CN105185106A - Road traffic flow parameter prediction method based on granular computing - Google Patents

Abstract

The present invention relates to the field of traffic information release and traffic management and control and discloses a road traffic flow parameter prediction method based on granular computing. The method comprises the steps of (1) replacing a data point by information particulate to be the basic unit of data mining analysis, (2) with granular computing ideology throughout the whole prediction framework, taking granular processing as a data processing method with a unified structure, allowing a policy maker to clearly understand the positions of various forms of systems in mutual interaction, grasping the communication mode of the systems, and establishing an enhanced harmonious environment among different ways, (3) with a fuzzy time series and the Gath-Geva cluster theory as the basis, by focusing on the commonalities of existing formal methods, recognizing the orthogonality of an existing good frame ( such as the probability theory and the probability density functions of various variables), with variable granularity concept as a basis, establishing the interval length analysis model of a granularity range according to a numerical entity, and thus realizing pattern recognition and speculation on the above basis.

Description

A kind of road traffic flow parameter prediction method based on Granule Computing

Technical field

The present invention relates to and relate to Traffic information demonstration and traffic administration and control field, particularly relate to a kind of road traffic flow parameter prediction method based on Granule Computing.

Background technology

Traffic flow parameter prediction is the important evidence of Traffic flow guidance and Traffic information demonstration.Under urban transportation background, traffic flow is based on time domain dynamic change, thus the prediction of traffic flow parameter is a problem based on time domain dynamic process.Traditional Forecasting Methodology, in precision and dynamic process, is difficult to search out equilibrium point, causes it to predict the outcome and departs from actual traffic supplemental characteristic variation tendency.

SongandChissom (1993) proposes the concept of Fuzzy time sequence first, compares with traditional fuzzy collection, and it has good variation characteristic in dynamic time territory.In recent years, Fuzzy time sequence has successfully solved the decision problem in multiple field, based on the concept of Fuzzy time sequence, scholars propose Fuzzy Time Series Model based on Fixed Time Interval and the Fuzzy Time Series Model in the dynamic change time interval in succession.The applied research of Fuzzy Time Series Model has achieved very large achievement, but still there is certain deficiency.In the past in research, scholars seldom consider the impact (or the time itself not being considered as variable) of time variable itself, and research is often confined to the quantity of information of data itself, and the related information of little mining data inherence, as the data regularity of distribution.Scholars have used support vector machine, to have delimited in research range gap length to seek higher precision of prediction based on the rough set of entropy, granular computing scheduling algorithm, but the prediction effect only obtained in research range, once research range is with after actual demand or time dynamic, estimation effect obviously declines, and applicability is short of to some extent.

Predict for traffic flow parameter, CN201210102006.0 discloses a kind of traffic flow forecasting method based on Fuzzy Kalman Filter, it is characterized in that comprising the following steps: lay detecting device in each track in each track, upstream, section and downstream, section and gather arithmetic for real-time traffic flow supplemental characteristic; Obtain each lane detector historical traffic same period stream supplemental characteristic; Kalman Filter Technology is adopted to build Dynamic Kalman Filtering traffic flow parameter forecast model; By Kalman filtering time update equation and state updating equation, obtain Kalman filtering parameter prediction result; History average traffic same period stream parameter is introduced in formula, build Fuzzy Kalman Filter traffic flow parameter forecast model; The arithmetic for real-time traffic flow parameter collected according to detecting device and history average traffic same period stream parameter, predict following time interval and traffic flow parameter afterwards.The disclosed Forecasting Methodology based on Fuzzy Kalman Filter of CN201210102006.0 is to utilize kalman filter method, the performance prediction of parameter is completed in conjunction with obscure idea, its precision of prediction achieves certain raising compared with classic method, but precision is non-very accurate, namely predict the outcome and there is the stochastic error of undulatory property with True Data, and error is unable to estimate.Can be clearer and more definite, although the value degree of accuracy of Forecasting Methodology prediction is high disclosed in CN201210102006.0, but necessarily there is deviation, is a kind of mistake of accurate type.

Summary of the invention

For not high and lack the technical matters of adaptable performance prediction for road traffic flow parameter prediction fiduciary level in prior art, the invention discloses a kind of road traffic flow parameter prediction method based on Granule Computing.

