SHORT-TERM TRAVEL TIME PREDICTION USING A NEURAL NETWORK

Transcription

1 SHORT-TERM TRAVEL TIME PREDICTION USING A NEURAL NETWORK Giovai Huiske ad Eric va Berkum Dept. of Civil Egieerig - Uiversity of Twete AE Eschede - The Netherlads g.huiske@ctw.utwete.l ad e.c.vaberkum@ctw.utwete.l 1 INTRODUCTION The growth i car mobility has lead to more ucertaity i travel times. As a result cardrivers have a icreasig demad for iformatio o these travel times (see e.g. Huiske ad Va Berkum, 2000). Travel times ca be measured usig automated vehicle idetificatio (AVI) techiques (floatig car data, automated licese plate recogitio, etc.). However, these techiques are rarely used sice they eed large ivestmets i roadside equipmet. Curretly travel times are estimated usig data from iductive loop detectors. Sice loop detectors yield spot measuremets of flow ad speed, travel times ca oly be estimated, ot actually measured. Furthermore, whe cardrivers are provided with iformatio o travel times, these travel times should ideally be the times that they will ecouter. Therefore we eed to predict travel times, based o previous measuremets. Curretly two methods are beig used, i.e. Static Travel Time Estimatios (STTE) ad Dyamic Travel Time Estimatios (DTTE) (see e.g. Housell ad Ishtiaq, 1997; Va Arem et al, 1997; ad Zee, 2001). This research proposes a ew travel time predictio method usig a Artificial Neural Network (ANN). The three methods STTE, DTTE, ad ANN methods were applied o the A13 motorway from The Hague to Rotterdam ad their performace was compared. 2 DATA ACQUISITION This sectio is about the data acquisitio site ad the data itself. A overview of the geographical site is give alog with a quatitative descriptio of the data sets. 2.1 The geographical site The chose geographical locatio (figure 1) is the motorway A13 from The Hague to Rotterdam oe directio oly. This motorway has a legth of 11.4 km, 5 o- ad offramps, a speed limit of 100 km/h, ad has oe gas statio (approximately halfway). The sectio of the motorway possesses 21 locatios where dual iductio loop detectors are situated.

2 Figure 1. The motorway sectio from The Hague to Rotterdam. 2.2 Data sets descriptio I this cotributio two mai issues will be addressed. First a method is selected that yields the best travel time estimate usig iductive loop data. Secod, three methods to predict travel times are beig compared. For the first part two data sets have bee acquired: oe licece plate recogitio set (Delft-Noord/Nootdorp [km 7.3 post] Rotterdam-Overschie [km post]) ad oe through the dual iductio loop detectors with the MARI [More Applicatie Routekeuze Iformatie] system. The licece plate recogitio set was acquired by time ad licece plate registratio of passig red vehicles at the startig ad re-idetificatio poit ad subsequet subtractio produced travel times. The collectio took place o October 8 th ad 9 th 1996: 07:00 09:30 ad 15:30 18:30 ad o October 11 th 1996: 15:00 18:45. Durig this period, also iductive loop data for the 21 locatios were collected, cotaiig flow ad speed o a oe-miute basis. For the secod part, i.e. the compariso of the three predictio methods, loop data was collected from November 11 th 1996 February 23 rd The set also cotais 1-miute aggregated data from loop detectors cotaiig iformatio o flow ad speed. 3 METHODS AND MODEL DEVELOPMENT 3.1 Estimatio of travel times usig iductive loop data From Bovy ad Thijs (2000) five algorithms (RT0 - RT4) were selected ad tested for the estimatio of travel times usig loop detector data (table 1). The estimates these methods yield were compared with the actual measured travel times usig licese plate recogitio RT_M_AVG (figure 2). Table 1. Travel time estimatio methods. Lik method Route method Smoothig Speed Mass balace Static Dyamic Iput Output RT0 x x x x x RT1 x x x x RT2 x x x RT3 x x x RT4 x x

3 RT4 9 RT3 Travel time [m] RT2 RT1 RT0 RT_M _A V G Time of day Figure 2. Estimated travel times by method. RT4 gave the best results (a average error of 23.0 secods o a mea travel time of 8 miutes ad 46 secods after shiftig). Sice for the traiig of the eural etwork, as well as for the compariso of the three predictio methods oly iductive loop data are available ad o actual travel time measuremets, RT4 is applied to costruct a dataset of travel times. 3.2 Predictio usig static travel time estimatio STTE uses last-kow lik travel times ad sums them. Liks (L) are defied as the distace betwee two cosecutive dual iductio loops. Suppose a vehicle eters a specific route (umber of liks = k) at time t = T 0. The most recet recorded loop speeds v (T 0 ), ( = 1..k), are assiged to half of the lik before ad after the specific loop locatio. STTE becomes: STTE T0 = L k 1 2 L 1 2 v 1 ( T 0 ) + L +1 L 1 + L k L k 1 () 1 = 2 2 v ( T 0 ) 2 v k ( T 0 ) So, lik travel times are assumed fixed as of the momet that the vehicle eters the route (T 0 ). 3.3 Predictio usig dyamic travel time estimatio DTTE at time T 0 ca oly be estimated by historical recostructio ad is doe iteratively. Suppose at T 0 vehicle 2 eters the motorway ad vehicle 1 is the last vehicle that left the motorway. From the recorded speed data the travel time that vehicle 1 did ecouter ca be recostructed. Now DTTE uses this travel time as a predictio for vehicle 2.

