Predicting Freeway Travelling Time Using Multiple- Source Data

Transcription

1 Artcle Predctng Freeway Travellng Tme Usng Multple- Source Data Kejun Long 1, Wuka Yao 2, Jan Gu 1 *and We Wu 1 1 Hunan Provncal Key Laboratory of Smart Roadway and Cooperatve Vehcle-Infrastructure Systems, Changsha Unversty of Scence & Technology, Changsha , Chna; longkejun@csust.edu.cn; jaotongwewu@csust.edu.cn; gujan87@126.com 2 School of Traffc and Transportaton Engneerng, Changsha Unversty of Scence & Technology, Changsha , Chna; @qq.com * Correspondence : gujan87@126.com; Tel.: Abstract: Freeway travellng tme s affected by many factors ncludng traffc volume, adverse weather, accdent, traffc control and so on. We employ the multple source data-mnng method to analyze freeway travellng tme. We collected toll data, weather data, traffc accdent dsposal logs and other hstorcal data of freeway G5513 n Hunan provnce, Chna. Usng Support Vector Machne (SVM), we proposed the travellng tme model based on these databases. The new SVM model can smulate the nonlnear relatonshp between travellng tme and those factors. In order to mprove the precson of the SVM model, we appled Artfcal Fsh Swarm algorthm to optmze the SVM model parameters, whch nclude the kernel parameter σ, non-senstve loss functon parameter ε, and penalty parameter C. We compared the new optmzed SVM model wth Back Propagaton (BP) neural network and common SVM model, usng the hstorcal data collected from freeway G5513. The results show that the accuracy of the optmzed SVM model s 17.27% and 16.44% hgher than those of the BP neural network model and the common SVM model respectvely. Keywords: Support Vector Machne; Travellng tme; Intellgent Transportaton System; Artfcal Fsh Swarm algorthm; Bg data. 1. Introducton Travel tme s one of the man ndexes that reflect the traffc operaton level of a freeway, and t s also the bass for Advanced Traveler Informaton System (ATIS), Traffc Gudance System(TGS), and Traffc Control System(TCS). The challenges and dffcultes of travel tme predcton are dentfed below. Dverse nfluencng factors such as weather, holdays, traffc accdents, out of sample predcton, and mechansms contrbutng to congeston. It s dffcult to descrbe and predct the nfluence mechansm by usng tradtonal conventonal mathematcal models. The complexty and ncompleteness of basc data. Although there are many flow detectors and vdeo detecton equpment on the freeway, captured data are ncompatble, redundant, and nclude error or loss. To avod these, technques whch utlze mult-source data to mprove the accuracy of travel tme predcton s extremely mportant. In Chna, practcal applcaton of travel tme predcton focuses manly on the followng two aspects. The frst aspect s the predcton of travel tme by map navgaton provders usng ther personalzed GPS data. Many map servce provders employ ther personalzed data for travel tme forecast servces and commercal products. For nstance, Ba-du, Gao-de, and other Chnese map provders collect real-tme GPS data from users whle provdng map navgaton servces. Then, a 2018 by the author(s). Dstrbuted under a Creatve Commons CC BY lcense.