Object of the present invention is realized by following technical proposals:

Based on a road traffic flow parameter prediction method for Granule Computing, it is characterized in that specifically comprising the following steps:

Based on a road traffic flow parameter prediction method for Granule Computing, the numerical range span of the traffic flow parameter that it specifically comprises the following steps: step one, basis detects, definition research range U=[U _l, U _u] and comformed information numbers of particles h, wherein U _lrepresent the round values more arbitrarily small than numerical value minimum value in overall data, U _urepresent the round values larger arbitrarily than numerical value maximal value in overall data; Step 2, in research range, delimit fuzzy set, and determine the membership between the traffic flow parameter data that detect and fuzzy set; Wherein fuzzy set number is identical with messenger particle number; Step 3, the logical relation determining between fuzzy set, obtain fuzzy relation group; Step 4, trend according to fuzzy relation group, adopt Fuzzy time sequence to carry out messenger particle interval estimation, thus dope the traffic flow parameter of subsequent time period.

Further, the process of above-mentioned messenger particle interval estimation is specially: logic of propositions relation, A _j→ A _j1, A _j2..., A _jpif logical relation group trend is ascendant trend, then the lower limit of interval estimation is m _j, the upper limit of interval estimation is (m _j1+ m _j2..., m _jp)/p; If logical relation group trend is downtrending, then the lower limit of interval estimation is (m _j1+ m _j2..., m _jp)/p, the upper limit of interval estimation is m _j; If logical relation group trend is not only on the rise but also have downtrending, then the lower limit of interval estimation is (m _j1+ m _j2..., m _jk)/k, the upper limit of interval estimation is (m _jk+1+ m _jk+2..., m _jp)/(p-k), wherein, j is the subscript of logical relation front end fuzzy set, A _jrepresent a jth fuzzy set, A _j1, A _j2..., A _jpthe rear end fuzzy set that presentation logic relation front end fuzzy set is corresponding, p altogether, k is the intermediate point between 1 ~ p, m _jfor A _jthe u that fuzzy set is corresponding _jmedian point.

Further, above-mentioned messenger particle is by data set D={x _k| k=1 ..., n} is formed, and particle characteristic comprises interval range length and contains data point number two features, and messenger particle is expressed as Ω=[a, b], and wherein a and b is data set D={x _k| k=1 ..., the boundary of n}, described boundary refers to the boundary up and down of particle, namely particle comprise the boundary up and down of data set.

Further, for Fuzzy time sequence F (the t-1)=A of different time points _iwith F (t)=A _j, fuzzy logical relationship regards the logical relation between F (t-1) and F (t) as, is designated as A _i→ A _j, A _ifor relation front end, A _jfor relation rear end; For same fuzzy set A, when several fuzzy logical relationship front end is identical, merge into fuzzy logical relationship group.

Further, as time series χ={ x _k| k=1 ..., n}, altogether n sample, corresponding time coordinate is θ={ t _k| k=1 ..., n}, by degree of membership u _i,kwith cluster centre η _ithe fuzzy clustering on time data collection that the minimization function formed completes, minimization function is configured to: $J_{GGt}=\underset{k=1}{\overset{n}{\Σ}}\underset{i=1}{\overset{c}{\Σ}}{u}_{i,k}^m\left|\right|z_k-{\mathrm{\η}}_i|{|}^2=\underset{k=1}{\overset{n}{\Σ}}\underset{i=1}{\overset{c}{\Σ}}{u}_{i,k}^m{D}^2(z_k,{\mathrm{\η}}_i);$ Wherein z _k=[t _k, x _k ^t] ^tit is the data point comprising time coordinate; M is weighted index, m > 1; C is cluster species number, c > 2; D ²(z _k, η _i) be data point z _kwith cluster centre η _ibetween distance; u _i,krepresent that in n sample, a kth data point is under the jurisdiction of the degree of membership of the i-th class.

Further, above-mentioned U=[U _l, U _u] be research range, this research range is divided into the interval of h unequal length, i.e. h messenger particle, determines that the method for this h burst length is as follows: Step1. determines species number c, calculates corresponding degree of membership; Species number c=[h/2], represent the maximum integer being no more than h/2 value, h is interval number, and calculates cluster centre η ₁, η ₂..., η _cand the degree of membership u of correspondence _i,k(i=1 ..., c; K=1 ..., n); Step2. according to degree of membership construction data subset; For cluster centre η ₁, η ₂..., η _cand corresponding degree of membership u _i,k(i=1 ..., c; K=1 ..., n), construction data subset is as follows: $D_2=\{x_k\∈D|u_{2,k}=\underset{1\≤i\≤c}{max\left\{u_{i,k}\right\}}\},...,D_c=\{x_k\∈D|u_{c,k}=\underset{1\≤i\≤c}{max\left\{u_{i,k}\right\}}\};$ Step3. messenger particle is built; data subset D _icluster centre, be utilize the upper and lower boundary of optimum of messenger particle building method computing information particle, messenger particle is Ω _i=[a _i, b _i]; Step4. the interval u of corresponding research range is determined ₁, u ₂..., u _h.

Further, said method also comprises and carries out pre-service to the traffic flow parameter detected.

