CN103927872A - Method for predicting multi-period travel time distribution based on floating vehicle data - Google Patents

Abstract

The invention discloses a method for predicting multi-period travel time distribution based on floating vehicle data. According to the method, on the basis of a traditional KNN algorithm and through the adoption of the learning ability of a Bayesian model, the robustness and accuracy of travel time predicting are improved; through the effective fusion of history data and real-time data, the travel time average and variance prediction of the continuous multiple time ranges are achieved, and the confidence interval of the travel time floating is built. Compared with a traditional KNN model, the travel time distribution information for half an hour can be predicted effectively in various complex city networks, the result is more reliable and precise.

Description

A kind of method of predicting multi-period journey time distribution based on floating car data

Technical field

The invention belongs to traffic travel time estimation and prediction field, a kind of method of predicting multi-period journey time distribution based on floating car data, relate in particular to a kind of method of predicting multi-period journey time distribution based on low sample frequency floating car data, belong to the multi-period journey time distribution forecasting method of a kind of improvement KNN with learning ability.

Background technology

Along with the quickening of Urbanization in China and the raising of automobile pollution, urban traffic blocking increasingly sharpens, and advanced transportation information service systems (ATIS) is a kind of effective way of alleviating traffic congestion.As Fundamentals of ATIS application, travel time information has caused great concern accurately and reliably.As an important indicator weighing traffic, travel time information provides a kind of quantizating index of effective measurement transportation network running status for traffic administration person, meanwhile, the transport information basis of advanced transportation information service systems application and crucial especially accurately.The prediction of multi-period journey time not only can help traveler in many ways in trip, formulating rational plan of travel, reduces travel time and travelling delay, reduces energy consumption, can also improve the operational efficiency of traffic system, effectively alleviates traffic congestion.

The outer research about Forecasting of Travel Time of Present Domestic can be divided into two large classes: highway and urban traffic network.Research for expressway up stroke time prediction is enriched very much, yet is quite limited about the research of urban road network Forecasting of Travel Time.Than relatively simple traffic environment on expressway, the Forecasting of Travel Time of urban road has more challenge.In urban traffic network, journey time has great uncertainty, mainly by following reason, is caused: the undulatory property of (1) transport need; (2) traffic control, the impact of each Signal on Highway Cross; (3) impact of the traffic events such as traffic hazard and control; (4) in urban traffic network, car type is many, driving behavior differentiation.Due to the uncertainty of journey time, cause the journey time average journey time of substantial deviation reality often of prediction, lost the value of tour reference.Therefore, the journey time distribution forecasting method under necessary research and development urban environment, provides city journey time average and uncertainty.

In recent years, floating car technology becomes one of advanced technology means of obtaining Traffic Information, so-called Floating Car is generally to refer to vehicle GPS positioning equipment to be installed and have been exercised motorbus and the taxi on major urban arterial highway, can registration of vehicle position, and the information such as direction and speed.Different from traditional static collect means, GPS Floating Car can be round-the-clock, dynamically collects in real time Traffic Information.Although GPS Floating Car has low cost, easily installation and maintenance, the advantage such as wide coverage is not affected by environment.Floating car data likely provides high-quality real-time traffic monitoring and the information of management and mobility service in the situation that cost is relatively low.

Summary of the invention

The present invention, in order to solve journey time forecast of distribution problem in complicated urban network, has proposed a kind of method of predicting multi-period journey time distribution based on low sample frequency floating car data.

The technical solution adopted in the present invention is: a kind of method of predicting multi-period journey time distribution based on floating car data, it is characterized in that, and comprise the following steps:

Step 1: the signal lamp of take is divided section as sign, take and floating car data is divided as the time interval for N minute, wherein N≤5;

Step 2: by GPS sampled point is carried out to interpolation, estimate the journey time of each Floating Car on each section;

Step 3: utilize KNN to predict the average of multi-period Link Travel Time;

Step 4: utilize Bayesian learning to predict that the error of multi-period Link Travel Time distributes;

Step 5: generate Link Travel Time and distribute, and on this basis, predicted travel time is interval.

As preferably, the N=5 described in step 1.

As preferably, the average of the multi-period Link Travel Time of prediction described in step 3 is to utilize the average of the multi-period Link Travel Time of KNN model prediction.

