CN103489039A - Expressway traffic flow fusing and forecasting method with online self-tuning and optimizing function - Google Patents

Abstract

The invention relates to the field of traffic flow state analysis and provides an expressway traffic flow fusing and forecasting method with the online self-tuning and optimizing function, wherein according to the expressway traffic flow fusing and forecasting method, online optimization and updating can be conducted on a model, the traffic flow characteristics which change along with time can be adapted, and the forecasting accuracy can be improved. The method includes the steps of obtaining expressway detection traffic data, setting up a sliding time window, reading the expressway detection traffic data through the sliding time window, processing the data in the current sliding time window, comparing forecasting accuracy of four types of monomer forecasting models in the current sliding time window, training the monomer forecasting model with the poorest forecasting accuracy under the current sliding time window, selecting the other three types of monomer forecasting models, conducting data fusing on forecasting results of the selected three types of monomer forecasting models, and storing the fused forecasting data.

Description

Freeway traffic flow amount fusion forecasting method with online self-tuning optimization ability

Technical field

The present invention relates to the traffic flow modes analysis field, relate in particular to a kind of freeway traffic flow amount fusion forecasting method.

Background technology

In recent years, social economy's fast development, the popularity rate of automobile is greatly improved, while is along with the lifting of China's freeway service level, increasing people select to drive or ride in an automobile to go on a journey by highway, this induces to the management of highway and has proposed new challenge, so the importance of freeway traffic flow amount prediction has just displayed at this.

Aspect the traffic behavior forecast model, the model proposed based on traditional mathematics method (as mathematics statistics and infinitesimal analysis etc.) and based on modern science and technology and method (as neural network and analogue technique etc.) and the model of proposition all has its advantage and limitation.The magnitude of traffic flow of highway is the coefficient result such as to induce by transport need, road network condition, traffic administration control program, share of public transportation, information, so be engraved in variation during the magnitude of traffic flow, in the middle of single forecast model, which does not also have can not keep in the same time precision of prediction definitely preferably, so predict traffic by the advantage under different condition that can merge different forecast models, to improve precision of prediction.Different forecast models respectively has its merits and demerits, between them, is not mutual repulsion, but connects each other, mutually supplements.The key of fusion forecasting method is dynamically determining of weights.Definite resonable degree of weights directly determines the height of this precision of prediction.

The Zheng Weizhong of Tsing-Hua University and Shi Qixin propose a kind of Bayes's combination neural net model and are applied to the prediction of short-term traffic flow, result shows: the estimated performance of model is better than single neural network model on the whole, and has guaranteed the stability of model prediction; Tan Manchun and Feng meat or fish are refined sues for peace running mean (ARIMA) and artificial neural network built-up pattern for short-time traffic flow forecast by autoregression, and result shows: the accuracy of the forecasting accuracy of Fusion Model when using separately separately; Cong Xinyu and Yu Huiqun propose historical trend model and multivariate regression model Weighted Fusion to set up combination forecasting, and utilize average weighted method, give larger weight to more accurate predicted value, thereby have improved the precision of model prediction.

As time goes on, the characteristic of freeway traffic flow can change, and the Fusion Model proposed both at home and abroad use is the algorithm of static and off-line, can't adjust online parameter value and the inner structure of the basic model in Fusion Model, this will cause the precision of prediction of forecast model to descend.

Summary of the invention

In view of this, the invention provides a kind of freeway traffic flow amount fusion forecasting method with online self-tuning optimization ability, model is carried out to the on-line optimization renewal, adapt to time dependent traffic stream characteristics, improve precision of prediction.