Further, above-mentioned pre-service is specially and adopts the average of closing on two time points to carry out supplementing and reparation of data as interpolation.

By adopting above technical scheme, the present invention has following beneficial effect:

In urban transportation background of today, road traffic stream mode is always in real-time dynamic change.The prediction of traffic flow parameter, can not depart from the impact of traffic flow modes.Traditional Forecasting Methodology, in precision and dynamic process, is difficult to search out equilibrium point, causes it to predict the outcome and departs from actual traffic supplemental characteristic variation tendency.Present forecast demand is the process of dynamic process under time domain, the present invention is in order to adapt to this feature, construct the Fuzzy Time Series Model based on messenger particle: with Granule Computing thought, using messenger particle substitution number strong point as the elementary cell of data mining analysis, by building the messenger particle with varying granularity, the dynamical forecasting problem of research traffic flow parameter.In actual traffic environment, the frequency of the traffic flow parameter dynamic change Accurate Prediction based on numerical value the is become difficult point of sphere of learning.But traffic decision-making person is often interval range to the demand of parameter prediction, judges traffic flow modes with this.Forecast model of the present invention, has just sought appropriate prediction scheme in actual demand and dynamic process, ensure that prediction accuracy, and the supply of the decision information of the person that serves communications policy better.

Compared with Forecasting Methodology disclosed in CN201210102006.0, Forecasting Methodology of the present invention is based on Granule Computing thought, in conjunction with Fuzzy time sequence and relevant cluster theory, complete the motion interval prediction of traffic flow parameter, the result namely predicted is a kind of interval range form.Although this patent does not provide a kind of accurate numerical value, by case verification, this patent can ensure that actual value one fixes within the scope of this, and namely this patent provides the accurate of a kind of fuzzy type.Compared with CN201210102006.0, difference is in essence that the mistake of accurate type provides numerical value, but but necessarily there is error, and performance prediction process medial error is random fluctuation.And fuzzy type is accurately replace numerical value with interval range, numerical value is arbitrary value of interval range, but ensure that numerical value can exceed this interval range scarcely, compared with the mistake of accurate type, be equivalent to the estimation of numerical prediction and fluctuation (error) to complete simultaneously.In actual applications, traffic decision-making person often according to the analytical standard of Assessment of Service Level for Urban Roads and contrasting of traffic flow parameter, thus judges traffic flow modes.In Assessment of Serviceability of Roads analytical standard, which provide the affiliated scope of traffic flow parameter under various traffic behavior, namely the demand of traffic decision-making person to parameter prediction is interval range.Therefore, in the Forecasting Methodology of two kinds of patents, the interval range prediction that this patent proposes more is close to actual demand, and practicality is higher.

Accompanying drawing explanation

Fig. 1 is take data point as the structural representation of data analysis unit.

Fig. 2 is take information as the structural representation of data analysis unit.

Fig. 3 is the structure process flow diagram of traffic flow parameter forecast model of the present invention.

Fig. 4 is trend analysis figure in traffic flow parameter of the present invention prediction example.

Embodiment

Below in conjunction with Figure of description, describe the specific embodiment of the present invention in detail.

The invention discloses a kind of road traffic flow parameter prediction method based on Granule Computing, its feature specifically comprise the following steps: step one, according to the numerical range span of traffic flow parameter detected, definition research range comformed information numbers of particles; Step 2, in research range, delimit fuzzy set, and determine the membership between the traffic flow parameter data that detect and fuzzy set; Wherein fuzzy set number is identical with messenger particle number; Step 3, the logical relation determining between fuzzy set, obtain fuzzy relation group; Step 4, employing Fuzzy time sequence carry out messenger particle interval estimation, thus dope the traffic flow parameter of subsequent time period.Fuzzy time sequence is attached to traffic flow parameter prediction field by the present invention, achieve the Accurate Prediction of traffic flow parameter, dope the interval range of the traffic flow parameter of next time period, it is 100% accurate that it predicts the outcome, compared to recursive prediction of the prior art, the modes such as neural network prediction, its accuracy rate brings up to 100%, simultaneously according to the feature of traffic parameter, such as speed, its needs obtain a value range can carry out judging whether to block up, do not need accurate data dot values, the point value that one exists error is obtained obviously with it, the present invention can obtain one 100% accurately value range be one and better select.The present invention is using messenger particle substitution number strong point as the elementary cell of data mining analysis, overall prediction framework is run through with Granule Computing thought, with Fuzzy time sequence and Gath-Geva Clustering Theory for relying on, by focusing on the general character of existing official method, be familiar with the orthogonality of existing good framework, based on varying granularity concept, set up the burst length analytical model building particle size range according to numerical value entity, implementation pattern identification and supposition accordingly.

