 2017-05-18 | 2017-07-21 | 福州大学 | The Short-time Traffic Flow Forecasting Methods of optimizable multi-core adaptive SVMs |
| CN108986470A (en) * | 2018-08-20 | 2018-12-11 | 华南理工大学 | The Travel Time Estimation Method of particle swarm algorithm optimization LSTM neural network |
Non-Patent Citations (2)
| Title |
|---|
| Quantifying uncertainty in short-term traffic prediction and its;Lei Lina, John C. Handleyb, Yiming Guc, Lei Zhud, Xuejin Wene, A;《Transportation Research》;20180519;全文 * |
| 基于机器学习的交通状态判别与预测方法研究;商强;《CNKI中国博士学位论文全文数据库》;20171115;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109598939A (en) | 2019-04-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109598939B (en) | A kind of prediction of short-term traffic volume method based on multitask multiple view learning model | |
| Dong et al. | Nats-bench: Benchmarking nas algorithms for architecture topology and size | |
| Feng et al. | Deeptp: An end-to-end neural network for mobile cellular traffic prediction | |
| Johnson et al. | Optimization by simulated annealing: an experimental evaluation; part II, graph coloring and number partitioning | |
| CN111860951A (en) | Rail transit passenger flow prediction method based on dynamic hypergraph convolutional network | |
| CN106952181A (en) | Electric Load Prediction System based on long Memory Neural Networks in short-term | |
| CN109446923A (en) | Depth based on training characteristics fusion supervises convolutional neural networks Activity recognition method | |
| CN109902880A (en) | A kind of city stream of people's prediction technique generating confrontation network based on Seq2Seq | |
| CN112382082A (en) | Method and system for predicting traffic running state in congested area | |
| CN108804220A (en) | A method of the satellite task planning algorithm research based on parallel computation | |
| CN1987906A (en) | Method for dynamicaly predicting land use change | |
| CN108108808A (en) | A kind of position predicting method and device based on depth belief network | |
| CN106910199A (en) | Towards the car networking mass-rent method of city space information gathering | |
| CN104217258A (en) | Method for power load condition density prediction | |
| CN113762595B (en) | Traffic time prediction model training method, traffic time prediction method and equipment | |
| CN110310479A (en) | A kind of Forecast of Urban Traffic Flow forecasting system and method | |
| CN108537691A (en) | A kind of region visit intelligent management system and method | |
| CN110515732A (en) | A kind of method for allocating tasks based on resource-constrained robot deep learning reasoning | |
| CN109063759A (en) | A kind of neural network structure searching method applied to the more attribute forecasts of picture | |
| CN110619419A (en) | Passenger flow prediction method for urban rail transit | |
| CN110517494A (en) | Forecasting traffic flow model, prediction technique, system, device based on integrated study | |
| CN112419711A (en) | Closed parking lot parking demand prediction method based on improved GMDH algorithm | |
| CN103605493B (en) | Sorting in parallel learning method based on Graphics Processing Unit and system | |
| CN116108383A (en) | Ship track prediction method based on improved goblet sea squirt multi-output support vector | |
| CN112036598A (en) | Charging pile use information prediction method based on multi-information coupling |
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 |