The present invention solves the problems of the technologies described above by following technological means:

There is the freeway traffic flow amount fusion forecasting method of online self-tuning optimization ability, comprise the steps:

1) obtain highway and detect traffic data, each group highway detects traffic data and comprises time, flow, speed and occupation rate;

2) set up time slip-window (t-m+1, t), the length that m is time window, need the group number of the highway detection history traffic data that reads; Read highway by time slip-window and detect traffic data, read 1 group of new data at every turn, and delete 1 group of data the oldest;

3) data in current time slip-window are processed: detect whether shortage of data is arranged, if having, carry out interpolation processing;

4) contrast the precision of prediction of 4 kinds of monomer forecast models in current time slip-window, the monomer forecast model that precision of prediction is the poorest is trained under current time slip-window,, training always, to the moment of next monomer precision of forecasting model contrast, is selected other 3 kinds of monomer forecast models simultaneously;

5) predicting the outcome of three kinds of monomer forecast models selected in step 4) carried out to data fusion;

6) preserve the predicted data after step 5) merges;

7) judge that highway detects traffic data and whether upgrades, as upgraded, read new highway detection traffic data, judge whether the data in current time slip-window lack, if having, carry out interpolation processing;

8) judge the current number of times of having been predicted, if the number of times of prediction does not reach m time, return to step 5), if the number of times of prediction reaches m time, return to step 4).

Further, described highway detection history traffic data is time, flow, speed and occupation rate.

Further, described step 3) specifically comprises the steps:

31), in time slip-window (t-m+1, t), check whether the data in this time period have disappearance, if just selected this segment data of disappearance is arranged;

32) set a time starting point, then the time field changed into to data field, as the interpolation horizontal ordinate, using flow, speed and occupation rate respectively as the interpolation ordinate;

33) according to the cubic spline interpolation interpolation.

Further, in described step 4), 4 kinds of monomer forecast models are respectively time series predicting model, Kalman prediction model, neural network prediction model and chaotic prediction model;

For time series predicting model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine optimum sample size;

For the training mechanism of Kalman prediction model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine the dimension of optimum state vector.

For the BP neural network prediction model, if its be the poorest monomer forecast model of precision of prediction, again train forecast model, determine new prediction model parameters.

For the chaotic prediction model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine optimum embedding dimension and time delay, meet the phase space of chaotic characteristic with structure.

Further, in step 5), the method that predicting the outcome of three kinds of monomer forecast models carried out to data fusion is as follows: with three kinds of monomer forecast models, predicted respectively, utilize error inverse proportion method to predict the outcome and give respectively weight, fusion is then predicted the outcome.

Freeway traffic flow amount fusion forecasting method with online self-tuning optimization ability of the present invention, space-time characterisation based on freeway traffic flow, can be by the basic model in the online self-tuning research on optimizing information fusion, the method selects four kinds to have the monomer forecast model that upgrades Training Capability, preferentially mechanism by accuracy comparison, select three kinds of forecast models as basic Fusion Model, unsuccessful forecast model continues preferentially to contrast the planned number of competition basis Fusion Model by online training a period of time in real time and other three kinds of pre-measuring and calculating models, unsuccessful forecast model continues online training in real time, the like, can adapt to very soon time dependent traffic stream characteristics, improve precision of prediction.

The accompanying drawing explanation

Fig. 1 shows the schematic flow sheet of the freeway traffic flow amount fusion forecasting method with online self-tuning optimization ability;

Fig. 2 shows the training process flow diagram of time series predicting model;

Fig. 3 shows the training process flow diagram of Kalman prediction model;

Fig. 4 shows the training process flow diagram of BP neural network prediction model;

Fig. 5 shows the training process flow diagram of chaotic prediction model;

Fig. 6 shows the process flow diagram of fusion forecasting result.

embodiment

Below with reference to accompanying drawing, the present invention is described in detail.