Above-mentioned messenger particle, as data mining analysis elementary cell of the present invention, is different from conventional data analysis method, the introducing of messenger particle, is convenient to the general character identifying data essence from extremely different problem aspects.Messenger particle is by data set D={x _k| k=1 ..., n} is formed, and particle characteristic to comprise between granulomere extent length and contains data point number two features.Messenger particle is expressed as Ω=[a, b], and wherein a and b can regard data set D={x as _k| k=1 ..., the boundary of n}.Wherein L (Ω) burst length that is particle, the function of computational length can be set to F ₁, then the feature representation of messenger particle can be defined as F ₁(L (Ω)).What particle also needed to describe is data amount check contained by particle, is designated as Card{x _k| x _k∈ Ω }, can function F be used ₂(Card{x _k| x _k∈ Ω }) represent.Then the computing method of boundary can be expressed as follows:

V(b)＝F ₁(|b-med(D)|)·F ₂(Card{x _k∈D|med(D)≤x _k≤b})(1)

V(a)＝F ₁(|a-med(D)|)·F ₂(Card{x _k∈D|a≤x _k≤med(D)})(2)

Wherein med (D) is the intermediate value of data set D.On this basis, optimum messenger particle bound computing method are as follows:

$V\left(b_{opt}\right)=\underset{b\≥med\left(D\right)}{max}V\left(b\right),V\left(a_{opt}\right)=\underset{a\≤med\left(D\right)}{min}V\left(a\right)---\left(3\right)$

Find out a meeting above formula _optand b _optas the bound of messenger particle.About length and data amount check function, can be calculated as follows:

F ₁(u)＝exp(-αu),F ₂(u)＝u(4)

Further, the above-mentioned structure to messenger particle (statement of the feature that formula (1) ~ (3) and front and back related text all should comprise at descriptor particle and feature and calculating, wherein calculating section relates to theory) (formula (6) ~ (10) are exactly Clustering Theory thought to relate to Fuzzy time sequence and Gath-Geva Clustering Theory, describe in detail how on seasonal effect in time series basis, apply this Clustering Theory, complete cluster, i.e. fuzzy partition.The wherein key concept of degree of membership inherently fuzzy clustering, probability density calculates and adopts Gaussian function analysis to be adjustment.The present invention, in order to be applicable to the structure of messenger particle, carries out adjustment to two kinds of theories as follows:

Fuzzy time sequence, be the effect of this factor of joining day on the basis of fuzzy set, redefine fuzzy set, construction method is as follows:

U={u is used in definition 1. ₁, u ₂..., u _nthe research range of problem of representation, then the fuzzy set under scope U is defined as follows:

$A=\frac{f_A\left(u_1\right)}{u_1}+\frac{f_A\left(u_2\right)}{u_2}+,...,+\frac{f_A\left(u_n\right)}{u_n}---\left(5\right)$

Wherein f _athe subordinate function being under the jurisdiction of fuzzy set A, f _arepresent u _ibe under the jurisdiction of the degree of fuzzy set A, 1≤i≤n, i.e. f _a→ [0,1].

Definition 2. hypothesis time variable set Y (t) (t=..., 0,1,2 ...), by corresponding data time point, be applied to each time subset f of fuzzy set A _i(t) (i=1,2 ...) in, F (t) is considered as the f in corresponding moment _i(t) (i=1,2 ...) and set, then F (t) be called time variable Y (t) (t=..., 0,1,2 ...) and on Fuzzy time sequence.

Definition 3. is for Fuzzy time sequence F (the t-1)=A of different time points _iwith F (t)=A _j, fuzzy logical relationship can be regarded as the logical relation between F (t-1) and F (t), is denoted as A _i→ A _j, A _ifor relation front end, A _jfor relation rear end.

Definition 4. is directed to same fuzzy set A, when several fuzzy logical relationship front end is identical, can merge into fuzzy logical relationship group, reduces computational analysis amount.Namely there is relation A _i→ A _j1, A _i→ A _j2..., fuzzy logical relationship group A can be formed _i→ A _j1, A _j2....

