AReviewofTrafficCongestionPredictionUsing …2020/08/08 · English language, 48 articles were...

Review ArticleA Review of Traffic Congestion Prediction UsingArtificial Intelligence

Mahmuda Akhtar and Sara Moridpour

Department of Civil and Infrastructure Engineering RMIT University Melbourne VIC 3000 Australia

Correspondence should be addressed to Mahmuda Akhtar s3799862studentrmiteduau

Received 8 August 2020 Revised 7 January 2021 Accepted 18 January 2021 Published 30 January 2021

Academic Editor Michael Bazant

Copyright copy 2021Mahmuda Akhtar and Sara Moridpour+is is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in anymedium provided the original work isproperly cited

In recent years traffic congestion prediction has led to a growing research area especially of machine learning of artificialintelligence (AI) With the introduction of big data by stationary sensors or probe vehicle data and the development of new AImodels in the last few decades this research area has expanded extensively Traffic congestion prediction especially short-termtraffic congestion prediction is made by evaluating different traffic parameters Most of the researches focus on historical data inforecasting traffic congestion However a few articles made real-time traffic congestion prediction +is paper systematicallysummarises the existing research conducted by applying the various methodologies of AI notably different machine learningmodels +e paper accumulates the models under respective branches of AI and the strength and weaknesses of the modelsare summarised

1 Introduction

Artificial intelligence (AI) is the most important branch ofcomputer science in this era of big data AIwas born 50 years agoand came a longwaymaking encouraging progress especially inmachine learning datamining computer vision expert systemsnatural language processing robotics and related applications[1] Machine learning is the most popular branch of AI Otherclasses of AI include probabilistic models deep learning arti-ficial neural network systems and game theory+ese classes aredeveloped and applied in a wide range of sectors Recently it hasbeen the leading research area in transportation engineeringespecially in traffic congestion prediction

Traffic congestion has a direct and indirect impact on acountryrsquos economy and its dwellersrsquo health According toAli et al [2] traffic congestion causes Pak Rs 1 millionevery day in terms of opportunity cost and fuel con-sumption due to traffic congestion Traffic congestion af-fects on individual level as well Time loss especially duringpeak hours mental stress and the added pollution to theglobal warming are also some important factors caused dueto traffic congestion

Ensuring economic growth and the road usersrsquo comfortare the two requirements for the development of a countrywhich is impossible without smooth traffic flow With thedevelopment in the transportation sector by collecting trafficinformation authorities are puttingmore attention on trafficcongestion monitoring Traffic congestion prediction pro-vides the authorities with the required time to plan in theallocation of resources to make the journey smooth fortravellers Traffic congestion prediction problem discussedin this paper can be defined as an estimation of parametersrelated to traffic congestion into the short-term future eg15 minutes to a few hours by applying different AI meth-odologies by using collected traffic data +ere are usuallyfive parameters to evaluate including traffic volume trafficdensity occupancy traffic congestion index and travel timewhile monitoring and predicting traffic congestionsDepending on the nature of the collected data a variety of AIapproaches are applied to evaluate the congestion param-eters +is article systematically discusses the models andtheir advantage and disadvantages +e primary motivationof this review is to gather the articles focusing solely ontraffic congestion prediction models +e keywords used in

HindawiJournal of Advanced TransportationVolume 2021 Article ID 8878011 18 pageshttpsdoiorg10115520218878011

the search process included ldquotraffic congestion predictionrdquoOR ldquotraffic congestion estimationrdquo OR ldquocongestion pre-diction modellingrdquo OR ldquoprediction of traffic congestionrdquo ORldquoroad congestion forecastrdquo OR ldquotraffic congestion forecastrdquoFor efficient screening research paper search was doneaccording to year using search engines like Scopus GoogleScholar and Science Direct After collecting all the peer-reviewed journal and conference papers written in theEnglish language 48 articles were found for review Anystudies focusing on the cause of traffic congestion trafficcongestion control traffic congestion impact traffic con-gestion propagation traffic congestion prevention etc wereexcluded from this manuscript

A general layout of the prediction approaches is pro-vided in Section 2 +e data collection sources and con-gestion forecasting models are explained in Sections 3ndash6 andthey provide the overall discussion and concluding remarks

2 General Layout

Traffic congestion forecasting has two basic steps of datacollection and prediction model development Every step ofthe methodology is important and may affect the results ifnot done correctly After data collection data processingplays a vital role to prepare the training and testing datasetsCase area differs for different research After developing themodel it is validated with other base models and groundtrue results Figure 1 shows the general components of trafficcongestion prediction studies +ese branches were furtherdivided into more specific sub-branches and are discussed inthe following sections

3 Data Source

Traffic datasets used in different studies can be mainly di-vided into two classes including stationary and probe dataStationary data can be further divided into sensor data andfixed cameras On the other hand probe data that were usedin the studies were GPS data mounted on vehicles

Stationary sensors continuously capture spatiotemporaldata of traffic However sensor operation may interruptanytime Authorities should always consider this temporaryfailure of the sensor while planning by using this data +eadvantage of the sensor data is that there is no confusion onthe location of the vehicles +e most used dataset wasPerformance Measurement System (PeMS) that collectshighway data across all major metropolitan areas of the Stateof California of traffic flow sensor occupancy and travelspeed in real-time Most of the studies used dataset from theI-5 highway in San Diego California every 5 minutes [3ndash6]Other systems included the Genetec blufaxcloud travel-timesystem engine (GBTTSE) [7] and the Topologically Inte-grated Geographic Encoding and Referencing (TIGER) linegraph [8]