As preferably, the average of utilizing the multi-period Link Travel Time of KNN model prediction described in step 3, KNN model when carrying out Single-step Prediction, the real time data R before the known t period _tand historical data the data of considering t-l time period to the t-1 time period during prediction, its implementation comprises following sub-step:

Step 3.1: judge whether it is Single-step Prediction, if so, calculating current state vector is real time data R _tand historical data between Euclidean distance then continue step 3.2; Otherwise, jump to step 3.4;

Step 3.2: according to Euclidean distance from historical data middle selection k-neighbour, then the historic state amount of the t time period journey time of definite prediction;

Step 3.3: using the inverse of Euclidean distance as weight, by k-neighbour weight estimation t period journey time average

Step 3.4: make t=t+1, by the predicted value of t period upgrade real time data R _tthereby, real time data state vector and historical data are upgraded, obtain R _t+1and H _t+1, judgement:

If t<t+p, returns to step 3.1, step 3.2, step 3.3 and step 3.4, carry out multi-period journey time mean prediction, by k-neighbour weight estimation t+1 time period journey time average

If t=t+p, prediction finishes, and p represents the time period number of prediction continuously.

As preferably, the error of the multi-period Link Travel Time of prediction described in step 4 distributes, be on the basis of the multi-period journey time average of traditional KNN model prediction, utilize Bayesian model to make improvements, predict that the error of multi-period Link Travel Time distributes.

As preferably, the Bayesian model of utilizing described in step 4 is predicted the error distribution ε of multi-period Link Travel Time _t, first hypothesis at one time section in, the predicated error x of KNN model _tnormal Distribution, two location parameter μ _twith all unknown and connect each other, and Normal Distribution and the distribution of falling gamma respectively, the joint distribution of the two is normal state-distribution of falling gamma, its implementation comprises following sub-step:

The degree of freedom v of step 4.1. initialization sample ₀, journey time estimation average m ₀initial value c with the variance of journey time ₀;

Step 4.2. prior estimate: utilize the posteriority of current t time period to distribute to estimate the prior distribution of t+1 time period;

Step 4.3. prediction: for the t+1 time period, utilize the average of prior distribution and the error mean f that variance information is predicted the t+1 time period _t+1, error mean variance C _t+1variance c with error _t+1;

Step 4.4. upgrades: As time goes on, if obtain the real-time sample of t time period, utilize it to prediction distribution, to carry out posteriority renewal; Otherwise the sample information using the posteriority lastest imformation of a upper period as the current period, carries out posteriority renewal; Step 4.5. makes t=t+1, if t<t+p+1 returns to step 4.2, carries out multi-period journey time mean prediction; If t=t+p+1, prediction finishes, and wherein p represents the time period number of prediction continuously.

The forecasting research that the present invention is directed to that existing Travel Time Estimation Method majority is only estimated Link Travel Time average, ignored the uncertainty of journey time in city road network, the research expressway up stroke time average of most limitations and distribution, city journey time distributes concentrates on the deficiency of single period, the method that has proposed to utilize Bayes to improve traditional KNN solves the problem of multi-period journey time forecast of distribution in crowded urban road network, when having improved the robustness that predicts the outcome, also guaranteed the reliability predicting the outcome.

Accompanying drawing explanation

Fig. 1: be the process flow diagram of the embodiment of the present invention.

Embodiment

What the present invention will solve is the forecasting problem that in complicated urban network, journey time distributes, and for the Floating Car gps data of low sample frequency, has proposed the multi-period journey time distribution forecasting method of a kind of improvement KNN with learning ability.

For the ease of those of ordinary skills, understand and enforcement the present invention, below in conjunction with drawings and Examples, the present invention is described in further detail, should be appreciated that exemplifying embodiment described herein, only for description and interpretation the present invention, is not intended to limit the present invention.

Input data of the present invention are Floating Car GPS tracing point and the road net data after map match, GPS track data is that a series of GPS points according to time sequencing sequence form, each GPS point is comprised of speed, time point and position etc., traffic network is comprised of a series of oriented sections and node, and there is space-time dynamic characteristic, if G=(N, A, Ω) represents a traffic network, wherein, N represents set of node, and A represents section collection, Ω represent a series of discrete time section △ ..., n △, ..., and n is integer, △ represents the time interval (the present embodiment is got 5min).Here section a _ij∈ A, take in city road network signal lamp as according to dividing.