Use Gath-Geva fuzzy clustering to carry out fuzzy partition to time series, temporal information is added wherein, obtain a fuzzy partition with time domain information.Assuming that a time series χ={ x _k| k=1 ..., n}, altogether n sample, corresponding time coordinate is θ={ t _k| k=1 ..., n}, this Clustering Theory essence is by degree of membership u _i,kwith cluster centre η _ithe fuzzy clustering on time data collection that the minimization function formed completes, minimization function is constructed as follows:

$J_{GGt}=\underset{k=1}{\overset{n}{\Σ}}\underset{i=1}{\overset{c}{\Σ}}{u}_{i,k}^m\left|\right|z_k-{\mathrm{\η}}_i|{|}^2=\underset{k=1}{\overset{n}{\Σ}}\underset{i=1}{\overset{c}{\Σ}}{u}_{i,k}^m{D}^2(z_k,{\mathrm{\η}}_i)---\left(6\right)$

Get local minimum, wherein z _k=[t _k, x _k ^t] ^tit is the data point comprising time coordinate; M is weighted index, general m > 1; C is cluster species number, general c > 2; D ²(z _k, η _i) be data point z _kwith cluster centre η _ibetween distance.The constraint condition of minimization function is as follows:

$0\&le;u_{i,k}\&le;1\&DoubleRightArrow;\&ForAll;i,k,0<\underset{k=1}{\overset{n}{\&Sigma;}}{u}_{i,k}\&le;n\&DoubleRightArrow;\&ForAll;i,\underset{i=1}{\overset{c}{\&Sigma;}}{u}_{i,k}=1\&DoubleRightArrow;\&ForAll;k---\left(7\right)$

Because t _kwith x _kseparate, distance D ²(z _k, η _i) can express with following formula:

$D^2(z_k,{\mathrm{\&eta;}}_i)=\frac{1}{{\mathrm{\&alpha;}}_{i}p\left(t_k\right|{\mathrm{\&eta;}}_i)}=\frac{1}{{\mathrm{\&alpha;}}_{i}p\left(t_k\right|{\mathrm{\&eta;}}_i^t)p\left(x_k\right|{\mathrm{\&eta;}}_i^x)}---\left(8\right)$

For the independence of time variable and numerical variable, the probability density in Gath-Geva Fuzzy Clustering Theory is adjusted, be about to distance D ²(z _k, η _i) in the probability density that relates to adopt and calculate with the following method:

Probability density function represent t _kbelong to the probability of the i-th class, Gaussian function can be utilized to calculate:

$p\left(t_k\right|{\mathrm{\&eta;}}_i^t)=G(t_k;v_i^t,{\mathrm{\&sigma;}}_{i,t}^2)=-\frac{1}{\sqrt{2{\mathrm{\&pi;\&sigma;}}_{i,t}^2}}\exp (-\frac{1}{2}\frac{{(t_k-v_i^t)}^2}{{\mathrm{\&sigma;}}_{i,t}^2})---\left(9\right)$

Wherein with represent average and variance respectively.And probability density function represent x _kbelong to the probability of the i-th class, Gaussian function can be utilized to calculate:

$p\left(x_k\right|{\mathrm{\&eta;}}_i^x)=G(x_k;v_i^x,F_i^x)=\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{\frac{r}{2}}\sqrt{\det F_i^x}}\exp (-\frac{1}{2}{(x_k-v_i^x)}^T\frac{1}{F_i^x}(x_k-v_i^x))---\left(10\right)$

Wherein with represent average and covariance respectively, r is covariance matrix line number.α _ibe coefficient, meet: $\underset{i=1}{\overset{c}{\&Sigma;}}{\mathrm{\&alpha;}}_i=1,{\mathrm{\&alpha;}}_i\&GreaterEqual;0,i=1,...,c.$

In sum, when the division hop count of given research range, this theory can complete cluster and corresponding fuzzy partition based on time series.

Further, based on above-mentioned two kinds of theories, length (burst length and data amount check) the constant current journey really of messenger particle is as follows:

The delimitation of messenger particle burst length can affect the precision finally predicted the outcome, and the present invention adopts the division methods of a kind of unequal intervals length, to improve the accuracy predicted the outcome.

First research range needs the interval number supposing to divide when dividing, i.e. messenger particle number.Thus provide c class cluster centre, then according to the degree of membership of correspondence, build c class subset, and calculate maximum messenger particle subset, the burst length of last comformed information particle.

Concrete division methods is as follows: suppose sequence sets D={x preset time _k| k=1 ..., n}, definition D _min=min{x _i| x _i∈ D}, D _max=max{x _i| x _i∈ D}.Suppose U=[U _l, U _u] be research range, this research range is divided into the interval (messenger particle number) of h unequal length, determines that the method for this h burst length is as follows:

Step1. determine species number c, calculate corresponding degree of membership.Species number c=[h/2], represent the maximum integer being no more than h/2 value, h is interval number (messenger particle number) and calculates cluster centre η ₁, η ₂..., η _cand the degree of membership u of correspondence _i,k(i=1 ..., c; K=1 ..., n);

Step2. according to degree of membership construction data subset.For cluster centre η ₁, η ₂..., η _cand corresponding degree of membership u _i,k(i=1 ..., c; K=1 ..., n), construction data subset is as follows: $D_2=\{x_k\&Element;D|u_{2,k}=\underset{1\&le;i\&le;c}{max\left\{u_{i,k}\right\}}\},...,D_c=\{x_k\&Element;D|u_{c,k}=\underset{1\&le;i\&le;c}{max\left\{u_{i,k}\right\}}\};$

Step3. messenger particle is built. data subset D _icluster centre, be utilize the upper and lower boundary of optimum of messenger particle building method computing information particle, messenger particle is Ω _i=[a _i, b _i];

Step4. the interval u of corresponding research range is determined ₁, u ₂..., u _h.