On the other hand probe data has the advantage ofcovering the entire road network A network consists ofdifferent structured roads +erefore studies especiallythose that considered the network wide area used probedata +e most used dataset was GPS data collecting every

second from approximately 20000 taxies of Beijing ChinaData included the taxi number the latitude-longitude of thevehicle timestamp when sampling and whether there was apassenger or not Data updating frequency of this datasetvaries from 10 s to 5min according to the quality of GPSdevice [4 5 9] Other probe data included low-frequencyProbe Vehicle Data (PVD) [10] and bus GPS data [11 12]However sometimes probe data show significant fluctua-tion Besides map matching is usually a must for probe dataBut data can minimize this limitation Probe data collectedfrom one city cannot be used directly for modelling othercity networks+is is because the data collected from BeijingChina includes latitude-longitude of the vehicle which isunique However a generalised model using probe data canbe generated for different cities

Other data sources eg data from tolling system anddata provided by transportation authority will add morereliable data as the sources are dependable However a lot ofthe times study area needs to be adjusted as in most casestolled road information is not available Tracking cellularphone movements without privacy breach can also be asource of data However the heterogeneity of the vehicledistribution will be hard to determine from this dataset ifnot impossible Besides due to pedestrian or cyclists trav-elling through the sidewalk there might be many outliers inthe dataset if modelling is done for a road network Datacollected from a questionnaire to the general publicdriversmay provide a misleading result [13]

31 Clustering Algorithms Some studies use clustering theacquired data before applying the main congestion modelsof prediction +is hybrid modelling technique is applied tofine-tune the input values and to use them in the trainingphase Figure 2 shows the commonly used AI clusteringmodels in this field of research +e models are describedbriefly in this section

Fuzzy C-Means (FCM) is a popular nondeterministicclustering technique in data mining In traffic engineeringresearches traffic pattern recognition plays an importantrole Besides these studies often face the limitation ofmissing or incomplete data To deal with these constraintsFCM has become a commonly applied clustering technique+e advantage of this approach is unlike original C-meansclustering methods it can overcome the issue of gettingtrapped in the local optimum [14] However FCM requiressetting a predefined cluster number which is not alwayspossible while dealing with massive data without any priorknowledge of the data dimension Besides this model be-comes computationally expensive with data size incrementDifferent studies have applied FCM successfully by im-proving its limitations Some studies changed the fuzzyindex value for each FCM algorithm execution [15] somecalculated the Davies-Bouldin (DB) index [10] while othersapplied the K-means clustering algorithm [16 17]

K-means clustering is an effective and relatively flexiblealgorithm while dealing with large datasets It is a popularunsupervised machine learning algorithm Depending onthe features cluster number varied from two [18] to 50

2 Journal of Advanced Transportation

[19ndash21] Like FCM K-means clustering requires a pre-defined cluster number and selecting K original clustercentres GAP [22] and WEKA toolbox [23] were used todefine the value For large datasets as the sample distri-bution is unknown in the beginning it is not always possibleto fulfil these two requirements A few studies used adaptive

K-means clustering overcoming the limitations andexploited the pattern using principal component analysis(PCA) [24 25]

DBSCAN is more of a general clustering application inmachine learning and data mining +is method overcomesthe limitation of FCM of predefining the cluster number It

AI clustering algorithms

Fuzzy C-meansDensity-based spatial

clustering of applications with noise

K-means

Figure 2 Commonly used AI clustering algorithms

Methodology

Traffic congestion prediction

Data preprocessing

Scope of study area

Modelling applying AI

Traffic parameters

Traffic volume

Traffic density

Sensor occupancy

Traffic speed

Traffic congestion

index

Model validation

AI models

Probabilistic reasoning

Shallow machine learning

Deep machine learning

Dataset

Stationary sensor data

Probe vehicle data

Clustering No clustering

Traffic congestion state leveling

Figure 1 +e layout of traffic congestion prediction system

Journal of Advanced Transportation 3

can automatically generate arbitrary cluster shapes sur-rounded by clusters of different characteristics and can easilyrecognise outlier However it requires two parameters topreset A suitable parameter determination method egtrial and error method [8] and human judgement [26] makesthe model computationally expensive and requires a clearunderstanding of the dataset

From the above discussion it is concluded that only 16out of 48 studies have done clustering before applyingprediction models Several time-series models and shallowmachine learning (SML) algorithms have used clusteringapproach However deep learning algorithms can processinput data on different layers of the model thus may notneed clustering beforehand

4 Applied Methodology

Traffic flow is a complex amalgamation of heterogenoustraffic fleet +us traffic pattern prediction modelling couldbe an easy and efficient congestion prediction approachHowever depending on the data characteristics and qualitydifferent classes of AI are applied in various studies Figure 3shows the main branchesmdashprobabilistic reasoning andmachine learning (ML) Machine learning comprised ofboth shallow and deep learning algorithms However withthe progress of this article these sections were subdividedinto detailed algorithms