Ask for an interview Fig. 1, the technical solution adopted in the present invention is: a kind of method of predicting multi-period journey time distribution based on floating car data, it is characterized in that, and comprise the following steps:

Step 1: the signal lamp of take is divided section as sign, take and floating car data is divided as the time interval for 5 minutes;

Step 2: by GPS sampled point is carried out to interpolation, estimate the journey time of each Floating Car on each section;

While sampling due to GPS, most Floating Car can not be passed through crossing just, thereby most of GPS sampled point drops on section.Therefore, embodiment goes out vehicle by the moment of two ends, section (being intersection) by Interpolate estimation, thereby estimates the Link Travel Time of each Floating Car on each section.

Step 3: the average of utilizing the multi-period Link Travel Time of KNN model prediction;

Example utilizes historical data and real time data to predict, the real time data R before the known t period _t=[r _t-l..., r _t-2, r _t-1] and historical data $H_t=\left[\begin{array}{cccc}{h}_{t-l}^1,& \&CenterDot;\&CenterDot;\&CenterDot;,& h_{t-2}^1,& h_{t-1}^1\\ h_{t-l}^2,& \&CenterDot;\&CenterDot;\&CenterDot;,& h_{t-2}^2,& h_{t-1}^2\\ \&CenterDot;& & \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;,& \&CenterDot;& \&CenterDot;\\ \&CenterDot;,& & \&CenterDot;,& \&CenterDot;\\ h_{t-l}^d,& \&CenterDot;\&CenterDot;\&CenterDot;,& h_{t-2}^d,& h_{t-1}^d\end{array}\right]$ (during prediction, consider the data of t-l time period to t-1 time period, get l=2 here, d represents the number of days of historical data), its implementation comprises following sub-step,

Step 3.1, if Single-step Prediction calculates real time data R _tand historical data between Euclidean distance can be expressed as: ${\mathrm{ED}}_i^{\mathrm{RH}}=\sqrt{{(r_{t-l}-h_{t-l}^i)}^2+\&CenterDot;\&CenterDot;\&CenterDot;+{(r_{t-1}-h_{t-1}^i)}^2},$ Continue step 2.2; Otherwise, jump to step 2.4;

Step 3.2, according to Euclidean distance from historical data middle selection k-neighbour (in this example, k=2), then can determine the historic state amount of the t time period journey time of prediction

Step 3.3, usings the inverse of Euclidean distance as weight, by k-neighbour weight estimation t period journey time average computing formula is $\widehat{m_t}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{h}_t^i/{\mathrm{ED}}_i^{\mathrm{RH}}}{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}1/{\mathrm{ED}}_i^{\mathrm{RH}}};$

Step 3.4, makes t=t+1, by the predicted value of t period real time data R in step of updating 2 _tthereby, real time data state vector and historical data matrix are upgraded, obtain current state vector and historical data $H_{t+1}=\left[\begin{array}{cccc}{h}_{t-(l-1)}^1,& \&CenterDot;\&CenterDot;\&CenterDot;,& h_{t-1}^1,& h^t\\ h_{t-(l-1)}^2,& \&CenterDot;\&CenterDot;\&CenterDot;,& h_{t-1}^2,& h_t^2\\ \&CenterDot;& & \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;,& \&CenterDot;& \&CenterDot;\\ \&CenterDot;,& & \&CenterDot;,& \&CenterDot;\\ h_{t-(l-1)}^d,& \&CenterDot;\&CenterDot;\&CenterDot;,& h_{t-1}^d,& h_t^d\end{array}\right],$ Judgement:

If t<t+p (predicts the journey time of continuous 6 time periods in this example, be p=6), return to step 3.1, step 3.2, step 3.3 and step 3.4, carry out multi-period journey time mean prediction, by k-neighbour weight estimation t+1 time period journey time average

If t=t+p, prediction finishes, and p represents the time period number of prediction continuously.

Step 4: on the basis of the multi-period journey time average of traditional KNN model prediction, utilize Bayesian model to make improvements, predict that the error of multi-period Link Travel Time distributes;

Because KNN is in the insecure situation of historical data, predict the outcome unreliable, robustness is poor.Therefore,, after KNN predicted travel time average, it is necessary utilizing Bayes to carry out predicated error distribution.

Normal average dynamic linear models DLM{1,1, V _t, W _tthe conventional a kind of model of Bayesian forecasting, due to W _tgenerally be difficult for obtaining, thus discount factor δ in model, introduced, common 0< δ <1, and have this shows that priori precision equals the posteriority precision of discount.This normal average dynamic model with discount factor, is called normal average discount Bayesian model, is designated as DBM{1,1, V, δ }.