2 2 of correlaton algorthm s proposed to obtan the travel tme predcton result at road sectons, whch depends on the market share of the map navgaton servce. The hgher frequency of people usng the navgaton servce, the more complete of GPS data and the hgher the predcton accuracy wll be. However, accordng to the Chnese market report, the market share of Ba-du and Gao-de servces are presently 29.3% and 32.6%respectvely. Therefore, the accuracy of results should be further mproved by ncreasng market share. The second aspect s the predcton based on the traffc detecton data of urban traffc managers and hstorcal data. In recent years, numerous fxed detector devces have been nstalled n most of urban roads and rural freeways for the predcton of travel tme, ncludng nductve loops, vdeo recorder, mcrowave, and laser detecton. However, unavodable damage to flow detecton equpment and transmsson error of partal data make traffc detecton data ncomplete redundant or error. In addton, dfferent detectors have dfferent data formats and data accuracy. Wth the rough use of mstake data for precse travel-tme predcton, the Traveler Informaton Servce system cannot recommend optmal travel routes or warn of potental traffc congeston and users cannot determne the optmal departure tme or estmate ther expected arrval tme based on predcted travel tmes. Theoretcal research on freeway travel tme predcton can be dvded nto two categores based on sngle source data and mult-source data. 1.1 Overvew ofpredcton Method of Sngle Source Data A sngle data source was earler method used for predct travel tme. Many researches predcton results were obtaned upon a sngle data source. Gpps, P. G. [1]used the occupancy and arrval tme to predct the travel tme n a road n Mehmet Y, Nkolas G [2]set statstcal predctve algorthms to predct the future travel tme. Shen, L., &Had, M. [3]employed data obtaned from detector n freeway. Kyung et al. [4,5]used nductve loop detectors to obtan the front poston and capture the nteractons between trucks and non-trucks. But fxed detector devces s easly affected by external envronment and cannot drectly access some mportant parameters, such as travel tme, etc. In addton,many researches consder usng GPS data to predct travel tme. Ramezan et al. [6] and Zhang et al. [7,8]consdered the dversty of GPS data and nvestgated the applcaton of Markov chan to travel tme estmaton and mplemented good predcton accuracy. Woodard, Nogn &Paul et al. [9]used the GPS data of the current hghest volume GPS data source, and appled the TRIP method to predct the travel tme. Based on GPS data sets, Bahuleyan, H., &Vanajaksh, L. D. [10] proposed a predcton method for urban trunk lnes whch was only sutable for traffc condtons n Inda. But the GPS data only gets the speed and real-tme poston nformaton and uncertanty due to the route of the vehcle, t affects the coverage and accuracy of detecton data. Above methods ndeed are nnovatons and mprovements n travel tme predcton, and results are more accurate. However, many predctons that use a sngle data source do not consder the mpact of other unexpected events or the result was not accurate enough because sngle data source cannot reflect traffc state of road network exactly. It wll result n certan errors between predcton results and true values. 1.2Overvew ofpredcton Method of Multple Source Data Nowadays, the development of traffc bg data envronment has progressed rapdly. Wth the support of a large amount of data, t s possble to clearly vsualze traffc flow changes under the jont acton of dfferent factors, that s, the traffc state presented, whch s more favorable. The constructon of the predctve model mproves the adaptablty and accuracy of the model,[11] and f the same state occurs, t can be predcted based on hstorcal results. The more populated the database s, the hgher the qualty and the hgher the lkelhood of fndng commonaltes and predctng accurate results wll be. Ths concept can be appled by searchng for common traffc states for predcton.

3 3 of Owng to progress n the dynamc traffc nformaton acquston system, varous traffc data can be collected more easly. And data fuson s fnshed n the dynamc traffc nformaton acquston system, whch s jontly determned by the advantages of mult-source data and the characterstcs of traffc condtons. And usng mult-source data for predcton can overcome the lmtatons of the sngle data source. In other words sngle data source has low qualty and s not comprehensve. The traffc state s descrbed from dfferent angles and drectons to mprove the accuracy of predcton and reduce dsturbance from unexpected factors. At present, many studes have been conducted on travel tme predcton, especally studes based on the hstorcal data travel tme of mult-source data. The common predctng methods and ther characters are summarzed n the table below. Table 1. Comparson of common predcton methods Predcton method Kalman flter Bayesan estmaton Statstcal decson theory Neural network Author Ln J W C V.2006[12] Zhou, J. 2014[13] Tang-Hsen Chang.2016[14] Fe, X.2011[15] Zhan, X. 2016[16] Wosyka,J.2012[17] Innamaa, S.2005[18] Ln J W C V.2005[19] Data source Travel tme data Floatng vehcle and fxed detector data Electronc Toll Collecton (ETC) and tradtonal Vehcle Detector data The real loop detector data of an I-66 segment n Northern Vrgna A large-scale tax trp dataset from New York Cty Two detectors datan Prague and also n the Czech Republc. travel tme data Travel data and Some mssng or corrupt travel data Compared to the sngle source data, the mult-source data method can extract deep nformaton wthn data, sgnfcantly reduce the cost of data acquston, and make up for lack of nformaton and packet loss of sngle source data. At present, bg data applcaton technology are wdely used n the traffc feld, many studes have been conducted n the feld of freeway travel tme predcton based on bg data analytcs. However, there are several defcences ncludng: The method pays attenton to machne learnng algorthms and lacks the mastery of the characterstcs of the traffc flow, resultng n the uncoordnated and unsutable correspondence between the data and the traffc flow. Wth the contnuous updatng of bg data, t provdes condtons for traffc travel tme predcton, but some advantages and characterstcs of these data are not notced, resultng n many useful data not beng used and mned Some model parameter calbraton s too subjectve, whch largely depend on researchers' experence; Some model mostly amed at a specfc example, and cannot be easly adapted to other stuatons. Therefore, n ths study, hstorcal data of a freeway toll staton were collected, and were categorzed usng the support vector machne (SVM) algorthm. Although the predecessors have done some work: Wu, Chun-Hsn[20]used the method of support vector regresson to predct the tme; Vanajaksh[21] obtaned the support vector for short-term predcton of travel tme by algorthm. Machne technology; Mendes-Morera [22]obtaned a regresson method for comparng long-term travel tme predcton through ntellgent data analyss. But ther analyss s based on machne learnng algorthms and does not better understand or mprove the transportaton system. So ths paper uses SVM model based on hstorcal data to predct the common traffc state and the method used for model constructon was smplfed. The practcablty of the predcton algorthm was enhanced to overcome assumptons and uncertantes n the exstng traffc flow theory.