To generalise traffic congestion forecasting studies usingdifferent models is not straight forward +e commonfactors of all the articles include the study area data col-lection horizon predicted parameter prediction intervalsand validation procedure Most of the articles took studiedcorridor segment as the study area [5 27ndash30] Other studyareas included the traffic network [31 32] ring road [9] andarterial road [33] Data collection horizon varied from 2years [34] to less than a day [35] in the studies Congestionestimation is done predicting traffic flow parameters egtraffic speed [4] density speed [5] and congestion index[31] to mention a few +e Congestion Index (CI) approachis suitable to monitor the congestion level continuously in aspatiotemporal dimension Studies those compared theirresults with the ground truth value or with other modelsused mean absolute error (MAE) (equation (1)) symmetricmean absolute percentage error (sMAPE) (equation (1))MAPE root-mean-squared error (RMSE) (equation (3))false positive rate (FPR) (equation (4)) and detection rate(DR) (equation (5)) Many studies used SUMO to validatetheir models

MAE 1n

1113944

n

i1Yi minus Yi

11138681113868111386811138681113868111386811138681113868 (1)

sMAPE 1n

1113944

n

i1

Yi minus Yi

11138681113868111386811138681113868111386811138681113868

Yi

11138681113868111386811138681113868111386811138681113868 + Yi

111386811138681113868111386811138681113868111386811138681113872 11138732

100 (2)

RMSE

1113936ni1 Yi minus Yi( 1113857

2

n

1113971

(3)

where Y original value Yi predicted value andn number of instances

FPR FP

TN + FP (4)

DR TP

FN + TP (5)

where FP TN FN and TP represent the false positive truenegative false negative and true positive respectively

+e rest of this section will discuss the methodology theauthors have applied in the studies

41 Probabilistic Reasoning Probabilistic reasoning is asignificant section of AI It is applied to deal with the field ofuncertain knowledge and reasoning A variety of these al-gorithms are commonly used in traffic congestion predictionstudies +e studies discussed hereunder probabilistic rea-soning is shown in Figure 4

411 Fuzzy Logic Zadeh is a commonly applied model indynamic traffic congestion prediction as it allows vaguenessinstead of binary outcomes In this method several mem-bership functions are developed those represent the degreeof truth With the vastness with time traffic data are be-coming complex and nonlinear Due to its ability to dealwith uncertainty in the dataset fuzzy logic has becomepopular in traffic congestion prediction studies

A fuzzy system comprises of several fuzzy sets which isbuilt of membership functions +ere are usually threecodification shapes to choose for the membership functions(MFs) of input triangular trapezoidal and Gauss function+e fuzzy rule-based system (FRBS) is the most commonfuzzy logic system in traffic engineering research It consistsof several IF-THEN rules that logically relate the inputvariables with output It can effectively deal with thecomplexity resulting from real-world traffic situations byrepresenting them in simple rules +ese rules combine therelations among different traffic states to detect the resultingtraffic condition [36] However with the growth in datacomplexity the total number of rules also grows lesseningthe accuracy of the whole system thus making it compu-tationally expensive To better manage this problem twotypes of fuzzy logic controls are applied In hierarchicalcontrol (HFRBS) according to the significance the inputvariables are ordered and MFs are employed Figure 5 showsa simple HFRBS structure MFs are optimized by applyingdifferent algorithms eg genetic algorithm (GA) [30] hy-brid genetic algorithm (GA) and cross-entropy (CE)

Artificial intelligence

Probabilisticreasoning


Deep machinelearning

Figure 3 Branches of artificial intelligence in this article


[28 37] compared the performance of evolutionary crisprule learning (ECRL) and evolutionary fuzzy rule learning(EFRL) for road traffic congestion prediction It was seenthat ECRL models outperformed EFRL in terms of averagedaccuracy and no of rules but was computationally expensive

+e Takagi-Sugeno-Kang (TSK) (FRBS) model is one ofthe simple fuzzy models due to its mathematical treatabilityA weighted average computes the output of this modelAnother simple FRBS model is Mamdani-type model +eoutput of this model is a fuzzy set which needs defuzzifi-cation which is time-consuming Due to its good inter-pretability it can improve the accuracy of fuzzy linguisticmodels Cao and Wang [3] applied this model to show thecongestion severity change among road grades A fewstudies used this method to fuse heterogenous parameters[7 13] +e TSK model works on improving the inter-pretability of an accurate fuzzy model TSK is applied for itsfast calculation characteristics [37]

+e fuzzy comprehensive evaluation (FCE) uses theprinciple of fuzzy transformation and maximum member-ship degree +is model consists of several layers which is auseful objective evaluation method assessing all relevantfactors +e number of layers depends on the objectivecomplicacy and the number of factors Kong et al [4] andYang et al [5] applied FCE in which the weights and thefuzzy matrix of multi-indexes were adapted according to thetraffic flow to estimate traffic congestion state Adaptivecontrol adjusts weight coefficient based on judgement ma-trix Certain weights are assigned to calculate the mem-bership degree of the parameters [35]

Other than GA and PSO Ant Colony Optimization(ACO) algorithmwas also introduced by Daissaoui et al [38]in fuzzy logic system +ey provided the theory for a smartcity where each vehicle GPS data was taken as a pheromoneconsistent with the concept of ACO +e objective was topredict traffic congestion one minute ahead from the in-formation (pheromone) provided by past cars However thearticle does not give any result on support to the model