This example hypothesis is at one time in section, the predicated error x of KNN _tnormal Distribution, two location parameter μ _twith all unknown and connect each other, and Normal Distribution and the distribution of falling gamma respectively, be respectively $\left({\mathrm{\&mu;}}_t\right|{\mathrm{\&sigma;}}_t^2)~N[m_t,{\mathrm{\&sigma;}}_t^2/k_t],{\mathrm{\&sigma;}}_t^2~\mathrm{IGa}[v_t/2,v_t\&CenterDot;c_t/2],$ The joint distribution of the two is normal state-distribution of falling gamma, is, wherein, m _t, c _t, v _tthe super parameter of this distribution, m _tfor average, c _tfor estimated value, v _tfor sample degree of freedom.

Bayesian forecasting comprises two basic equations, i.e. state equation μ _t=μ _t-1+ w _twith observation equation Y _t=μ _t+ v _t.Its implementation comprises following sub-step,

Step 4.1, initialization: the initial information of model is v wherein ₀the initial degree of freedom that represents sample; m ₀the initial estimation average that represents Forecasting of Travel Time error, c ₀the initial value that represents the variance of Forecasting of Travel Time error, D ₀represent initial information collection.

Step 4.2, prior estimate: utilize the posteriority of current t time period to distribute to estimate the prior distribution of t+1 time period, the prior distribution of error mean and variance is respectively with

wherein, R _t+1=C _t+ W _t+1, c _tthe variance that represents predicated error, c _t=k _tr _t+1(v _t-2)/v _t.

Step 4.3, prediction: for the t+1 time period, can utilize the information such as the average of prior distribution and variance can be used for predicting the error mean f of t+1 time period _t+1, the variance C of error mean _t+1variance c with error _t+1, have

$\left(Y_{t+1}\right|{\mathrm{\&sigma;}}_t^2,D_t)~N[f_{t+1},Q_{t+1}],\left({\mathrm{\&sigma;}}_{t+1}^2\right|D_t)~\mathrm{IGa}[v_t/2,v_t\&CenterDot;c_t/2],$ C _t=k _tq _t+1(v _t-2)/v _t, wherein, f _t+1=m _t, Q _t+1=R _t+1+ V _t+1.

Step 4.4, upgrades: As time goes on, if obtain the real-time sample of t+1 period, utilize it to prediction distribution, to carry out posteriority renewal; Otherwise the sample information using the posteriority lastest imformation of a upper period as the current period, carries out posteriority renewal; Make e _t+1=Y _t+1-f _t+1, A _t+1=R _t+1/ Q _t+1(A _t+1(0<A _t+1<1) be adaptive adjustment coefficient.When prior imformation is more reliable, A _t+1→ 0; And when observed reading is more important, A _t+1→ 1), have

$\left({\mathrm{\&mu;}}_{t+1}\right|{\mathrm{\&sigma;}}_{t+1}^2,D_{t+1})~N[m_{t+1},C_{t+1}]$ With $\left({\mathrm{\&sigma;}}_{t+1}^2\right|D_{t+1})~\mathrm{IGa}[v_{t+1}/2,v_{t+1}\&CenterDot;c_{t+1}/2],$

$c_{t+1}=(v_t\&CenterDot;c_t+(n_{t+1}-1)\&CenterDot;S_{t+1}^2+\frac{k_t\&CenterDot;n_{t+1}}{k_t+n_{t+1}}\&CenterDot;{e_{t+1}}^2)/v_{t+1},$

Wherein, m _t+1=m _t+ A _t+1e _t+1, C _t+1=A _t+1v _t+1, k _t+1=k _t+ 1, variance for t+1 time period sample.

Step 4.5, makes t=t+1, if t<t+p+1 returns to step 4.2, carries out the prediction of multi-period error distribution; If t=t+p+1, prediction finishes.

For multi-step prediction, because t+1 not yet obtains real time data to the t+p period, therefore when posteriority upgrades, the sample information using the posteriority lastest imformation of a upper period as the current period herein, do is like this that the transport information of adjacent time interval is more or less the same because seasonal effect in time series one step autocorrelation is very high.

Step 5: generate Link Travel Time and distribute, and on this basis, predicted travel time is interval.

