CN105355038A - Method for predicting short-term traffic flow through employing PMA modeling - Google Patents

Abstract

The invention discloses a method for predicting the short-term traffic flow through employing PMA modeling. The method comprises the steps: preliminary preparation- defining terms; PMA modeling-building an independent model; PMA modeling-building a similar model; and PMA modeling-carrying out the verification and evaluation of fusion and prediction results. The method has the following advantages in that (1) the method integrates independent and similar flow modes while making prediction; (2) the method does not need a large number of traffic data sets; (3) compared with a conventional prediction method, the method is stronger in generalization capability for predicting complex and unpredictable urban traffic flow.

Description

A kind of method utilizing PMA modeling and forecasting short-term traffic flow

Technical field

The invention belongs to technical field of intelligent traffic, particularly relate to a kind of method utilizing PMA modeling and forecasting short-term traffic flow.

Background technology

Traffic flow forecasting is a pith of traffic programme, traffic control, intelligent transportation system, and Accurate Prediction magnitude of traffic flow environment is seen as the effective ways of region traffic control for a long time.These methods briefly can be divided into i) short-term and ii) long-term traffic flow forecasting.Long-term forecasting provides the traffic flow forecasting of several months or several years, and is generally used for long-term traffic programme or construction.Short-term forecasting for short-term as the magnitude of traffic flow in 1 hour is made prediction.Especially, short-term traffic flow prediction can support dynamic traffic control in advance.Therefore, forecasting techniques has caused the attention of traffic engineer and researchist.According to the data type of application and the final potential use of prediction, technology different in a large number has been used to short-term traffic flow prediction.These technology comprise running mean method, k near neighbor method, and autoregression MA (ARIMA) model or cycle ARIMA (SARIMA) method, neural network (NNS) and complex technique are as DA method.

The exploitation of model of traffic flux forecast depends primarily on history and present traffic flow data.Traffic flow forecasting problem is the time series forecasting task of a standard, its target be estimate traffic direction future value and the magnitude of traffic flow before and funtcional relationship now between observation data, and short-term traffic flow prediction is a MISO system, illustrate the relation of magnitude of traffic flow environment in the past and future traffic flow environment.Nearest document concentrates on three aspects: i) propose method that is new or that strengthen (" the Predictionofshort-termtrafficvariablesusingintelligentsw arm-basedneuralnetworks " of the people such as KitYanChan and TharamDillon, " Short-termtrafficflowforecastingbasedongreydelaymodel " of the people such as HuanGuo and XinpingXiao, " the Inferringgasconsumptionandpollutionemissionofvehiclesthr oughoutacity " of the people such as JingboShang and YuZheng), ii) (" the Neural-network-basedmodelsforshort-termtrafficflowforeca stingusingahybridexponentialsmoothingandlevenberg – marquardtalgorithm " of the people such as KitYanChan and TharamSDillon that effectively predict the outcome is merged, " Multi-phasetimeseriesmodelsformotorwayflowforecasting " of the people such as MohsenDavarynejad and YubinWang, " Anaggregationapproachtoshort-termtrafficflowprediction " of the people such as Man-ChunTan and SCWong), and iii) effective preconditioning technique (" the Trafficflowforecastingneuralnetworksbasedonexponentialsm oothingmethod " of the people such as KitYanChan and TSDillon is proposed, " the Developmentofneuralnetworkbasedtrafficflowpredictorsusin gpre-processeddata " of KitYanChan and CedricKFYiu, " the Atrafficpredictionmethodbasedonannandadaptivetemplatemat ching " of QGao and GLi).