4 3.4 Predictio usig artificial eural etworks ANNs are based upo biological eural etworks by mimickig their architectural structure ad iformatio processig i a simplified maer. They both cosist of processig elemets called euros that are highly itercoected makig the etworks parallel iformatio processig systems. They are capable of tasks such as patter recogitio, perceptio ad motor cotrol that are cosidered poorly performed by covetioal liear processig. These parallel systems are also kow to be robust ad to have the capability to capture highly o-liear mappigs betwee iput ad output. ANN applicatios i trasport ca be foud i e.g. Dougherty (1995) ad Huiske (1998 ad 2001). Here Multi Layer Feedforward (MLF) eural etworks were used. The MLF is a supervised learig etwork meaig that durig the traiig phase all iputs are mapped o desired outputs. The error, i.e. the differece betwee the actual ad the desired output, is a criterio that is used to adjust the weights of the euros iteratively so that the total error of all iputoutput pairs is miimised. The algorithm resposible for this method is called the learig rule. More comprehesive iformatio o MLF eural etworks ca be foud i ay textbook o ANNs. MLF etworks will be traied with the most accurate travel times as targets. The umber of iput variables is: 21 (iductio loops) * 2 (quatities: speed ad itesity) + 1 (time of day) = 43. This high umber of iputs resulted i time cosumig traiig phases so preprocessig was used to cut the iput dow to 12 iputs. The data set was divided ito 3 equally sized subsets (for cross-validatio purposes) where oe subset was used as test set to fid the optimum umber of hidde euros ad epochs (traiig cycles) ad to prevet the MLF etwork from overfittig. 4 RESULTS & CONCLUSIONS The predictios of the three methods were compared with the travel times determied by RT4. Travel time predictio becomes iterestig whe free flow coditios o loger hold. Therefore predictio was oly executed whe travel times exceeded 500 secods (i.e. average speed uder 82 km/h). To compare the method s performaces several measures of error were determied (see formula 2): the Mea Relative Error (MRE) [%], the Mea Absolute Relative Error (MARE) [%], the Mea Time Error (MTE) [secods], the Mea Absolute Time Error (MATE) [secods], ad R-squared. MRE = t ˆ t MARE = ˆ t t ˆ t MTE = ˆ t ˆ t t MATE = t ˆ t (2) where is the umber of cases, t ˆ is the travel time to be predicted (target value from RT4), ad t is the travel time geerated by the model.

5 The results are give i table 2 that shows that MLF sigificatly outperforms DTTE, which i tur sigificatly outperforms STTE. The MRE values are also classified ito 5%-error itervals ad from this ca be cocludes that roughly two-thirds, a half, ad oe-third of the predictio cases fall i the [-5%, 5%] error domai (betwee the 2 vertical lies) for MLF, DTTE, ad STTE, respectively (figure 4). Table 2. MRE, MARE, MTE, MATE ad R-squared results. MRE [%] MARE [%] MTE [sec] MATE [sec] R-squared STTE DTTE MLF STTE DTTE MLF Figure 4. Distributio of MRE. REFERENCES Bovy P.H.L. ad Thijs R. (2000). Estimators of travel time for road etworks, ew developmets, evaluatio results, ad applicatios, Delft Uiversity of Techology. Dougherty M.S. (1995). A review of eural etworks applied to trasport. Trasp. Res.-C, 3(4), pp Housell N.B. ad Ishtiaq S. (1997). Jourey time forecastig for dyamic route guidace systems i icidet coditios. It. J. Forecastig, 13(1), pp

6 Huiske G. (1998). Literature review: eural etwork applicatios i traffic ad trasport. Research report: Uiversity of Twete, The Netherlads. Huiske G. (2001). Short-term forecastig of traffic flow o freeways. Proc. of the 9 th World Coferece o Trasport Research, July 2001, Seoul, Korea. Huiske G. ad Va Berkum E.C. (2000). DAB i the Netherlads? Proc. of the 8 th Meetig of the Euro Workig Group Trasportatio EWGT, September 2000, Rome, Italy. Va Arem B., Va Der Vlist M.J.M., Muste M. ad Smulders S.A. (1997). Travel time estimatio i the GERDIEN project. It. J. Forecastig, 13(1), pp Zee J.C. (2001). Oude reistijde actueel. M.Sc. report, Uiversity of Twete, Eschede, The Netherlads [I Dutch].