4 Preprnts ( NOT PEER-REVIEWED Posted: 25 October of Data Collecton and Preprocessng Data Descrpton Data for ths study was collected n the FreewayG5513 from Changsha to Yyang, Hunan Provnce, startng from the Changsha toll staton and endng at Yyang toll staton. Ths freeway G5513 s a standard freeway wth a two-way four lane of 100 km/h and a roadbed wdth of 26 m. The total length was approxmately 63 km, the daly average flow reached 58,000 vehcles, and the peak flow durng long vacaton was up to 96,600 vehcles. Because of heavy traffc, the freeway has been rated as one of the sx most congested sectons n the Hunan Provnce. There are nne toll statons along the road, from east to west, that s Changsha West, Youren, Guanshan, Jnzhou, Nngxang, Xangjang West, Quanjao, Chaoyang, and Yyang North, llustrated as Fgure Fgure 1. Layout of the freeway and toll staton The man data set collected n ths study ncludes: Toll data of the whole toll statons along G5513 n February 2018 (vehcles enterng and leavng toll staton), wth a total of 561,081 data tems, ncludng the name of the toll staton, the tme of vehcle enterng and leavng the toll staton, vehcle type and weght. Weather nformaton surveyng staton located near the freeway, whch was collected from the Chnese Weather Network n February 2018, wth a total of 672 data tems, ncludng 24 hour daly weather, temperature, relatve humdty, precptaton, and wnd drecton. Freeway blockage record statstcs, whch was obtaned from the freeways management department, a total of 260 freeway blockage nformaton reports were collected n February, March, Aprl, and May, ncludng blockage locaton, reasons for the blockage, blockage start tme, and blockage end tme. Freeway traffc control measures report, whch was obtaned from the Traffc Polce Department, wth a total of 7 data tems collected on Aprl 5 Qngmng tradtonal natonal Festval, May 1 Internatonal Labor Day, and other holday control nformaton Data Preprocessng Many abnormal data tems were found n the data, whch need to be preprocessed before use. Data sharng the same entry and ext toll

5 5 of On the freeway, some drvers turn around n the servce area or other sectons to avod the charges and even exchange the toll tags, whch s lkely to make the entry and ext of the vehcles at toll statons consstent. Therefore, t s necessary to determne whether a data tem s consstent wth the toll gates and elmnate nvald data tems. Abnormal tme record data Owng to the falure of the tme system assocated wth toll staton to synchronze or system falure, the tme of accessng the toll staton can be earler than the tme of extng the toll staton. In addton, there are other factors that can lead to long travel tme, such as the breakdown of vehcle on road, accdents, and the stuaton where drver may have a long rest n servce area. All of these stuatons wll result n unusual tme record data. In the process of data preprocessng, abnormal tme data record can be elmnated by screenng. Mssng data There are two man reasons for mssng data: on the one hand, t s manly from equpment problems or road envronment ncludngthe unstable scannng frequency of detector, faulty of transmsson equpment, and traffc jam On the other hand, elmnatng wrong data tems wll also lead to partal data mssng. Lack of data wll cause the road real traffc condtons to change drectly or ndrectly. Therefore, t s essental to make up for the mssng data for hstorcal data. Because of the strong contnuty of the traffc flow travel tme parameter, the trend n the change of the traffc flow travel tme parameter wth tme s consstent, although ts fluctuaton wll change as the collecton perod changes. Therefore, the followng data fll formula s obtaned, as gven n equaton (1) data t data t data t data t () = ( 1) + ( 2) + ( 3) Where, data(t)represents the current mssng data, data(t-1), data(t-2) and data(t-3) arethe traffc flow travel tme data of the past perod, two cycles, and three cycles are respectvely represented. 3. Support Vector Machne Model 3.1 Problem Descrptonof Freeway Travel Tme Predcton Travel tme of a freeway has strong contnuty n a certan tme range. That s, there are some complex functonal relatonshps between the current travel tme and the past travel tme. By analyzng the changes n travel tme, we can obtan rules and establsh a real-tme predcton model updated every 5 mn. Then, the accuracy and relablty of the predcted results can be mproved by usng an optmzaton algorthm to fnd the optmal soluton of the model. The change n the freeway travel tme n dfferent tme perods s not a smple lnear relatonshp, and t wll nether ncrease ndefntely, nor decrease ndefntely. But t wll only change contnuously wthn a floatng nterval. Therefore, usng a smple least squares regresson predcton or smlar methods s not suffcent to predct the travel tme. The SVM nonlnear regresson theory can be employed to solve ths problem. SVM uses nonlnear transformaton to map the orgnal varables to a hgh-dmensonal feature space, so that the problem of nonlnear separablty n the orgnal sample space s transformed nto hgh-dmensonal feature space. The lnear separable problem and the applcaton of expanson theorem of the kernel functon n the calculaton process do not requre the explct expresson of nonlnear mappng. In addton, snce the lnear learnng machne s establshed n the hgh-dmensonal feature space, t can be compared to the lnear model. The comparson not only ncreases the complexty of the calculaton, but also solves the problem of dmensonal dsaster. Owng to changes n the traffc envronment, sudden traffc accdents, weather, and other specal events, the SVM algorthm wll eventually be transformed nto a quadratc programmng problem. In theory, a global optmal soluton can be obtaned, thus solvng the tradtonal neural network. The network can avod the local optmal problem, and should adequately accommodate (1)