But some significant problems that existing method exists also are acknowledged.The first, the Data Collection that these methods use is in highway and highway, and the magnitude of traffic flow change of there seems very stable.The second, the data that they use are collected from single point source, very limited in more complicated cases.3rd, existing Forecasting Methodology does not consider that Temporal Data is as the interactivity of complexity and the comprehensive of data in densely populated urban district.Precisely, Traffic Flow Forecasting problem in urban area predict with freeway traffic compared with more complicated, owing to being subject to various constraint as the restriction of telltale etc., and in the flow collection of urban district, containing mass data point, these data points are difficult to the relation processed between them by method before.Incoherent transport hub may also have similar Changing Pattern and present identical flow mode.Traditional method is not considered similar flow mode and is utilized these similaritys, and they usually reduce the angle of data source or by Complex Modeling, these reasons make them be applied in reality.

Chinese patent CN102496284A discloses a kind of road traffic flow collection and Forecasting Methodology, comprise toroidal inductor, vehicle detection module, magnitude of traffic flow acquisition module, traffic flow data pre-service and prediction, road traffic flow data prediction and forecasting software carry out on host computer (PC), and read the traffic flow data in acquisition module (SD card) by network interface.For improving forecasting reliability, in road traffic flow data prediction and Forecasting Methodology, wavelet analysis is first adopted to carry out noise eliminating in conjunction with least square method to traffic flow data; Then the BP neural network model of traffic flux forecast of improvement is adopted, realize the prediction to the magnitude of traffic flow, for the control timing scheme and traffic planninng of optimizing road traffic provide foundation, this invention is by combination of hardware algorithm, predicting traffic flow amount, but have employed BP neural network forecast model, most accurately predicting cannot be reached.Chinese patent CN103870890A discloses the Forecasting Methodology of a kind of freeway net magnitude of traffic flow distribution, suppose to predict the freeway net following generation of each access station and following traffic attraction of each outlet station under scale sometime, present situation flow distribution and present situation utilization coefficient is calculated by the historical traffic distribution of website, computational prediction flow distribution afterwards, predicted the outcome, the degree of accuracy of algorithm is lower, only can be effective to the high speed that flow is not intensive, and applicability is poor.

Summary of the invention

For solving Forecast of Urban Traffic Flow prediction out of true, the problem of modeling complexity, we have proposed a kind of method utilizing PMA modeling and forecasting short-term traffic flow, and adopt the present invention can reach Forecasting Methodology simple, result is object accurately.

The present invention is achieved by the following technical solutions:

For achieving the above object, we have proposed a kind of method utilizing PMA modeling and forecasting short-term traffic flow, step is as follows:

(1) tentatively prepare, definition term:

Sampling flow: traffic flow data free in collect, each sampling flow relates to the sum through vehicle within certain specific time cycle.Each cycle is Δ t minute, the magnitude of traffic flow time slot (t-Δ t, t] in collect, t is integer; Q (t) is equally also integer, represents the _tthe magnitude of traffic flow in cycle;

Flow mode: for each sampling flow, q (t) is source time sequence; 2 time related sequences form periodic sequence S _a(t) and day sequence S _b(t), as follows:

1. .S _a(t) be comprise q (t) before k _athe set of individual cycle traffic data on flows;

$S_a\left(t\right)=\{q(t-k_a\mathrm{\Δ}t),q(t-(k_a-1)\mathrm{\Δ}t),...,q(t-\mathrm{\Δ}t),\hat{q}\left(t\right)\};$

2. .S _bt () is for comprising q (t) k _bthe set of magnitude of traffic flow record in the same time cycle before it;

$S_b\left(t\right)=\{q(t-1440k_b),q(t-1440(k_b-1)),...,q(t-1440),\hat{q}\left(t\right)\};$

(2) PMA modeling-set up independent model:

I. the plane flow mode of independent model: S _at () is horizontal flow mode, S _bt () is vertical flow mode, for predicted value;

Ii. mode weights:

$\begin{array}{cc}\mathrm{\&rho;}(m-i)=\frac{1}{L}\underset{k_a=1}{\overset{L}{\&Sigma;}}y(m,\mathrm{\&delta;})/y(m-i,\mathrm{\&delta;}-k_a)& 0<i<m\end{array}$

L represents the length of PFP, and δ represents row, m is the data set number of target day.