As discussed before with the development of optimi-sation algorithms optimisation of the fuzzy logic systemrsquosmembership functions is becoming diverse With time thesimplest form of FRBS-TSK has become popular due to itsgood interpretability Some other sectors of transportationwhere fuzzy logic models are popular include traffic lightsignal control [39 40] traffic flow prediction (Zhang and Ye[41]) traffic accident prediction [42] and modified fuzzylogic for freeway travel time estimation (Zhang and Ge [43])

+e fuzzy logic system is the only probabilistic reasoningmodel that can have an outcome of more than congestednoncongested state of the traffic state+is is one of the mainadvantages that has made this methodology popularHowever no study has provided any reasonable logic onselecting the membership function which is a significantlimitation of fuzzy logic models

412 Hidden Markov Model +e hidden Markov model(HMM) is a combination of stochastic characteristics ofMarkov process and discrete characteristics of Markovchains It is a stochastic time-series event recognitiontechnique Some studies have applied Markov chain modelfor traffic pattern recognition during congestion prediction[21 25 44] Pearson correlation coefficient (PCC) is com-monly applied among the parameters during pattern con-struction Zaki et al [32] applied HMM to select theappropriate prediction model from several models theydeveloped applying the adaptive neurofuzzy inference sys-tem (ANFIS) +ey obtained optimal state transition by fourprocessing steps initialization recursion termination andbacktracking +e last step analysed the previous step todetermine the probability of the current state by using theViterbi algorithm Based on the log-likelihood of the initialmodel parameter defined by expectation maximization(EM) algorithm of HMM with the traffic pattern a suitablecongestion model was selected for prediction Mishra et al[23] applied the discretised multiple symbol HMM (MS-HMM) prediction model named future state prediction(FSP) +ey evaluated model adaptability for different roadsegments A label was generated containing hidden states ofMS-HMM and the output was used for FSP to result in thenext hidden state label

In traffic engineering especially while utilising probevehicle data HMM is very useful inmap-matching Sun et al[45] applied HMM for mapping the trajectory of observedGPS points in nearby roads +ese candidate points weretaken as hidden states of HMM +e candidate points closerto the observation point had higher observation probability

Bayesian network


Fuzzy logic Hidden markov model

Gaussian mixture model

Figure 4 Subdivision of probabilistic reasoning models

Variable 1

Variable 2

Subsystem 1Rule base 1

OutputSubsystem 2Rule base 2

Variable 3

Figure 5 A simple structure of HFRBS


Transition probability of two adjacent candidates was alsoconsidered to avoid the misleading results generated fromabrupt traffic situations

HMM shows accuracy in selecting a traffic pattern or atraffic point It has the advantage that it can deal with thedata with outliers However points with a short samplinginterval seem to be matched well and long intervals andhigher similar probe data decreased the model accuracyStudies have found a significant mismatch for long samplinginterval dataset and similar road networks

+e GPS tracking system has been widely developed inthis era of the satellite +us making HMM modelling iscurrently more relevant for map matching Other sectors oftransport where HMM is applied include traffic prediction[46] modified HMM for speed prediction [47] and trafficflow state transition [48] etc

413 Gaussian Distribution Gaussian processes haveproven to be a successful tool for regression problemsFormally a Gaussian process is a collection of randomvariables any finite number of which obeys a joint Gaussianprior distribution For regression the function to be esti-mated is assumed to be generated by an infinite-dimensionalGaussian distribution and the observed outputs are con-taminated by additive Gaussian noise

Yang [29] applied Gaussian distribution for trafficcongestion prediction in their study+is study was dividedinto three parts First the sensor ranking was doneaccording to the volume quality by applying p test In thesecond part of the study the congestion occurring prob-ability was determined from a statistics-based method Inthe learning phase of this part two Gaussian probabilitymodels were developed from two datasets for every point ofinterest In the decision phase on which model the inputtraffic volume value fitted was evaluated and a predictionscore presenting congestion state was determined from theratio of two models Finally the probability of congestionoccurring at the point of interest was found by combiningand sorting the prediction score from all the ranked sen-sors Zhu et al [49] also presented the probability of trafficstate distribution Selection of mean and variance pa-rameters of Gaussian distribution is an important step Inthis study the EM algorithm was applied for this purpose+e first step generated the log-likelihood expectation forthe parameters whereas the last step maximised it Sunet al [45] approximated the error in GPS location in theroad with Gaussian Distribution taking mean 0 +e errorwas calculated from the actual GPS point matching pointon the road section and standard deviation of GPS mea-surement error

From the abovementioned studies it is seen that theGaussian distribution model has a useful application inreducing feature numbers without compromising the qualityof the prediction results or for location error estimationwhile using GPS data Gaussian distribution is also appliedin traffic volume prediction [50] traffic safety [51] andtraffic speed distribution variability [52]

414 Bayesian Network A Bayesian network (BN) alsoknown as a causal model is a directed graphical model forrepresenting conditional independencies between a set ofrandom variables It is a combination of probability theoryand graph theory and provides a natural tool for dealing withtwo problems that occur through applied mathematics andengineeringmdashuncertainty and complexity [53]