6 6 of the nfluencng factors due to these sudden changes to mprove the accuracy of the travel tme predcton result. Therefore, ths study used the nonlnear support vector machne regresson theory[23]. 3.2 Model Overvew The SVM s a machne learnng method, whch s based on the statstcal learnng theory developed by Vapnk. The theory has been further extended to dversfed applcaton algorthms, ncludng the lnear SVM classfcaton algorthm, the nonlnear SVM classfcaton algorthm, the lnear SVM regresson algorthm, and the nonlnear SVM regresson algorthm[24]. These SVM algorthms have been wdely used n many felds owng to ther smple structure and hgh computatonal effcency. Consder a tranng sample set of tranng samples, ( ) l {,, } 1 S= x y x R y R, whch s = non-lnear, where x s the nput column vector of the tranng sample, 1 2 d T,,, x = x x x y R T K x, y = x * ( x ). ( ) ( ) s the correspondng output of the kernel functon j Equaton (1) n the followng s the lnear regresson equaton establshed n the hgh dmensonal space, and ε s ntroduced as a lnear nsenstve loss functon: ( ( ),, ) L f x y ε Where, ( x) s the nonlnear mappng functon, f ( ) f ( x) = ω ( x) + b (2) ( ) 0, f x y ε = f ( x) y ε, f ( x) y > ε (3) x s the predcton functon, whch returns the predcted value, and y s the correspondng real value. Under the above constrants, we can fnd the optmal classfcaton hyper-plane, that s, fnd the soluton to the followng optmzaton problem. 2 ω mn 2 T st.. ω ( x) + b y ε, = 1,2,, l Ths problem can be solved by solvng the saddle pont of the Lagrange functon, and ts dual theory can be appled to solve the dual problem. 1 mn α α α α K x x ε α α y α α 2 = 1 j= 1 = 1 = 1 l * * st.. ( α α ) = 0, α, α 0, = 1,2,, l = 1 l l l l * * * * ( )( j j) (, j) + ( ) ( ) To solve the dual problem, a relaxaton factor can be set for each data pont. After ntroducng these two relaxaton factors, ξ * *, ξ ( ξ, ξ 0, 1,2,, l) =,the functon can be optmzed as: (4) (5) l ω * mn + C ( ξ ξ ) 2 = 1 T stω.. ( x) + b y ξ + ε, = 1,2,, l T * st.. y ω ( x) b ξ + ε, = 1,2,, l (6)

7 7 of In the above equaton, C s the penalty factor; the smaller the value, the smaller the penalty to the error data. Next, the Lagrangan multpler method can be used to solve the optmzaton algorthm, and the nonlnear regresson functon can be further used to solve the double optmzaton problem. l ( ) ( ) ( * T = + = ) ( ) * ( ) ( * j + = ) (, ) + (7) f x ω x b α α x x b α α K x y b = 1 = 1 l Fgure 2.Structure of SVM model For the freeways, there s essentally no sgnfcant error n the travel tme between the low peak perod and even the flat peak perod. However, the mpact of dfferent traffc condtons on the travel tme s nevtable durng peak perods. Therefore, the two cases should be dscussed separately. Furthermore, whether or not the workng day has a dfferent nfluence on the travel tme of vehcles travelng on the hghway. The commutng tme, travel purpose, and travel mode wll be also dfferent. Therefore, these two ponts should be vewed separately. In addton, road weather condtons and traffc control wll have certan nfluence on the predcton results and should be consdered. Based on the SVM model, nput workdays, non-workng days, the mornng and evenng peak perods, and off-peak hours as orgnal values, fnally can get sx tme perods: the mornng peak hours on workdays, the evenng peak hours on workdays, off-peak hours on workdays and so on. Weather and traffc control factors for the four scenaros can also be analyzed. However, the dfference n hghway traffc condtons between the workng day and the non-workng day, the mornng and evenng peak perods and flat peak perod s not consdered due to the lmtaton of the length of the artcle. Ths study used the travel tme predcton of evenng peak hours n the classfed workng days and non-workng day peak hours as an example by comparng the traffc data of multple workng days and non-workng days. 3.3 Model Constructon The freeway travel tme predcton model s based on the SVM algorthm and s constructed based on the relatonshp between the current travel tme of the road segment and the past travel tme of the road segment, the current weather, and the possblty of traffc control. In ths study, data from two toll statons wth dfferent dstances from east to west of G5513 were selected for analyss. Moreover, as the frst-class passenger car (7 passenger car) accounts for the vast majorty of the data, the travel tme of the frst-class passenger car was taken as the predcton object, and the analyss tme nterval s 5 mnutes. The characterstc of the toll staton s presented n table 2.