Iii.PMA predicted value:

W is the width (see Fig. 2) of PFP.The corresponding PMA size L specified of each PMA predicted value _i× W _i

(3) PMA modeling-set up scale model:

WWL algorithm: between two time serieses (equal length), we use and define similarity by mixing of forming of Euclidean distance and Pearson correlation coefficient apart from calculating to generate distance matrix;

PFP _similarity(t)＝{s(t)，s ¹(t)，...，s ⁱ(t)，...，s ^d(t)}.

Here s (t) represents target flow mode, s ⁱt () represents i-th mode similar to s (t), digital d represents the parameter of WWL algorithm;

(4) PMA modeling-fusion:

Select neural network model to predict the outcome obtain before two to merge;

$h_i=\tan sig(\underset{j=1}{\overset{2}{\&Sigma;}}{w}_{ij}{p}^j-{\mathrm{\&theta;}}_1)$

$net=purelin(\underset{j=1}{\overset{n}{\&Sigma;}}{w}_i{h}_i-{\mathrm{\&theta;}}_2)$

N represents neuronic number in hidden layer, h _irepresent the output of hidden layer, net represents final output;

(5) predict the outcome test evaluation.

Preferably, before being necessary at least one sky in above-mentioned steps (4).

The invention has the beneficial effects as follows:

(1) the method combines independence and similar flow mode simultaneously when making prediction;

(2) without the need to a large amount of traffic data collection;

(3) the method prediction complicated uncertain Forecast of Urban Traffic Flow in before Forecasting Methodology compared with there is stronger generalization ability.

Accompanying drawing explanation

Fig. 1 is the different sizes of PFP in PMA training process;

Fig. 2 is independent model and the training RMSE of scale model under Different L EssT.LTssT.LTL, W> value;

Fig. 3 is the performance comparison of the relation that training dataset and test data concentrate concealed nodes number and RMSE, submodel and Fusion Model;

Fig. 4 is that RMSE and the MPAE of different model compares.

Embodiment

In order to make technical matters solved by the invention, technical scheme and beneficial effect clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

Embodiment 1

Utilize a method for PMA modeling and forecasting short-term traffic flow, step is as follows:

(1) tentatively prepare, definition term:

Sampling flow: traffic flow data free in collect, each sampling flow relates to the sum through vehicle within certain specific time cycle.Without loss of generality, we select each cycle Δ t minute, the magnitude of traffic flow time slot (t-Δ t, t] in collect, t is integer.Q (t) is equally also integer, represents the magnitude of traffic flow in the t cycle.

Flow mode: for each sampling flow, q (t) is source time sequence.By analyze observation traffic flow data, can find magnitude of traffic flow mode almost every day loop cycle.Therefore, 2 time related sequences form periodic sequence S _a(t) and day sequence S _b(t), as follows:

1. S _a(t) be comprise q (t) before k _athe set of individual cycle traffic data on flows.

$S_a\left(t\right)=\{q(t-k_a\mathrm{\&Delta;}t),q(t-(k_a-1)\mathrm{\&Delta;}t),...,q(t-\mathrm{\&Delta;}t),\hat{q}\left(t\right)\}$

2. S _bt () is for comprising q (t) k _bthe set of magnitude of traffic flow record in the same time cycle before it.

$S_b\left(t\right)=\{q(t-1440k_b),q(t-1440(k_b-1)),...,q(t-1440),\hat{q}\left(t\right)\}$

(2) PMA modeling-set up independent model:

I. the plane flow mode of independent model:

We use a kind of new PMA model for solving short-term traffic flow forecasting problem.PMA is the innovatory algorithm of a kind of running mean (MA).Especially, PMA runs in plane flow mode (two dimensional data structure). for predicted value;

Predict to utilize independence and similar flow mode simultaneously, we illustrate two kinds of models, independent model utilizes the traffic flow data in self past to carry out modeling to Independent fluxes mode, and scale model utilizes similar teacher's traffic flow data to carry out modeling to similar flow mode.They utilize neural network to combine generation forecast device.It should be noted that two kinds of models all use PMA algorithm.Next we will provide the data constituting method of scale model, independent model, PMA training method and Model Fusion.