In sum, the determination of burst length can be reduced to following two kinds of situations:

(I) when h is odd number, if then u ₁=[U _l, (U _l+ med (D ₁))/2], u ₂=[(U _l+ med (D ₁))/2, med (D ₁)], u _2i-1=[med (D _i-1), (b _i-1+ a _i)/2], u _2i=[(b _i-1+ a _i)/2, med (D _i)], u _h=[med (D _(h-1)/2), U ₀] (i=2 ..., (h-1)/2), otherwise u ₁=[U _l, med (D ₁)], u _2i=[med (D _i), (b _i+ a _i+1)/2], u _2i+1=[(b _i+ a _i+1)/2, med (D _i+1)], u _h-1=[med (D _(h-1)/2), (U _u+ b _(h-1)/2)/2], u _h=[(U _u+ b _(h-1)/2)/2, U _u] (i=2 ..., (h-3)/2).

(II) when h is even number, u ₁=[U _l, med (D ₁)], u _2i=[med (D _i), (b _i+ a _i+1)/2], u _2i+1=[(b _i+ a _i+1)/2, med (D _i+1)], u _h=[med (D _h/2), U _u] (i=1 ..., (h-2)/2).

Further, based on length configuration model between above-mentioned prediction thought, basic theory, granulomere, the traffic flow parameter forecast model based on Granule Computing builds and is made up of following four steps:

Step1. research range is defined and comformed information particle burst length

Research range is an information matrix, and the traffic flow parameter detected by each moment traffic detector between A point to B point is as the element of in matrix and complete the division of messenger particle burst length:

$U=\left(\begin{array}{cccc}{u}_{11}& u_{12}& ...& u_{1m}\\ u_{21}& u_{22}& ...& u_{2m}\\ ...& ...& ...& ...\\ u_{n1}& u_{n2}& ...& u_{nm}\end{array}\right)---\left(11\right)$

Step2. fuzzy set and Fuzzy time sequence is built

According to the concept of Fuzzy time sequence, on the basis of Given information particle burst length, fuzzy set can be set up as follows:

A _i＝1/u _i+0.5/u _i-1+0.5/u _i+1，(i＝1,...,h)(12)

Wherein u ₀and u _h+1get infinitely-great numerical value.

Step3. the logical relation between fuzzy set is determined

Logical relation between fuzzy set bad direct description, but time domain corresponding relation during owing to building between fuzzy set and messenger particle, can be described by the degree of consistency between fuzzy set and messenger particle, be denoted as Poss (Ω, A _i), computing method are as follows:

Poss(Ω,A _i)＝sup _x∈Ω[Ω(x)tA _i(x)](13)

Wherein sup function representation capping, is equivalent to ask degree of association maximization problems.By the statement of the method, time series can change messenger particle time series into.

In above-mentioned analysis, there are a kind of special circumstances, suppose Ω ₇the A that (expression is divided into 7 intervals) is corresponding _iconsistent degree is respectively 0,0,0,0,0.5,1,1, and maximum is A ₆and A ₇, then Ω ₇logical relation map should be 0.5A ₆+ 0.5A ₇.After determining logical relation, logic need be carried out according to the mapping of each messenger particle logical relation and write, such as, suppose that the mapping of each messenger particle is respectively A ₁, A ₃, A ₃, A ₄, A ₃, A ₄, 0.5A ₆+ 0.5A ₇, then logical relation group can be written as: A ₁→ A ₃; A ₃→ A ₃, A ₄; A ₄→ A ₃, 0.5A ₆+ 0.5A ₇three groups.

Step4. interval estimation

On the basis obtaining the logical relation between fuzzy set, logic of propositions relation, A _j→ A _j1, A _j2..., A _jpmethod of interval estimation is mainly three kinds of principles:

If principle 1. logical relation group trend is ascendant trend, i.e. j1, j2 ..., jp > j, then the lower limit of interval estimation is m _j(u _jmedian point), the upper limit of interval estimation is (m _j1+ m _j2..., m _jp)/p, thus the interval estimation result obtaining current point in time.