8 8 of Table 2. Analyss objects of freeway travel tme predcton Toll staton Start pont End pont Dstance (klometer) 1 Changsha West Guanshan Changsha West Nngxang 23.2 The structure of the SVM s smlar to the neural network. The output s a lnear combnaton of ntermedate nodes, and each ntermedate node corresponds to a support vector. To determne the optmal classfcaton functon, ths study takes the four travel tmes of the tme perod before the predcton tme as the nput, namely tk 1, tk 2, tk 3, tk 4. t = g t, t, t, t (8) ( ) k k 1 k 2 k 3 k 4 k s the current tme perod, and t k represents the average travel tme of all vehcles n the current predcted tme perod. In the predcton process, varables such as weather, traffc accdent that affects the travel tme, holday or non-holday, and day of the week, are evaluated as follows: Varable category Table 3.Varables type and ts meanngs Varable name Varable type Varable value Meteorologcal Weather status Dscrete Tme Accdent Holday Weekly Affectng travel tme traffc accdents Dscrete Dscrete Dscrete Clear; Cloudy; Fog; Overcast; Lght ran; mod ran; hvy ran Varable meanng Clear; Cloudy; Fog; Overcast; Lght ran; mod ran; hvy ran 0 N 1 Y Monday... Sunday 0 N 1 Y Many traffc accdents occur on the freeway every day. To smplfy the parameters, we dvde traffc accdents nto two categores: traffc accdents that affect the travel tme and those that do not. In ths study, traffc accdents that affect the travel tme was regarded as nvarant. The followng fgure shows a comparson of travel tme affected by an accdent and normal travel tme on the afternoon of February 17.

9 9 of 15 travel tme/s Normal travel tme on February Accdent travel tme on February :00 17:15 17:30 17:45 18:00 18:15 18:30 18: Fgure 3. Comparson of accdents and normal travel tme Snce SVM s a machne learnng model, sample tranng s requred before predcton. Multple groups of any four consecutve travel tmes, daly weather, traffc accdents that affect the travel tme, holdays, and weekday data were used as tranng samples to obtan a traned model. In the traned model, tk 1, tk 2, tk 3, tk 4, weather, accdent, holday, and week are used to predct the travel tme of the next tme perod. When a certan number of tranng samples s acheved, real-tme data nput can be adopted to predct future results. Moreover, the model can be constantly modfed based on the relaton between the predcted data and the predcted data to predcton accuracy. 3.4 Parameter Calbraton and Optmzaton Parameter selecton s very mportant to fnd the optmal hyper-plane n the SVM model used n ths study. Exstng studes manly adopted the tradtonal grd search method, drect determnaton method, one-dmensonal search method, and nverse rato method to determne the nsenstve loss functon parameter ε and penalty parameter C. However, there are many shortcomngs assocated wth these methods, and the resultng errors wll sgnfcantly nfluence the accuracy of the predcton results. Moreover, n the SVM model, kernel functon selecton s also an mportant factor that nfluences the performance of the SVM. The radal bass kernel functon K( x, y ) = e (RBF) s an adaptve kernel functon for low-dmensonal space data and hgh-dmensonal space, whch have good convergence domans, and ths functon can be descrbed as an deal kernel functon. Therefore, the RBF was selected as the classfcaton predcton kernel functon of the SVM, n whch a kernel parameter σ needs to be optmzed. Therefore, three parameters need to be optmzed, namely the core parameter σ, the non-senstve loss functon parameter ε, and the punshment parameter C. The kernel parameter σ s the dstrbuton or range of the tranng sample data. The non-senstve loss functon parameter ε affects the number of support vectors. The larger the value of ε, the lower the regresson precson, and the fewer the support vectors. The penalty parameter C s used to control the degree of punshment of samples beyond the allowable error range. The hgher the value, the heaver the punshment of samples. We used the artfcal fsh swarm algorthm to optmze the parameters of the regresson model. The artfcal fsh swarm algorthm has unque advantages n parameter optmzaton and overcomes the blndness of tradtonal algorthms n parameter optmzaton and the defects of the lnear model and neural network n parameter selecton. It can be sad that the parallel performance of the artfcal fsh swarm algorthm can ensure that the model parameters converge faster to the global optmzaton extreme[24,25]. x x 2 σ