We claim S _at () is horizontal flow mode, S _bt () is vertical flow mode, and propose a kind of new thinking to set up plane flow mode (PFP).S _a(t), S _b(t), S _b(t-k _aΔ t) front k _aindividual cycle and S _a(t-1440k _b) front k _bit together constitutes plane flow mode (PFP), and the timestamp of every two neighbor q has distinguished Δ t minute or day.We can find k _aΔ t is substantially equal to 1440, and this illustrates that certain sampling flow will occur repeatedly and be recycled in PFP.

PMA algorithm is used for being predicted the outcome.Conveniently, we will use y (i, j) to represent the value of the i-th row jth row in plane flow mode.

Ii. mode weights:

Each flow mode has contribution to forecasting accuracy, and therefore we define a kind of mode weights ρ for describing the contribution of different modalities in target prediction.

$\begin{array}{cc}\mathrm{\&rho;}(m-i)=\frac{1}{L}\underset{k_a=1}{\overset{L}{\&Sigma;}}y(m,\mathrm{\&delta;})/y(m-i,\mathrm{\&delta;}-k_a)& 0<i<m\end{array}$

Here L represents the length (see Fig. 1) of PFP, and δ represents row, m is the data set number of target day.

Iii.PMA predicted value:

Similar to the computing method of a kind of simple smoothing technique MA.

Here W is the width (see Fig. 1) of PFP.The corresponding PMA size L specified of each PMA predicted value _i× W _i.

Such as, PFP is considered by three kinds of S _amode and four kinds of S _bmode is formed, and mode contiguous is up and down the Different periods in a day, if desired value be 4. we use mode { 1,2,3} and { 10,8} obtains mean predicted value, and we will obtain value 2 and 9, but they both keep off 4 simply.We will obtain two weights 0.478 and 0.639 to use mode weights, and we then can become (10 × 0.478+8 × 0.639)/2=4.946 at final predicted value.Obviously, 4.946 is immediate values.

PMA algorithm:

1: first, for PFP, utilize formula (1) to calculate mode weights ρ;

2:<X, Y> are the size of training dataset;

3：forL＝1...Xdo

4：forW＝1...Ydo

5: for each <L, Y>, through type (2) calculates the PMA predicted value of all training samplings;

6: calculate each predicted value and actual value between average RMSE;

7: utilize minimum RMSE method to upgrade best PMA large scale;

8：endfor

9：endfor

10: use best PMA size to predict.

In design PMA training method, it is important attribute that PMA size is concentrated at training data, needs carefully to choose.PMA algorithm describes in detail in algorithm 1.In order to obtain PMA size, this size can be concluded and do well in new situation, and data sampling is divided into two subsets usually: 1) training dataset, 2) test data set.

(3) PMA modeling-set up scale model:

In most of the cases, the time series of identical type shows similar Changing Pattern, especially in some specific region (such as same area).Such as, more than 1000 cars/hour flow can not occur in the major trunk roads of samll cities and towns.Therefore, we utilize mixing distance calculation criterion, and this criterion is made up of Euclidean distance and Pearson correlation coefficient, carrys out the similarity of evaluation time sequence.And the N number of Similar Time Series Based on Markov Chain selecting rank the most front is predicted.Notice and can compare time series needs at least before one day if do not had, then do not have known data to use.