Referring to Fig. 1, there is the freeway traffic flow amount fusion forecasting method of online self-tuning optimization ability, comprise the steps:

1) obtain highway and detect traffic data, each group highway detects traffic data and comprises time, flow, speed and occupation rate; The freeway traffic data field definition of the present embodiment is as shown in table 1:

Table 1 freeway traffic data field definition list

Field name	Data type	Field length	Allow empty	The field implication
					DeviceID	nvarchar	255	√	The microwave detector coding
[Time]	smalldatetime	4	√	Time
					TravelDir	int	4	√	Direction of traffic
Flow	int	8	√	Flow
					Speed	float	8	√	Speed
Share	float	8	√	Occupation rate

2) set up time slip-window (t-m+1, t), t is the current time, the length that m is time window, the value that the group that needs the highway that reads to detect traffic data is counted m considers the data group number of each forecast model needs of traffic characteristics of highway and determines, by test experiments, in the present embodiment, the m value is 60; Read highway by time slip-window and detect traffic data, read 1 group of new data at every turn, and delete 1 group of data the oldest;

3) whether the data in current time slip-window are processed: detecting has shortage of data, if having, current prediction process or on carry out interpolation processing in once prediction process; Specifically comprise the steps:

31), in time slip-window (t-m+1, t), check whether the data in this time period have disappearance, if just selected this segment data of disappearance is arranged;

32) set a time starting point, then the time field changed into to data field, as the interpolation horizontal ordinate, using flow, speed and occupation rate respectively as the interpolation ordinate;

33) according to the cubic spline interpolation interpolation.

In when prediction, usually the slickness of traffic data is had relatively high expectations, as the piecewise linearity low order interpolation of approximating function, can not satisfy condition.Suppose an approximating curve, it has the value identical with function on given node, between two nodes, is a cubic curve, and the whole piece curve has the Second Order Continuous derivative, Here it is cubic spline function.Its mathematical definition is as follows:

Definition: cubic spline function is designated as to S (x), and it is the function be defined on interval [a, b], meets:

(1) S (x) is at each minizone [x _i-1, x _i] on be a cubic polynomial function;

(2) in whole interval, [a, b] is upper, and its second derivative exists and be continuous, continuous in the second derivative of each Nodes.

Cubic spline function is the most basic, most important splines, belongs to the algebraic polynomial of segmentation, and it can pass through any, limited a plurality of nodes on plane, and can guarantee the single order of curve, the continuity of second derivative.Relatively be applicable to changing situation slowly.Cubic spline interpolation is a kind ofly can either overcome the high-order moment interpolation fault, can ensure again the Piecewise Interpolation Method of certain slickness, and it can solve the nonlinear problem in reality preferably.

4) contrast the precision of prediction of 4 kinds of monomer forecast models in current time slip-window, the monomer forecast model that precision of prediction is the poorest is trained under current time slip-window, training always is to the moment of next monomer precision of forecasting model contrast, and selected other 3 kinds of monomer forecast models are as basic Fusion Model simultaneously; 4 kinds of monomer forecast models are respectively time series predicting model, Kalman prediction model, neural network prediction model and chaotic prediction model;

For time series predicting model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine optimum sample size; Referring to Fig. 2, if the precision of time series predicting model is undesirable, so possible reason is that the stationarity of sample of Model Selection is bad, and this just need to train forecast model again, determines optimum sample size.Its training airplane is made as:

A) choose optimum sample size

Can be by the number of times of stationary test by sequence under the more different sample sizes of training under the set time window, choose wherein make calling sequence pass through the stationary test number of times when maximum corresponding sample size be optimum sample size.

When carrying out stationary test, adopt runs test method.What runs test judged is assumed to be: " order that sample data occurs does not have obvious trend, and sequence just has stationarity ".Suppose that the sample statistic adopted has:

N1: the sum that a kind of symbol occurs;

N2: the sum that another kind of symbol occurs;

γ: the sum of the distance of swimming;

Wherein: γ is test statistics.

The shorter small sample time-series (being that N1 and N2 are less than 15) for sequence length, can, after definite level of signifiance α (α=0.05 usually), by checking corresponding " runs test γ distribution table ", judge that whether γ is at distance of swimming sum lower limit γ _lwith distance of swimming sum upper limit γ _ubetween, if so, accept hypothesis, judge that former sequence is stationary sequence, on the contrary no.