10 10 of The frst step n the optmzaton process of the artfcal fsh swarm algorthm s to feed n the tranng value and the tranng target through the SVM model to calculate the ftness of the ndvdual. The most adaptable ndvdual s regarded as the optmal value of the current fsh group and the correspondng parameters σ, ε, and C of the current optmal value are saved. In the subsequent teraton, σ, ε, and C correspondng to the maxmum ftness value are taken as the fnal optmzaton results. 4. Case Study 4.1 Data Selecton The data used n ths study was collected n February 2018 on G5513 (from Changsha West to Guanshan/Nngxang Staton)n Changsha, Hunan Provnce, Chna. The travel tme was detected for all days and the detecton nterval s 5 mnutes. Furthermore, 288 sequences are ncluded n one day. The daly evenng peak (17:00-19:00) data of G5513 (Changsha West to Guanshan /Nngxang Staton) was selected as an example after comparng data of multple workng days and non-workng days, whch contans 204 to 228 tems. Other varable data tems also need to be fltered accordng to the above data. The dmenson feature values are based on the tme seres and the data requrements accordng to the predcton model. In ths study, the regresson SVM model s used to establsh the model parameters, and the artfcal fsh swarm algorthm s used to establsh the model parameter optmzaton algorthm. The optmzaton results are presented n Table 4. The optmzaton process for the optmal value of the penalty parameter C s shown n Fgure 4. Table 4. Optmzaton of parameter values Penalty parameter, Nuclear parameter, Insenstve loss fun Secton C σ cton parameter, ε Changsha West-Guanshan Changsha West-Nngxang Fgure 4. The parameter optmzaton curve The parameters of the artfcal fsh swarm algorthm are set as follows: the maxmum number of teratons of the artfcal fsh s 100; the populaton sze s 5; the maxmum number of trals s 5; the crowdng factor δ s 0.618; the perceved dstance s 0.5; and the movng step s Results and Comparatve Analyss The data adopted n ths study were obtaned durng the Chnese Sprng Festval from February15 to 21, 2018.Therefore, the data was dvded nto workng days and holdays. There were

11 11 of sets of data from February1 to 14,2018(14 days data) n each group of toll statons; 264 groups (11 days) were randomly selected as tranng data nput, and 72 groups (3 days) were adopted as predcton numbers. There were 168 sets of data from February15to 21,2018(11 days data) n each group of toll statons;96 groups (4 days) were randomly selected as tranng data nput, and 72 groups (3 days) were adopted as predcton numbers. The detecton tme s from 17:00 to 19:00 P.M. and the value was taken as test data. BP neural network, SVM, and optmzed SVM were used for the predcton. The root mean square error (RMSE), the mean absolute percent error (MAPE) and the covarance protocol (CP) were selected as the error evaluaton crtera n the predcton process[26]. The RMSE s a comprehensve evaluaton ndcator of the predcton effect, the MAPE s the predcton relatve error, whle the CP s the error component analyss ndcator. The followng fgures show the forecastng effect dagram of the holday evenng peak Fgure 5.Results of freeway travel tme predcton Table 5.Error evaluaton of forecastng freeway travel tme Workng day Method Changsha West- Guanshan Staton Changsha West-Nngxang Staton RMSE MAPE CP RMSE MAPE CP BP neural network SVM