WWL (we are as which) algorithm: between two time serieses (equal length), we use and define similarity by mixing of forming of Euclidean distance and Pearson correlation coefficient apart from calculating to generate distance matrix.These two matrixes of minimization can ensure the highest similarity degree of the magnitude of traffic flow and terrestrial reference simultaneously.First, by calculating Pearson correlation coefficient $pcc(s^i,s^j)=\frac{{\&Sigma;}_{t=1}^m\left(s^i\right(t)-\stackrel{\&OverBar;}{s^i})\left(s^j\right(t)-\stackrel{\&OverBar;}{s^j})}{\sqrt{{\&Sigma;}_{t=1}^m{\left(s^i\right(t)-\stackrel{\&OverBar;}{s^i})}^2{\&Sigma;}_{t=1}^m{\left(s^j\right(t)-\stackrel{\&OverBar;}{s^j})}^2}}$ Filter time series, pcc value more than 0.8 expression shape closely.Then, by equation ed (s ⁱ, s ^j)=|| s ⁱ-s ^j|| calculate Euclidean distance and guarantee that the magnitude of traffic flow is close as much as possible, and sort by once method:

PFP _similarity(t)＝{s(t)，s ¹(t)，...，s ⁱ(t)，...，s ^d(t)}.

Here s (t) represents target flow mode, s ⁱt () represents i-th mode similar to s (t), digital d represents the parameter of WWL algorithm.Remember s ⁱbefore t () is necessary at least one sky.

Empirically, a data on flows, more similar in shape, then flow mode is more useful.In addition, run to simplify PMA, the similarity that we calculate according to WWL algorithm, by them from rear to front alinement, with the PFP in independent model _individualt () is the same, PMA algorithm is used for acquisition and predicts the outcome.

(4) PMA modeling-fusion:

Select suitable model to predict the outcome obtain before two to merge.There is many popular methods to apply and merge, such as linear regression, logistic regression, neural network etc.For validity and indirect, we select neural network to merge to connect a model that we have proposed.

Thus, scale model and independent model can obtain a series of predicted value (w=1 represents scale model, and w=2 represents independent model), m is the numbering of Temporal Sampling result here, the predicted value of expression time m.Naturally, we have true magnitude of traffic flow sequence, R=(r ₁, r ₂..., r _m) be target of prediction, then we can set up data set as (P ¹, P ², R).

In design NN model, the activation function of hidden layer and output layer is respectively tansig and purelin.Because LM algorithm convergence is the fastest, therefore we use Levenberg-MarquardtBP algorithm as learning algorithm, are best learning criterion in the case.

$h_i=\tan sig(\underset{j=1}{\overset{2}{\&Sigma;}}{w}_{ij}{p}^j-{\mathrm{\&theta;}}_1)$

$net=purelin(\underset{j=1}{\overset{n}{\&Sigma;}}{w}_i{h}_i-{\mathrm{\&theta;}}_2)$

Here n represents neuronic number in hidden layer, h _irepresent the output of hidden layer, net represents final output.Although accurate criterion is for determining the size of optimum training data set and test data set, we advise that training dataset is larger.

(5) predict the outcome test evaluation:

We utilize the data on flows collection of actual measurement, have done the performance that proposed PMA model is evaluated in a large amount of experiments.In this part, first we will set with experiment by decryption collection, next by Evaluations matrix measurement performance, and last experimental results.

1. survey region

Data for testing consist of record the actual measurement vehicle on May 5th, 1 day 1 April in 2013 of the traffic surveillance and control system record of metropolitan area.Data set comprises 140, and 440,933 actual measurement vehicles are by record, and relate to whole 308 cameras being positioned at major trunk roads, we are divided into two parts, training dataset and test data set.The training dataset of above model contains on May 2,1 day to 2013 April in 2013, and test data set is the traffic records on May 3rd, 2013 to May 5.The experimental data of our all camera watch regions is carried out average.Traffic flow data has carried out being polymerized also average in each cycle Δ t grouping.

2. pre-service

We carry out pre-service to data set, form flow mode, calculate the quantity of vehicle in each time cycle, advise Δ t off time to be set to 60.If because camera damages or other reasons, sampling flow loses (being 0 in this situation), and we fill up it by the value of average hinge always.After pre-processing, we obtain amounting to 10780 horizontal flow mode and 7392 vertical modes, and we predict the magnitude of traffic flow of following one-period in an experiment.

3. goodness of fit statistics

We use two to evaluate statistical method the most widely and obtain the accuracy predicted the outcome.