When N1 or N2 surpass 15, can be similar to normal distribution, now the statistic of use is:

In formula, ${\mathrm{\μ}}_{\mathrm{\γ}}=\frac{{2N}_1{N}_2}{N}+1,{\mathrm{\σ}}_{\mathrm{\γ}}={\left[\frac{{2N}_1{N}_2({2N}_1{N}_2-N)}{N^2(N-1)}\right]}^{1/2},$ N=N ₁+ N ₂, for the level of signifiance of α=0.05, if | Z|≤1.96 (by 2 σ principles), accept hypothesis, judge that former sequence has stationarity, on the contrary no.

B) model parameter solves

This algorithm adopts the square estimation technique to carry out model parameter and solves.Specifically can be calculated by following formula:

In formula,

with

mean respectively autocorrelation function and the model parameter of sequence, p means the order of model, when calculating autocorrelation function, generally is taken as or N/10, wherein N means sample size, selects here autocorrelation function can be solved by following formula:

${\hat{R}}_k=\frac{1}{N}\underset{t=1}{\overset{N-k}{\mathrm{\Σ}}}{x}_t{x}_{t+k}(k=\mathrm{1,2},...N-1);$

In formula, x _tmean the observed reading in time series.

For the training mechanism of Kalman prediction model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine optimum state vector dimension.

Referring to Fig. 3, if the precision of Kalman prediction model is undesirable, so possible reason is that the number of Model Selection state variable is not optimum, this just need to train forecast model again, the present invention considers to choose one or more observer of upstream according to distance, determines the state vector dimension of optimum optimum.Its training airplane is made as:

A) give initial state vector X (0) and covariance matrix assignment P (0);

B) calculating observation value matrix A (t);

C) according to prediction covariance equation P (t|t-1)=B (t-1) P (t-1|t-1) B ^t(t-1)+Q (t-1) calculates covariance matrix P (t|t-1);

D) according to K (t)=P (t|t-1) A ^t(t) [A (t) P (t|t-1) A ^t(t)+R (t)] ^-1calculate kalman gain matrix K (t);

E) optimum of computing mode vector X (t) forecast valuation

F) basis

provide the result of prediction.

G) constantly adjust the dimension of state vector according to predicting the outcome, so that the forecast model prediction effect reaches optimum.

For the BP neural network prediction model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine new prediction model parameters.

Referring to Fig. 4, if the precision of BP neural network prediction model is undesirable, so possible reason is that the parameter that this model was trained in the past has been not suitable for the data under the current time window, so be necessary again the BP neural network prediction model to be trained.Its training airplane is made as:

A) under the current time window, the initialization model parameter;

B) take out a sample sequence X _p, this sequence is joined to input end;

C) the output o at hidden layer and output layer by each neuron _pj ^lcalculate;

D) calculate the error of output layer, and using error sum of squares as the accurate side function that judges whether neural network restrains

$E_P=\frac{1}{2}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{(T_{\mathrm{pj}}-{o_{\mathrm{pj}}}^l)}^2;$

In formula: T _pjfor output layer j neuronic idea output.If meet the demands learn to finish, otherwise proceed to next step;

E), from the output layer to the ground floor, calculate successively each neuronic partial gradient δ of each layer _pj ^l;

F) calculate every layer of neuronic adjustment amount

revise each layer of weights

$\mathrm{\Δ}{w}_{\mathrm{ij}}^l=a{{\mathrm{\δ}}_{\mathrm{pj}}}^l{o_{\mathrm{pj}}}^{l-1}$

$w_{\mathrm{ij}}^l(k+1)=w_{\mathrm{ij}}^l\left(k\right)+\mathrm{\Δ}{w}_{\mathrm{ij}}^l;$

In formula, a is the learning rate coefficient, usually value between 0.01-1;

G) k=k+1, input new sample, returns to b), until E _ptill meeting the demands.

For the chaotic prediction model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine optimum embedding dimension and time delay, meet the phase space of chaotic characteristic with structure.