12 12 of 15 Holday Optmzed SVM BP neural network SVM Optmzed SVM From Table 5, t can be observed that the workng day was predcted by the RMSE, the MAPE, and the CP data between the two toll statons. It could be found that all three models can be used for predctng travel tme. Although the predcton error of the BP neural network may be larger than those of the SVM and the optmzed SVM models, there s no devaton between the error of the SVM and the optmzed SVM model. However, when forecastng holdays wth large traffc and long travel tme, the RMSE of the optmzed SVM model s sgnfcantly better than those of the BP neural network and the SVM model. In the predcton of Changsha West-Guanshan, the accuracy of the optmzed SVM model usng artfcal fsh swarm algorthm s 17.27% hgher than that of the BP neural network model and 16.44% hgher than the conventonal SVM model. In the predcton of Changsha West-Nngxang, the accuracy of the optmzed SVM model usng artfcal fsh swarm algorthm s 23.80% and 8.01% hgher than those of the BP neural network model and the conventonal SVM model, respectvely. The optmzed SVM model descrbed n ths paper has hgher travel tme predcton accuracy n the road segment, and the mappng law of the nput and output are better represented by the optmzed SVM model. In terms of the relatve predcton errors of the three predcton models, the MAPE of the optmzed SVM model s lower than the predcton errors of the BP neural network and the conventonal SVM model when usng holday and workng day data. Ths ndcates that the optmzed SVM model descrbed n ths paper has certan advantages n terms of the travel tme predcton model of the road segment, and the data requrements are lower. 4.3 Analyss of Influencng Factors of Travel Tme The freeway travel tme s determned by varous factors, such as weather, traffc accdent, holday, and week day. However, owng to lmtatons of sample data, only traffc accdents and holdays were consdered Effect of traffc accdents on travel tme Frst, the optmzed SVM model descrbed n ths paper s superor to the BP neural network and SVM model n terms of the CP. In the predcton of Changsha West-Guanshan, the CP of the optmzed SVM model s 21.40% hgher than that of the BP neural network model, whch s 7.28% hgher than that of the conventonal SVM model; In the predcton of Changsha West-Nngxang, the CP of the optmzed SVM model s 10.9% and 6.53% hgher than those of the BP neural network model and the conventonal SVM model, respectvely. The results presented n ths paper ndcates that the optmzed SVM model has better nclusveness and stablty when unexpected factors such as traffc accdents that affect the travel tme are encountered, thereby avodng the need for repeated tral and error to address network problems.

13 13 of Second, freeway traffc accdents wll cause the traffc capacty of certan sectons of the road network to decrease, and queues wll be formed near the accdent ste, whch ncreases the travel tme of the vehcle Effect of holdays on travel tme A comparson of the workng day and holday forecastng error evaluaton crtera presented n the prevous secton ndcates that the BP neural network has a larger predcton error than the workng day predcton results of the other models probably due to the problem of constructon of the network structure. However, the conventonal SVM and the optmzed SVM models have smlar predcton error results. In the above analyss, there are large gaps n the holday predcton results of the three dfferent models. It can observed that holdays have sgnfcant nfluence on the travel tme. The effect of holdays on the travel tme, whch was obtaned from the analyss of the orgnal toll staton data, s that t sgnfcantly ncreases the volume of traffc on the hghway network. 5. Conclusons Ths study performed an n-depth analyss of freeway travel tme predcton to provde hgh-qualty travel experence for users and found that the freeway travel tme s affected by travel tme, weather, and traffc. The effect of dfferent factors was analyzed, such as accdents and holdays. Bad weather reduces the overall traffc rate, whch ncreases the travel tme. Traffc accdents lead to reduced road traffc capacty, whch affects the travel tme. Free passage on hghways durng holdays and the ncreased demand for travel result n ncreased vehcle flow, whch also affects the travel tme. In ths study, basc data was analyzed, and the traffc state predcton method based on SVM data mnng technology was proposed to transform the problem nto a quadratc programmng problem usng artfcal fsh swarm algorthm, whch reduces the computatonal and local optmal problems of tradtonal neural networks. The parameters of the SVM were optmzed usng tradtonal network optmzaton, and a global optmal soluton was obtaned. Results show that the accuracy of the optmzed SVM model s 17.27% and 16.44% hgher than those of the BP neural network model and the conventonal SVM model, respectvely. Accurate predcton of the travel tme on the freeway was realzed, whch can provde data support for montorng, early warnng, and decson analyss for the freeway operaton status. In ths study, nfluencng factors such as weather, traffc, accdents and holdays were ncluded n the optmzaton of the SVM predcton model. However, owng to lmtatons of the number of samples, the model was not fully traned. Therefore, a certan error occurred n the predcton results. In the future, t s necessary to categorze traffc accdents, clarfy the mpact of each type of accdent on the travel tme, categorze ncrease n the holday traffc, and clarfy the mpact of each level of accdent on the travel tme. Furthermore, the number of tranng samples, database capacty, and predcton accuracy should be contnuously ncreased. In ths study, only data from freeway toll statons was valdated, and applcaton to actual large-scale road networks should be further explored n the future. Author Contrbutons: Ths work was conducted by Kejun Long and We Wu wth the help of graduate student Wuka Yao. It was manly drafted by Kejun Long and Wuka Yao, and checked and revsed by We Wu and Jan Gu. Kejun Long and Wuka Yao desgned and analyzed the proposed model. Wuka Yao and Jan Gu performed the smulaton. Kejun Long s responsble for the Englsh polsh and proofreadng of the work. Fundng: Ths research was funded by Natonal Natural Scence Foundaton of Chna (NSFC),grant number , Hunan Provncal Key Laboratory of Smart Roadway and Cooperatve Vehcle-Infrastructure Systems,grant number2017tp1016.