1) root-mean-square error (RMSE) is used to measure a kind of method of average error of predicting the outcome, and can be calculated by following formula:

$RMSE=\sqrt{\frac{1}{N}\underset{n=1}{\overset{N}{\&Sigma;}}{(y_n-{\hat{y}}_n)}^2}$

2) average absolute value percentage error (MPAE) is a kind of method measuring the proportional error that predicts the outcome, and can be calculated by following formula

$MPAE=\frac{1}{N}\underset{n=1}{\overset{N}{\&Sigma;}}\frac{\left|y_n-{\hat{y}}_n\right|}{y_n}\&times;100\%$

Y herein _nwith represent observed reading and the predicted value of n-th observation respectively, N represents the sum of observation.

4. PMA is evaluated

We evaluate the performance of PMA algorithm in this two Seed model, independent model and scale model, and their fusion.For each experiment, we predict all camera regions, and the mean value of Report Evaluation statistics.

First, we change the parameter <L of independent model and scale model in formula (1) and formula (2), W> (mentioning above, the size for planar sequence).Can be reflected the absolute error of the magnitude of traffic flow by RMSE, we select it to help train our model.Fig. 2 (a) and (b) show independent model and the RMSE of scale model under Different L EssT.LTssT.LTL, W> value, and wherein RMSE calculates based on 24 observations (60 minutes, 1 day).Based on the result shown in figure, when RMSE is minimum, we select L=6 for independent model, W=2, select L=7, W=2 as the optimum dimension of PMA for scale model.

Secondly, for NN Fusion Model, in hidden layer, the sequence of the neuronic NN of different number is trained.Neuronic number changes between 4 to 12, calculates the RMSE of training dataset and test data set simultaneously.According to the generalization ability of test data set self, the value of RMSE is lower, and network model is better.Fig. 3 (a) shows the relation curve of RMSE and hidden layer neuron number.In Fig. 3 (a), we find that best hidden layer neuron number is 8.Therefore, 2-8-1NN model is selected as predicting.The performance comparison of submodel and Fusion Model is shown in Fig. 3 (b).We can block RMSE and significantly reduce and suitable improve of MPAE.

5. Comparative result

Some single source models comprise mA, ARIMA, and NN model, and polymerization model is used to time series S _a(t) and S _b(t).We compare them and utilize the prediction of PMA model in test data set.

1) (or constant) model can use the simplest form for traffic flow forecasting:

${\hat{q}}_{Na}\left(t\right)=q(t-1)$

Herein for the predicted value in the t of time slot.

2) MA model: MA Model sequence can be calculated by following formula:

${\hat{y}}_{t+1}=\frac{y_t+y_{t-1}+t_{t-2}+...+t_{t-k+1}}{k}$

K is the number [12] of MA item herein.MA technology only can process the given data in a last k cycle, each average in the number of data point constant in time.

3) ARIMA model: the common ARIMA model of sequence (r, d, s), d represents the exponent number of difference herein, and exponent number r and s is AR and MA operational symbol.

4) artificial NN: for single source model contrast with PMA, NN model is for fit non-linear relation:

${\hat{q}}_{NN}\left(t\right)=f_1(q(t-1),q(t-2),...,q(t-l))$

NN model be input as front l continuous time gap magnitude of traffic flow record, its output is the traffic flow forecasting of time slot t.The number of input l and the number of NN hidden neuron are also by optimum experimental.5) polymerization model (DA): it has merged three Seed models (MA, ES and ARIMA), uses simple neural net method fusion forecasting result.

Note different model ( mA, ARIMA, NN, DA) need the historical data of different length.In addition, sample size needs to select for each model.Such as, for the time span of model training data can not exceed predicted time first 4 hours, for the training data of ARIMA model, and before at least will comprising 2 days.For each model, we by observing best matching or prediction Selection parameter, and compare predicting the outcome on same test data set.