If principle 2. logical relation group trend is downtrending, i.e. j1, j2 ..., jp≤j, then the lower limit of interval estimation is (m _j1+ m _j2..., m _jp)/p, the upper limit of interval estimation is m _j, thus obtain the interval estimation result of current point in time.

If principle 3. logical relation group trend is not only on the rise but also have downtrending, i.e. j1, j2 ..., jk≤j and jk+1 ..., jp > j, then the lower limit of interval estimation is (m _j1+ m _j2..., m _jk)/k, the upper limit of interval estimation is (m _jk+1+ m _jk+2..., m _jp)/(p-k), thus obtain the interval estimation result of current point in time.

The following detailed description of how to build forecast model and to complete performance prediction.It specifically comprises the following steps: data prediction, delimits fuzzy set, builds the relation between fuzzy set, interval estimation.

The detailed description of model construction and application is carried out: adopt in March, 2011 Beijing's Three links theory microwave detection data to carry out the analysis of traffic flow parameter performance prediction by example.It is as shown in table 1 that microwave detects data raw data form:

Table 1 microwave detects raw data (partial data) Table1microwavedetectiondata

Shown in upper table is only the partial data of Beijing's Three links theory partial period record obtained.The traffic flow parameter that microwave raw sensor data contains has the parameter such as the magnitude of traffic flow (being obtained by the statistics number of CAR_ID), occupation rate, speed, for the ease of sample calculation analysis, in these only data based on this parameter of speed, verification model estimation effect.Raw data gets 1-19 time point data, and 20-22 time point data are as verification msg.

On the basis of existing basic data, the pre-service of data is carried out for speed parameter information, speed data as part-time point is 0 value, mainly there is undetected situation because of during detecting device identification, the average can closing on two time points is carried out supplementing of data as interpolation and repairs, and improves the integrity degree of data message.After data prediction, obtain result as shown in table 2:

Table 2 microwave detection speed preprocessed data (partial data) Table2preprocessingofmicrowavedetectiondata (snapshot)

Utilize model of the present invention, carry out the estimation prediction of speed, computation process is as follows:

Step1. research range is defined and comformed information numbers of particles.

According to numerical range span, definition research range U=[35,95], in example, hypothesis is divided into h=7 messenger particle, namely in research range, the species number c=[7/2]=3 of cluster.Research range dividing condition is:

U＝{u ₁,u ₂,...,u ₇}。

Step2. fuzzy set delimited

In research range, delimit fuzzy set, assume that 7 messenger particles in example, general fuzzy set number is identical with it.Utilize messenger particle determination FUZZY SET APPROACH TO ENVIRONMENTAL as follows:

A _i＝1/u _i+0.5/u _i-1+0.5/u _i+1，(i＝1,...,h)(14)

According to said method, with reference to institute of the present invention established model, fuzzy set corresponding membership result of calculation is as shown in table 3:

Table 3 data membership situation

Table3data’sdegreeofmembership

Parameter alpha value is different, and data membership situation may exist certain difference, can utilize the value of historical data calibrating parameters α in practical application, and choosing value is generally array in element value.

Step3. the logical relation between fuzzy set is built

Utilize formula (13) to determine logical relation between fuzzy set and these relations are arranged arrangement, obtaining the fuzzy relation group needed in model, as shown in table 4:

Logical relation group Table4logicrelationshipgroupsbetweenfuzzysets between table 4 fuzzy set

Step4. interval estimation

The estimation of messenger particle feature structure that what Fuzzy time sequence had been predicted is, i.e. the estimation of particle length.The estimated value of the time point comprised in messenger particle all arbitrary value can substitute in burst length, and the error of generation is all in model allowed band.Messenger particle number is determined according to actual needs, and the time point that messenger particle comprises is fewer, and grain refine degree is higher, and estimated accuracy is also higher, but calculated amount also can fastly increase thereupon.When huge data, the time point contained by messenger particle is more, but huge data are generally concerned about is not concrete time point, but the estimated accuracy of time period, with the interval estimation not contradiction of messenger particle, also can reduce workload simultaneously.According to interval estimation principle in this paper, in conjunction with the logical relation group between the fuzzy set obtained, 7 time period (messenger particle) interval estimation results can be obtained as follows:

Table 5 interval estimation result Table5estimationresultsinIntervals

According to upper table, subsequent time period predict the outcome as [79.5,98.5], owing in example dividing the data of 19 time points in order to 7 messenger particles, representing the time point that a time period contains is 2-3.The result more than showing to estimate represents that the data value of a following 2-3 time point should in [79.5,98.5].The known estimated result of True Data value putting 20-22 observing time is accurate.

Certainly, due to the impact of the factor such as accidental, general forecast 3 points are not enough to the applicability proving model.Utilize microwave data, proceed prediction experiment again, only get front 19 data of predicted time point as historical data at every turn, draw anticipation trend analysis chart as shown in Figure 4.