Asencio-Cortes et al [54] applied an ensemble of sevenmachine learning algorithms to compute the traffic con-gestion prediction +is methodology was developed as abinary classification problem applying the HIOCC algo-rithm Machine learning algorithms applied in this studywere K-nearest neighbour (K-NN) C45 decision trees(C45) artificial neural network (ANN) of backpropagationtechnique stochastic gradient descent optimisation (SGD)fuzzy unordered rule induction algorithm (FURIA)Bayesian network (BN) and support vector machine (SVM)+ree of these algorithms (C45 FURIA and BN) canproduce interpretable models of viewable knowledge A setof ensembled learning algorithms were applied to improvethe results found from these prediction models +e en-semble algorithm group included bagging boosting (Ada-Boost M1) stacking and Probability +reshold Selector(PTS) +e authors found a significant improvement inPrecision for BN after applying ensemble algorithms On theother hand Kim andWang [34] applied BN to determine thefactors that affect congestion initialization on different roadsections +e developed model of this study gave a frame-work to assess different scenario ranking and prioritizing

Bayesian network is seen to perform better withensembled algorithms or while modified eg other trans-port sectors of traffic flow prediction [55] and parameterestimation at signalised intersection [56 57]

415 Others Other than the models mentioned above theKalman Filter (KF) is also a popular probabilistic algorithmWith the increment of available data data fusion methodsare becoming popular +e fusion of historical and real-timetraffic data can achieve a higher level of traffic congestionprediction accuracy In this regard KF is commonly appliedExtended KF (EKF) is an extension of KF which can be usedto stochastically filter the nonlinear noises to improve themean and covariance of an estimated state +erefore afterdata fusion it updated the estimated covariance error byremoving outliers [7]

Wen et al [8] applied GA in traffic congestion predictionfrom spatiotemporal traffic environment Temporal asso-ciation rules were extracted from the traffic environmentapplying GA-based temporal association rules (GATARs)+eir proposed Hybrid Temporal Association Rules Miningmethod (HTARM) included DBSCAN and GATARmethods +e DBSCAN application method was discussedpreviously in this article While encoding using GATARroad section number and congestion level were included inthe chromosome+e decoding was done to obtain temporalassociation rules and was sorted according to confidence andsupport value in the rule pool For both simulated and real-


world scenarios the proposed HTARM method out-performed GATAR in terms of extracting temporal asso-ciation rules and prediction accuracy However the clusternumber difference showed a big difference in the two sce-narios Besides with the increment of road network com-plexity the prediction accuracy decreased

Table 1 summarises the methodologies and differentparameters used in various studies we have discussed so far

42 Shallow Machine Learning Shallow machine learning(SML) algorithms include traditional and simple ML algo-rithms +ese algorithms usually consist of a few manytimes one hidden layer SML algorithms cannot extractfeatures from the input and features need to be definedbeforehand Model training can only be done after featureextraction SML algorithms and their application in trafficcongestion studies are discussed in this section and shown inFigure 6

421 Artificial Neural Network Artificial neural network(ANN) was developed mimicking the function of the hu-man brain to solve different nonlinear problems It is a first-order mathematical or computational model that consists ofa set of interconnected processors or neurons Figure 7shows a simple ANN structure Due to its easy imple-mentation and efficient forecasting ability ANN has becomepopular in the field of traffic congestion prediction researchHopfield network feedforward network and back-propagation are the examples of ANN Feedforward neuralnetwork (FNN) is the simplest NN where the input data goto the hidden layer and from there to the output layerBackpropagation neural network (BPNN) consists of feed-forward and weight adjustment of the layers and is the mostcommonly applied ANN in transportation management Xuet al [31] applied BPNN to predict traffic flow thus toevaluate congestion factor in their study +ey proposedoccupancy-based congestion factor (CRO) evaluationmethod with three other evaluated congestion factors basedon mileage ratio of congestion (CMRC) road speed (CRS)and vehicle density (CVD) +ey also evaluated the effect ofdata-size on real-time rendering of road congestionComplex road network with higher interconnectionsshowed higher complication in simulation and rendering+e advantage of the proposed model was that it took littleprocessing time for high sampling data rendering +emodel can be used as a general congestion prediction modelfor different road networks Some used hybrid NN forcongestion prediction Nadeem and Fowdur [11] predictedcongestion in spatial space applying the combination of oneof six SML algorithms with NN Six SML algorithms in-cluded moving average (MA) autoregressive integratedmoving average (ARIMA) linear regression second- andthird-degree polynomial regression and k-nearest neigh-bour (KNN) +e model showing the least RMSE value wascombined with BPNN to form hybrid NN +e hidden layerhad seven neurons which was determined by trial and errorHowever it was a very preliminary level work It did notshow the effect of data increment in the accuracy

Unlike the previous studies those focused on traffic flowparameters to conduct traffic congestion prediction re-search Ito and Kaneyasu [60] analysed driversrsquo behaviour inpredicting congestion +ey showed that vehicle operatorsact differently on different phases of the journey +ey usedone layered BPNN to learn the behaviour of female driversand extract travel phase according to that +e resultsshowed an average efficiency of 82 in distinguishing thetravel phase

ANN is a useful machine learning model which has aflexible structure +e neurons of the layer can be adaptedaccording to the input data As mentioned above a generalmodel can be developed and applied for different road typesby using the advantage of nonlinearity capturing ability ofANN However ANN requires larger datasets than theprobabilistic reasoning models which results in highcomplexity

ANN shows great potential in diverse parameter anal-ysis ANN is the only model that has recently been appliedfor driver behaviour analysis for traffic congestion ANN ispopular in every section of transport- traffic flow prediction[61 62] congestion control [63] driver tiredness [64] andvehicle noise [65 66]