We are following one-period generation forecast.Fig. 4 gives and uses RMSE and the MAPE of different forecast model under test data set.For we use end value as predicted value simply.For MA model, we arrange k=3 is optimal parameter.For ARIMA model, our parameters is (1,1,0), and now RMSE is minimum.We arrange α=0.1 and set up DA method with γ=0.1 in ES model.For last a kind of, we select 3-12-1NN model to be used for comparing.Fig. 4 (a) and (b) show the result that PMA predicts and are better than predicting the outcome of other 5 kinds of methods.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, all any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within protection scope of the present invention.

Claims (2)

1. utilize a method for PMA modeling and forecasting short-term traffic flow, it is characterized in that, step is as follows:

(1) tentatively prepare, definition term:

Sampling flow: traffic flow data free in collect, each sampling flow relates to the sum through vehicle within certain specific time cycle.Each cycle is Δ t minute, the magnitude of traffic flow time slot (t-Δ t, t] in collect, t is integer; Q (t) is equally also integer, represents the _tthe magnitude of traffic flow in cycle;

Flow mode: for each sampling flow, q (t) is source time sequence; 2 time related sequences form periodic sequence S _a(t) and day sequence S _b(t), as follows:

1. .S _a(t) be comprise q (t) before k _athe set of individual cycle traffic data on flows;

$S_a\left(t\right)=\{q(t-k_a\mathrm{\&Delta;}t),q(t-(k_a-1\left)\mathrm{\&Delta;}t\right),...,q(t-\mathrm{\&Delta;}t),\hat{q}\left(t\right)\};$

2. .S _bt () is for comprising q (t) k _bthe set of magnitude of traffic flow record in the same time cycle before it;

$S_b\left(t\right)=\{q(t-1440k_b),q(t-1440(k_b-1\left)\right),...,q(t-1440),\hat{q}\left(t\right)\};$

(2) PMA modeling-set up independent model:

I. the plane flow mode of independent model: S _at () is horizontal flow mode, S _bt () is vertical flow mode, for predicted value;

Ii. mode weights:

$\mathrm{\&rho;}(m-i)=\frac{1}{L}\underset{k_a=1}{\overset{L}{\&Sigma;}}y(m,\mathrm{\&delta;})/y(m-i,\mathrm{\&delta;}-k_a),0<i<m$

L represents the length of PFP, and δ represents row, m is the data set number of target day.

Iii.PMA predicted value:

$\mathrm{PMA}\left(\mathrm{\&delta;}\right)=\frac{1}{W}\underset{k_b=1}{\overset{W}{\mathrm{\&Sigma;}}}y(m-k_b,\mathrm{\&delta;})\&times;\mathrm{\&rho;}(m-k_b)$

W is the width (see Fig. 2) of PFP.The corresponding PMA size L specified of each PMA predicted value _i× W _i

(3) PMA modeling-set up scale model:

WWL algorithm: between two time serieses (equal length), we use and define similarity by mixing of forming of Euclidean distance and Pearson correlation coefficient apart from calculating to generate distance matrix;

PFP _similarity(t)＝{s(t)，s ¹(t)，...，s ⁱ(t)，...s ^d(t)}.

Here s (t) represents target flow mode, s ⁱt () represents i-th mode similar to s (t), digital d represents the parameter of WWL algorithm;

(4) PMA modeling-fusion:

Select neural network model to predict the outcome obtain before two to merge;

$h_i=\tan sig(\underset{j=1}{\overset{2}{\&Sigma;}}{w}_{ij}{p}^j-{\mathrm{\&theta;}}_1)$

$net=purelin(\underset{i=1}{\overset{n}{\&Sigma;}}{w}_i{h}_i-{\mathrm{\&theta;}}_2)$

N represents neuronic number in hidden layer, h _irepresent the output of hidden layer, net represents final output;

(5) predict the outcome test evaluation.

2. a kind of method utilizing PMA modeling and forecasting short-term traffic flow as claimed in claim 1, is characterized in that, before being necessary at least one sky in described step (4).