Due to predicated error Normal Distribution, according to formula can know also Normal Distribution of corresponding journey time, can be expressed as can provide rational fiducial interval according to its average and variance.

Should be understood that; the above-mentioned description for preferred embodiment is comparatively detailed; can not therefore think the restriction to scope of patent protection of the present invention; those of ordinary skill in the art is under enlightenment of the present invention; do not departing from the scope situation that the claims in the present invention protect; can also make and replacing or distortion, within all falling into protection scope of the present invention, the scope of asking for protection of the present invention should be as the criterion with claims.

Claims (5)

1. based on floating car data, predict the method that multi-period journey time distributes, it is characterized in that, comprise the following steps:

Step 1: the signal lamp of take is divided section as sign, take and floating car data is divided as the time interval for N minute, wherein N≤5;

Step 2: by GPS sampled point is carried out to interpolation, estimate the journey time of each Floating Car on each section;

Step 3: utilize KNN to predict the average of multi-period Link Travel Time;

Step 4: utilize Bayesian learning to predict that the error of multi-period Link Travel Time distributes;

Step 5: generate Link Travel Time and distribute, and on this basis, predicted travel time is interval.

2. the method for predicting multi-period journey time distribution based on floating car data according to claim 1, is characterized in that: the average of the multi-period Link Travel Time of prediction described in step 3 is to utilize the average of the multi-period Link Travel Time of KNN model prediction.

3. according to claim 2ly based on floating car data, predict the method that multi-period journey time distributes, it is characterized in that: the average of utilizing the multi-period Link Travel Time of KNN model prediction described in step 3, KNN model when carrying out Single-step Prediction, the real time data R before the known t period _tand historical data the data of considering t-l time period to the t-1 time period during prediction, its implementation comprises following sub-step:

Step 3.1: judge whether it is Single-step Prediction, if so, calculating current state vector is real time data R _tand historical data between Euclidean distance then continue step 3.2; Otherwise, jump to step 3.4;

Step 3.2: according to Euclidean distance from historical data middle selection k-neighbour, then the historic state amount of the t time period journey time of definite prediction;

Step 3.3: using the inverse of Euclidean distance as weight, by k-neighbour weight estimation t period journey time average step 3.4: make t=t+1, by the predicted value of t period upgrade real time data R _tthereby, real time data state vector and historical data are upgraded, obtain R _t+1and H _t+1, judgement:

If t<t+p, returns to step 3.1, step 3.2, step 3.3 and step 3.4, carry out multi-period journey time mean prediction, by k-neighbour weight estimation t+1 time period journey time average

If t=t+p, prediction finishes, and p represents the time period number of prediction continuously.

4. according to claim 2ly based on floating car data, predict the method that multi-period journey time distributes, it is characterized in that: the error of the multi-period Link Travel Time of prediction described in step 4 distributes, on the basis of the multi-period journey time average of traditional KNN model prediction, utilize Bayesian model to make improvements, predict that the error of multi-period Link Travel Time distributes.

5. according to claim 4ly based on floating car data, predict the method that multi-period journey time distributes, it is characterized in that:

The Bayesian model of utilizing described in step 4 is predicted the error distribution ε of multi-period Link Travel Time _t, first hypothesis at one time section in, the predicated error x of KNN model _tnormal Distribution, two location parameter μ _twith all unknown and connect each other, and Normal Distribution and the distribution of falling gamma respectively, the joint distribution of the two is normal state-distribution of falling gamma, its implementation comprises following sub-step:

The degree of freedom v of step 4.1. initialization sample ₀, journey time estimation average m ₀initial value c with the variance of journey time ₀;

Step 4.2. prior estimate: utilize the posteriority of current t time period to distribute to estimate the prior distribution of t+1 time period;

Step 4.3. prediction: for the t+1 time period, utilize the average of prior distribution and the error mean f that variance information is predicted the t+1 time period _t+1, error mean variance C _t+1variance c with error _t+1;

Step 4.4. upgrades: As time goes on, if obtain the real-time sample of t time period, utilize it to prediction distribution, to carry out posteriority renewal; Otherwise the sample information using the posteriority lastest imformation of a upper period as the current period, carries out posteriority renewal; Step 4.5. makes t=t+1, if t<t+p+1 returns to step 4.2, carries out multi-period journey time mean prediction; If t=t+p+1, prediction finishes, and wherein p represents the time period number of prediction continuously.