422 RegressionModel Regression is a statistical supervisedML algorithm It models the prediction real numberedoutput value based on the independent input numericalvariable Regression models can be further dividedaccording to the number of input variables +e simplestregression model is linear regression with one input featureWhen the feature number increases the multiple regressionmodel is generated

Jiwan et al [27] developed a multiple linear regressionanalysis (MLRA) model using weather data and trafficcongestion data after preprocessing using Hadoop At first asingle regression model was developed for all the variablesusing R After a 3-fold reduction process only ten variableswere determined to form the final MLRA model Zhang andQian [22] conducted an interesting approach to predictmorning peak hour congestion using household electricityusage patterns +ey used LASSO regression to correlate thepattern features using the advantage of linearly relatedcritical feature selection capability

On the other hand Jain et al [33] developed both linearand exponential regression model using IBM SPSS softwareto find the relevant variables +e authors converted het-erogenous vehicles into passenger car unit (PCU) for sim-plification +ree independent variables were considered toestimation origin-destination- (O-D-) based congestionmeasures +ey used PCC to evaluate the correlation amongthe parameters However simply averaging O-D node pa-rameters may not provide the actual situation of dynamictraffic patterns

Regression models consist of some hidden coefficientswhich are determined in the training phase +e most ap-plied regression model is the autoregressive integratedmoving average (ARIMA) ARIMA has three parameters- pd and q ldquoprdquo is the auto regressive order that refers to how


many lags of the independent variable needs to be con-sidered for prediction Moving average order ldquoqrdquo presentsthe lag prediction error numbers Lastly ldquodrdquo is used to make

the time-series stationary Alghamdi et al [67] took d as 1 asone differencing order could make the model stationaryNext they applied the autocorrelation function (ACF) and

Support vector machine


Artificial neural network Regression model Decision tree

Figure 6 Subdivision of shallow machine learning models

Table 1 Traffic congestion prediction studies in probabilistic reasoning

Methodology Road type Datasource Input parameters Target domain No of congestion

state levelslowast Reference

Hierarchical fuzzy rule-based system Highway corridor Sensor

Occupancy Speed 2 Zhang et al [30]

Speed Speed

4

Lopez-garciaet al [37]

Evolutionary fuzzy rulelearning Traffic flow Traffic density Onieva et al

[28]

Mamdani-type fuzzylogic inference

Highway trunkroad branch road mdash Speed

Congestion Index

Cao and Wang[3]

Density

Wang et al [58]Fuzzy inference Highway corridor CameraTravel timeTraffic flow

SpeedFuzzy comprehensiveevaluation Highway corridor Probe Traffic volume Saturation 5 Kong et al [4]

Speed Density speed Yang et al [5]

Hidden Markov model

Highway network SensorEmission matrix Traffic pattern

selection mdash Zaki et al [32]

Emission matrix Traffic patterndetermination mdash Zaki et al [25]Transition matrix

Main road Probe

Observationprobability Mapping GPS data mdash Sun et al [45]Transitionprobability

Gaussian distribution Highway corridor Sensor Traffic volume Optimal featureselection mdash Yang [29]

Bayesian network

Build-up area Simulation Road and busincrement

Congestionprobability mdash

Yi Liu et al [59]

Bridge Sensor

IntensityAsencio-Cortes

et al [54]Occupation

Average speedAverage distance

Highway network Sensor

Networkdirection

Congestionprobability

Kim and Wang[34]

Day and timeweatherIncidentsTraffic flowOccupancy

SpeedLevel of serviceCongestion state

Extended Kalman filter Highway Camera Travel time Data fusion mdash Adetiloye andAwasthi [7]

+e table accumulates the data source scope of the study area input and resulting parameters and how many cognitive traffic states were considered in thestudieslowast2 freecongested 4 freelightmediumsevere 5 very freefreelightmediumsevere


the partial autocorrelation function (PACF) along with theminimum information criteria matrix to determine thevalues of p and q +ey only took the time dimension intoaccount However the results inclined with the true patternfor only one week and needed to be fine-tuned consideringprediction errors Besides the study did not consider thespatial dimension

Regression models are useful to be applied for time seriesproblems +erefore regression models are suitable fortraffic forecasting problems However these models are notreliable for nonlinear rapidly changing the multidimensiondataset +e results need to be modified according to pre-diction errors

However as already and further will be discussed in thisarticle most of the studies used different regression modelsto validate their proposed model [6 11 25 68 69]

With the increment of dataset and complexity associatedwith it regression models are becoming less popular intraffic congestion prediction Currently regression modelsare frequently used by modifying with other machinelearning algorithms eg ANN and kernel functions Someother sectorsrsquo regression models are applied including hy-brid ARIMA in traffic speed prediction for specific vehicletype (Wang et al [70] traffic volume prediction [71] andflow prediction applying modified ARIMA [72]

423 Decision Tree A decision tree is a model that predictsan output based on several input variables +ere are twotypes of trees the classification tree and the regression treeWhen these two trees merge a new tree named classificationand regression tree (CART) generates Decision tree uses thefeatures extracted from the entire dataset Random forest is asupervised ML classification algorithm that is the average ofmultiple decision tree results +e features are randomlyused while developing decision trees It uses a vast amount ofCART decision trees +e decision trees vote for the pre-dicted class in a random forest model