As can be seen from anticipation trend figure, the broken line graph trend of predicted data is substantially identical with original True Data broken line graph trend.The fiducial interval of model selective analysis data of the present invention, focuses on and the contrasting of the variation tendency of actual value.During continuous prediction, according to the input of historical data Renewal model, performance prediction can be realized by shown in example.

Coefficient given in the above embodiments and parameter; be available to those skilled in the art to realize or use of the present invention; the present invention does not limit and only gets aforementioned disclosed numerical value; without departing from the present invention in the case of the inventive idea; those skilled in the art can make various modifications or adjustment to above-described embodiment; thus protection scope of the present invention not limit by above-described embodiment, and should be the maximum magnitude meeting the inventive features that claims are mentioned.

Claims (8)

1., based on a road traffic flow parameter prediction method for Granule Computing, it is characterized in that specifically comprising the following steps: step one, numerical range span according to the traffic flow parameter detected, definition research range and comformed information numbers of particles , wherein represent the round values more arbitrarily small than numerical value minimum value in overall data, represent the round values larger arbitrarily than numerical value maximal value in overall data; Step 2, in research range, delimit fuzzy set, and determine the membership between the traffic flow parameter data that detect and fuzzy set; Wherein fuzzy set number is identical with messenger particle number; Step 3, the logical relation determining between fuzzy set, obtain fuzzy relation group; Step 4, trend according to fuzzy relation group, adopt Fuzzy time sequence to carry out messenger particle interval estimation, thus dope the traffic flow parameter of subsequent time period.

2., as claimed in claim 1 based on the road traffic flow parameter prediction method of Granule Computing, it is characterized in that the process of described messenger particle interval estimation is specially: logic of propositions relation, if logical relation group trend is ascendant trend, then the lower limit of interval estimation is , the upper limit of interval estimation is ; If logical relation group trend is downtrending, then the lower limit of interval estimation is , the upper limit of interval estimation is ; If logical relation group trend is not only on the rise but also have downtrending, then the lower limit of interval estimation is , the upper limit of interval estimation is , wherein, for the subscript of logical relation front end fuzzy set, represent the individual fuzzy set, the rear end fuzzy set that presentation logic relation front end fuzzy set is corresponding, altogether it is individual, for between intermediate point, for fuzzy set is corresponding median point.

3., as described in claim 1 or 2 based on the road traffic flow parameter prediction method of Granule Computing, it is characterized in that described messenger particle is by data set form, particle characteristic comprises interval range length and contains data point number two features, and messenger particle is expressed as , wherein with for data set boundary, described boundary refers to the boundary up and down of particle, namely particle comprise the boundary up and down of data set.

4., as described in claim 1 or 2 based on the road traffic flow parameter prediction method of Granule Computing, it is characterized in that the Fuzzy time sequence for different time points with , fuzzy logical relationship is regarded as with between logical relation, be designated as , for relation front end, for relation rear end; For same fuzzy set , when several fuzzy logical relationship front end is identical, merge into fuzzy logical relationship group.

5., as described in claim 1 or 2 based on the road traffic flow parameter prediction method of Granule Computing, it is characterized in that working as time series , altogether individual sample, corresponding time coordinate is , pass through degree of membership and cluster centre the fuzzy clustering on time data collection that the minimization function formed completes, minimization function is configured to: ; Wherein it is the data point comprising time coordinate; weighted index, ; cluster species number, ; it is data point with cluster centre between distance; represent in individual sample individual data point is under the jurisdiction of the degree of membership of class.

6., as described in claim 1 or 2 based on the road traffic flow parameter prediction method of Granule Computing, it is characterized in that for research range, this research range is divided into the interval of individual unequal length, namely individual messenger particle, determines this the method of individual burst length is as follows: Step1. determines species number , calculate corresponding degree of membership; Species number , expression is no more than the maximum integer of value, for interval number, and calculate cluster centre and the degree of membership of correspondence ; Step2. according to degree of membership construction data subset; For cluster centre and corresponding degree of membership , construction data subset is as follows: , ..., ; Step3. messenger particle is built; it is data subset cluster centre, be ; Utilize the upper and lower boundary of optimum of messenger particle building method computing information particle, messenger particle is ; Step4. determine that corresponding research range is interval .

7., as claimed in claim 1 based on the road traffic flow parameter prediction method of Granule Computing, it is characterized in that described method also comprises and pre-service is carried out to the traffic flow parameter detected.

8., as claimed in claim 1 based on the road traffic flow parameter prediction method of Granule Computing, it is characterized in that described pre-service is specially the average adopting and close on two time points and carries out supplementing of data as interpolation and repair.