Referring to Fig. 5: if the precision of chaotic prediction model is undesirable, so possible reason is that the phase space of this Construction of A Model has not had chaotic characteristic, this just need to train forecast model again, determines optimum embedding dimension m and delay time T so that structure meets the phase space of chaotic characteristic.Its training airplane is made as:

A), under the current time window, calculate embedding dimension m and the time delay τ of ordered series of numbers _m, obtain phase space reconstruction:

X(t)={x(t),x(t-τ),…,x(t-(m-1)τ)},t=1,2,...,M；

In formula, the number that M=N-(m-1) τ is the phase space reconstruction mid point;

B) to be that current phase space is no have chaotic characteristic in judgement, if chaotic characteristic is arranged, proceeds to next step, otherwise continue to adjust, embeds dimension m and time delay τ _m;

C), if there is chaotic characteristic, according to the phase space of reconstruct, calculate maximum Lyapunov exponent λ ₁;

D) according to maximum Lyapunov exponent, traffic flow forecasting is carried out in definition;

E), according to prediction effect, constantly adjust and embed dimension m and time delay τ _m, so that the forecast model prediction effect reaches optimum.

5) predicting the outcome of three kinds of monomer forecast models selected in step 4) carried out to data fusion: with three kinds of monomer forecast models, predicted respectively, utilize error inverse proportion method to predict the outcome and give respectively weight, fusion is then predicted the outcome; Fusion forecasting mechanism is as shown in Figure 6:

Under the same time window, basic Fusion Model still exist precision of prediction height minute, therefore, in order further to improve precision of prediction, reduce predicated error, need to carry out the model fusion to basic Fusion Model.The thought merged is to fully take into account precision that basic Fusion Model predicts the outcome within front several periods and the suitable environment of model itself.To predicting the outcome, merged, computing formula is:

$\hat{y}\left(t\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_i\left(t\right)\·{\hat{y}}_i\left(t\right);$

In formula:

i kind algorithm is in the t predicted value in the moment;

weight;

This shows, most important step is determining of weight, and it has determined that certain model output information is to the role that finally predicts the outcome, and will directly have influence on the precision of model fusion.The weight of expectation should be constantly to adjust according to the variation of predicated error, so that best the predicting the outcome of precision can be played maximum effect to final output.For this reason, the definition dynamic error is:

$e_{d,i}\left(t\right)=\frac{1}{k}[e_{\mathrm{ar},i}\left(t\right){+e}_{\mathrm{ar},i}(t-1)+...+e_{\mathrm{ar},i}(t-k)];$

In formula: e _d,i(t) _ i model is in the dynamic error of t period, and it is actually t k interior i model e of period before _{ar, i}(t) average;

K_ error accumulation number, determine according to the sum of predicted data the value that it is suitable usually;

E _{ar, i}(t) absolute relative error of _ t period i model prediction result.

E _{ar, i}(t) computing formula is:

$e_{\mathrm{ar},i}\left(t\right)=\left|\frac{y\left(t\right)-{\hat{y}}_i\left(t\right)}{y\left(t\right)}\right|;$

In formula: the measured data of y (t) _ t period

model is in the predicted value of t period

After obtaining the dynamic error of every kind of forecast model, just can determine accordingly the fusion weight w of each model prediction result _i(t).W _i(t) be one along with e _d,i(t-1) change and the continuous function changed.The present invention adopts the inverse proportion method to determine weight, and its principle is that weight and error size are inversely proportional to, and error is large gives little weight, and error is little gives large weight.

Its flow chart description is as follows:

A) obtain the predicated error of nearly 10 times and the initialization fusion weight w of three kinds of prediction fusion forecasting models _i(t) be 0;

B) obtain respectively 10 error sums of three kinds of forecast models, and whether the error in judgement sum be 0, if 0, again giving this model error sum assignment is 0.01;

C) take each model error sum as data according to obtain the weight of each forecast model by the inverse proportion method;

D) utilize weight to be weighted summation to predicting the outcome of each model, acquisition predicts the outcome, and is deposited in database.