Wang et al [9] proposed a probabilistic method ofexploiting information theory tools of entropy and Fanorsquosinequality to predict road traffic pattern and its associatedcongestion for urban road segments with no prior knowl-edge on the O-D of the vehicle +ey incorporated roadcongestion level into time series for mapping the vehiclestate into the traffic conditions As interval influenced the

predictability an optimal segment length and velocity wasfound However with less available data an increasednumber of segments increased the predictability Anothertraffic parameter travel time was used to find CI by Liu andWu [73] +ey applied the random forest ML algorithm toforecast traffic congestion states At first they extracted 100sample sets to construct 100 decision trees by using boot-strap +e number of feature attributes was determined asthe square root of the total number of features Chen et al[16] also applied the CART method for prediction andclassification of traffic congestion +e authors appliedMoranrsquos I method to analyse the spatiotemporal correlationamong different road network traffic flow +e modelshowed effectiveness compared with SVM and K-meansalgorithm

Decision tree is a simple classification problem-solvingmodel that can be applied for multifeature data eg Liuand Wu [73] applied weather condition road conditiontime period and holiday as the input variables +ismodelrsquos knowledge can be represented in the form of IF-THEN rules making it an easily interpretable problem It isalso needed to be kept in mind that the classification resultsare usually binary and therefore not suitable where thecongestion level is required to be known Other sectors oftransport where decision tree models applied are trafficprediction [74] and traffic signal optimisation with Fuzzylogic [75]

424 Support Vector Machine +e support vector machine(SVM) is a statistical machine learning method +e mainidea of this model is to map the nonlinear data to a higherdimensional linear space where data can be linearly classifiedby hyperplane [1] +erefore it can be very useful in trafficflow pattern identification for traffic congestion predictionTseng et al [13] determined travel speed in predicting real-time congestion applying SVM +ey used Apache Storm toprocess big data using spouts and bolts Traffic weathersensors and events collected from social media of closeproximity were evaluated together by the system +eycategorised vehicle speed into classes and referred them aslabels Speed of the previous three intervals was used to trainthe proposed model However the congestion level cat-egororised from 0 to 100 does not carry a specific knowledgeof the severity of the level especially to the road usersIncrement in training data raised accuracy and computa-tional time +is may ultimately make it difficult to makereal-time congestion prediction

Traffic flow shows different patterns based on the trafficmixture or time of the day SVM is applied to identify theappropriate pattern Currently modified SVM mostly has itsapplication in other sectors as well eg freeway exitingtraffic volume prediction [58] traffic flow prediction [76]and sustainable development of transportation and ecology[77]

Most of the studies compared their developed modelwith SVM [22 78 79] Deep machine learning (DML) al-gorithms showed better results compared to SVM Table 2refers to the studies under this section

Output layer

Hidden layer

Input layer

Figure 7 A simple ANN structure


43 Deep Machine Learning DML algorithms consist ofseveral hidden layers to process nonlinear problems +emost significant advantage of these algorithms is they canextract features from the input data without any priorknowledge Unlike SML feature extraction and modeltraining are done together in these algorithms DML canconvert the vast continuous and complex traffic data withlimited collection time horizon into patterns or featurevectors From last few years DML has become popular intraffic congestion prediction studies Traffic congestionstudies that used DML algorithms are shown in Figure 8 anddiscussed in this section

431 Convolutional Neural Network Convolutional neuralnetwork (CNN) is a commonly applied DML algorithm intraffic engineering Due to the excellent performance ofCNN in image processing while applying in traffic pre-diction traffic flow data are converted into a 2-D matrix toprocess +ere are five main parts of a CNN structure intransportation the input layer convolution layer poollayer full connection layer and output layer Both theconvolution and pooling layer extracts important features+e depth of these two layers differs in different studiesMajority of the studies converted traffic flow data into animage of a 2-Dmatrix In the studies performed byMa et al[80] and Sun et al [45] each component of the matrixrepresented average traffic speed on a specific part of thetime While tuning CNN parameters they selected aconvolutional filter size of (3 times 3) and max-pooling of size(2 times 2) of 3 layers according to parameter settings of LeNetand AlexNet and loss of information measurement

Whereas Chen et al [68] used a five-layered convolution offilter size of (2 times 2) without the pooling layer +e authorsapplied a novel method called convolution-based deepneural network modelling periodic traffic data (PCNN)+e study folded the time-series to generate the inputcombining real-time and historical traffic data To capturethe correlation of a new time slot with the immediate pastthey duplicated the congestion level of the last slot in thematrix Zhu et al [49] also applied five convolution-pooling layers as well as (3 times 3) and (2 times 2) sizes respec-tively Along with temporal and spatial data the authorsalso incorporated time interval data to produce a 3-D inputmatrix Unlike these studies Zhang et al [6] preprocessedthe raw data by performing a spatiotemporal cross-cor-relation analysis of traffic flow sequence data using PCC+en they applied a model named spatiotemporal featureselection algorithm (STFSA) on the traffic flow sequencedata to select the feature subsets as the input matrix A 2-layered CNN with the convolutional and pooling size assame as the previous studies was used However STFSAconsiders its heuristics biases and trade-offs and does notguarantee optimality