14 14 of Acknowledgments: The authors express ther thanks to all who partcpated n ths research for ther cooperaton. The authors would lke to gve great thank to the hard work by the peer revewers and edtor. Conflcts of Interest: The authors declare no conflct of nterest. References 1. Gpps,P.G. The estmaton of a measure of vehcle delay from detector output. Newcastle-Upon-Tyne Unversty, England1977, 18 p. 2. Yldrmoglu, M; Gerolmns, N. Experenced travel tme predcton for congested freeways. Transp. Res. Part B-Method, 2013, 53(4): Shen,L; Had,M. Practcal approach for travel tme estmaton from pont traffc detector data. J. Adv Transportaton, 2013, 47(5): Hyun,K; Tok,A, Rtche S G. Long dstance truck trackng from advanced pont detectors usng a selectve weghted Bayesan model. Transport. Res. Part C-Emerg. Technol, 2016, 82: Hyun, K. K.; Jeong, K. Assessng crash rsk consderng vehcle nteractons wth trucks usng pont detector data. Accd. Anal. Prevent, 2018, 17p. 6. Ramezan, M.; Gerolmns, N. On the estmaton of arteral route travel tme dstrbuton wth Markov chans. Transp. Res. Part B, 2012, 46(10): Zhang, J. T.; Zhou, J. An Arteral Travel Tme Estmaton Model Based on Dscrete Tme Markov Chans. Syst. Eng, 2014,5: Zhang, J. T.; Zhou, J. Travel tme estmaton model based on spatal Markov chans. Syst. Eng,2015,12 : Woodard, D.; Nogn,G. Predctng travel tme relablty usng moble phone GPS data. Transport. Res. Part C-Emerg. Technol, 2017, 75: Bahuleyan, H.;Vanajaksh, L. D. Arteral path-level travel-tme estmaton usng machne-learnng technques. J. Comput. Cvl. Eng, 2016, 31(3), Tan, Y. Current Stuaton of Short-term Flow Forecastng and Dscusson on Forecastng Method n Bg Data Envronment. ITS Chna:2017: Ln J W C V. Incremental and onlne learnng through extended kalman flterng wth constrant weghts for freeway travel tme predcton. ITSC. IEEE, 2006: Zhou, J.; Zhang, C.B. Travel Tme Predcton Model for Urban Road Network based on Mult-source Data. Proceda - Socal and Behavoral Scences, 2014, Tang-Hsen Chang; Albert Y. Chen. Freeway Travel Tme Predcton Based on Seamless Spato-temporal Data Fuson: Case Study of the Freeway n Tawan. Transportaton Research Proceda,2016, Fe X; Lu C C. A Bayesan dynamc lnear model approach for real-tme short-term freeway travel tme predcton. Transp. Res. Part C, 2011, 19(6): Zhan, X.; Ukkusur, S. V. A Bayesan mxture model for short-term average lnk travel tme estmaton usng large-scale lmted nformaton trp-based data. Autom.Constr,2016, 72, Wosyka,J; Přbyl, P. Real-tme travel tme estmaton on hghways usng loop detector data and lcense plate recognton. Elektro. IEEE, 2012: Innamaa, S. Short-Term Predcton of Travel Tme usng Neural Networks on an Interurban Hghway. Transportaton, 2005, 32(6): Lnt, J.W.C.V; Hoogendoorn, S P. Accurate freeway travel tme predcton wth state-space neural networks under mssng data. Transp. Res. Part C, 2005, 13(5): Wu, Chun-Hsn; Jan-Mng Ho. Travel-tme predcton wth support vector regresson. IEEE.Trans.Intell.Transp.Systems.5.4,2004: Vanajaksh, L; Rlett, L R. Support Vector Machne Technque for the Short Term Predcton of Travel Tme. Intellgent Vehcles Symposum. IEEE, 2007: Mendes-Morera, João. Comparng state-of-the-art regresson methods for long term travel tme predcton. Intell. Data. Anal (2012): Wang, X.;Chen, X. H.; Yang, X. M. Short term predcton of expressway travel tme based on k nearest neghbor algorthm.chn.j.hghw.transport,2015,28(1), L, S.; Yuan, Z. C.; Wang, C. Optmzaton of support vector machne parameters based on group ntellgence algorthm. CAAI TIS,2018,13(01):70-84.

15 15 of Wang, Q.; Lu, Z.; Peng, Z.A PSO-SVM Model for short-term travel tme predcton based on Bluetooth Technology. J. Harbn. Inst. Technol, 2015,22(3), Yang, Z. S. Study on the Synthetc Lnk Travel Tme Predcton Model of Key Theory of ITS.J.Traff.Transp.Eng,2001,01:65-67.