6) preserve the predicted data after step 5) merges;

7) judge that highway detects traffic data and whether upgrades, as upgraded, read new highway detection traffic data, judge whether the data in current time slip-window lack, if having, carry out interpolation processing; ;

8) judge the current number of times of having been predicted, if the number of times of prediction does not reach m time, return to step 5), if the number of times of prediction reaches m time, return to step 4).

Finally explanation is, above embodiment is only unrestricted in order to technical scheme of the present invention to be described, although with reference to preferred embodiment, the present invention is had been described in detail, those of ordinary skill in the art is to be understood that, can modify or be equal to replacement technical scheme of the present invention, and not breaking away from aim and the scope of technical solution of the present invention, it all should be encompassed in the middle of claim scope of the present invention.

Claims (5)

1. there is the freeway traffic flow amount fusion forecasting method of online self-tuning optimization ability, it is characterized in that: comprise the steps:

1) obtain highway and detect traffic data, each group highway detects traffic data and comprises time, flow, speed and occupation rate;

2) set up time slip-window (t-m+1, t), the length that m is time window, need the highway that reads to detect the group number of traffic data; Read highway by time slip-window and detect traffic data, read 1 group of new data at every turn, and delete 1 group of data the oldest;

3) data in current time slip-window are processed: detect whether shortage of data is arranged, if having, carry out interpolation processing;

4) contrast the precision of prediction of 4 kinds of monomer forecast models in current time slip-window, the monomer forecast model that precision of prediction is the poorest is trained under time slip-window, training always, to the moment of next monomer precision of forecasting model contrast, is selected other 3 kinds of monomer forecast models simultaneously;

5) data fusion is carried out in predicting the outcome of three kinds of monomer forecast models will selecting;

6) preserve the predicted data after step 5) merges;

7) judge that highway detects traffic data and whether upgrades, as upgraded, read new highway detection traffic data, judge whether the data in current time slip-window lack, if having, carry out interpolation processing; ;

8) judge the current number of times of having been predicted, if the number of times of prediction does not reach m time, return to step 5), if the number of times of prediction reaches m time, return to step 4).

2. the freeway traffic flow amount fusion forecasting method with online self-tuning optimization ability as claimed in claim 1, it is characterized in that: described highway detection history traffic data is time, flow, speed and occupation rate.

3. the freeway traffic flow amount fusion forecasting method with online self-tuning optimization ability as claimed in claim 2, it is characterized in that: described step 3) specifically comprises the steps:

31), in time slip-window (t-m+1, t), check whether the data in this time period have disappearance, if just selected this segment data of disappearance is arranged;

32) set a time starting point, then the time field changed into to data field, as the interpolation horizontal ordinate, using flow, speed and occupation rate respectively as the interpolation ordinate;

33) according to the cubic spline interpolation interpolation.

4. the freeway traffic flow amount fusion forecasting method that there is as claimed any one in claims 1 to 3 the online self-tuning optimization ability, it is characterized in that: in described step 4), 4 kinds of monomer forecast models are respectively time series predicting model, Kalman prediction model, neural network prediction model and chaotic prediction model;

For time series predicting model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine optimum sample size;

For the training mechanism of Kalman prediction model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine optimum state vector dimension.

For the BP neural network prediction model, if its be the poorest monomer forecast model of precision of prediction, again train forecast model, determine new prediction model parameters.

For the chaotic prediction model, if it is the poorest monomer forecast model of precision of prediction, again train forecast model, determine optimum embedding dimension and time delay, meet the phase space of chaotic characteristic with structure.

5. the freeway traffic flow amount fusion forecasting method with online self-tuning optimization ability as claimed in claim 4, it is characterized in that: in step 5), the method that predicting the outcome of three kinds of monomer forecast models carried out to data fusion is as follows: with three kinds of monomer forecast models, predicted respectively, utilize error inverse proportion method to predict the outcome and give respectively weight, fusion is then predicted the outcome.

Images