CNN shows good performance where a large dataset isavailable It has excellent feature learning capability with lesstime-consuming classification ability +erefore CNN canbe applied where the available dataset can be converted intoan image CNN is applied in traffic speed prediction [81]traffic flow prediction [6] and modified CNN with LSTM isalso applied for traffic prediction [82] However as men-tioned above no model depth and parameter selectionstrategies are available

Table 2 Traffic congestion prediction studies in shallow machine learning

Methodology Road type Datasource Input parameters Target domain

No ofcongestionstate levels

Reference

Artificial neuralnetwork

Road network Sensor Occupancy Congestion factor 3 Xu et al [31]Simulation Density

Highwaycorridor

Sensor Distance Speed 2Nadeem and Fowdur

[11]Speed

SimulationSpeed Traffic congestion

state 2 Ito and Kaneyasu[60]+rottle opening

Steering input angle

Regression model

Highwaycorridor Sensor

Temperature

Traffic congestionscore mdash Jiwan et al [27]

HumidityRainfall

Traffic speedTime

Arterial road Camera Volume Congestion Index 4 Jain et al [33]Subarterial road Speed

Decision treeRing road

ProbeAverage speed Traffic predictability mdash Wang et al [9]

Road network Speed Moran index 5 Chen et al [16]Trajectory

Support vectormachine


Speed

Travel speed mdash Tseng et al [13]Density

Traffic volumedifference

Rainfall volume


432 Recurrent Neural Network Recurrent neural network(RNN) has a wide usage in the sequential traffic data pro-cessing by considering the influence of the related neighbour(Figure 9) Long short-term memory (LSTM) is a branch ofRNN In the hidden layer of LSTM there is a memory blockthat includes four NN layers which stores and regulates theinformation flow In recent years with different data col-lection systems with extended intervals LSTM has becomepopular Due to this advantage Zhao et al [12] developed anLSTM model consisted of three hidden layers and tenneurons using long interval data+ey set an adequate targetand fine-tuned the parameters until the training modelstabilized+e authors also applied the congestion index andclassification (CI-C) model to classify the congestion bycalculating CI from LSTM output data Most of the studiesuse the equal interval of CI to divide congestion states +isstudy did two more intervals of natural breakpoint andgeometric interval to find that the latest provided the mostinformation from information entropy Lee et al [69] ap-plied 4 layers with 100 neurons LSTM model of 3D matrixinput +e input matrix element contained a normalisedspeed to shorten the training time While eliminating thedependency the authors found that a random distribution oftarget road speed and more than optimally connected roadsin the matrix reduced the performance To eliminate thelimitation of temporal dependency Yuan-Yuan et al [79]trained their model in the batch learning approach +einstance found from classifying test dataset was used to trainthe model in an online framework Some studies introducednew layers to modify the LSTMmodel for feature extractionZhang et al [83] introduced an attention mechanism layerbetween LSTM and prediction layer that enabled the featureextraction from a traffic flow data sequence and captured theimportance of a traffic state Di et al [84] introducedconvolution that provides an input to the LSTM model toform the CPM-ConvLSTM model All the studies appliedthe one-hot method to convert the input parameters Adamstochastic gradient descent (SGD) and leakage integral echostate network (LiESN) are a few optimisation methodsapplied to fine-tune the outcome

A few studies combined RNN with other algorithmswhile dealing with vast parameters of the road network Inthis regard Ma et al [85] applied the RNN and restrictedBoltzmann machine (RNN-RBM) model for networkwidespatiotemporal congestion prediction Here they usedconditional RBM to construct the proposed deep architec-ture which is designed to process the temporal sequence byproviding a feedback loop between visible layer and hidden

layer +e congestion state was analysed from traffic speedand was represented in binary format in a matrix as inputAlso Sun et al [45] combined RNN of three hidden layerswith its two other variants LSTM and gated recurrent unit(GRU) +e hidden layers included the memory blockcharacteristic of LSTM and the cell state and hidden statewere incorporated by GRU

As the sample size is increasing vastly RNN is becomingpopular as a current way of modelling RNN has a short-term memory +is characteristic of RNN helps to modelnonlinear time series data +e training of RNN is alsostraight forward similar to multilayer FNN However thistraining may become difficult due to the conversion in adeep architecture with multiple layers in long-term de-pendency In case of long-term dependency problemsLSTM is becoming more suitable to be applied as LSTM canremember information for a long period of time RNN hasits application in other sectors of the transport too egtraffic passenger flow prediction [86] modified LSTM in realtime crash prediction [87] and road-network traffic pre-diction [88]

433 Extreme Learning Machine In recent years a novellearning algorithm called the extreme learning machine(ELM) is proposed for training the single layer feed-forwardneural network (SLFN) In ELM input weights and hiddenbiases are assigned randomly instead of being exhaustivelytuned +erefore ELM training is fast +erefore taking thisadvantage into account Ban et al [19] applied the ELMmodel for real-time traffic congestion prediction +ey de-termine CI using the average travel speed A 4-fold cross-validation was done to avoid noise in raw data +e model

Output layer

Output layer

Output layer

Figure 9 A simple RNN structure


Convolutional neural network

Long short-term memory

Extreme learning machine

Figure 8 Subdivision of deep machine learning models


found optimal hidden nodes to be 200 in terms of com-putational cost in the study An extension of this study wasdone by Shen et al [78] and Shen et al [89] by applying akernel-based semisupervised extreme learning machine(kernel-SSELM) model +is model can deal with theunlabelled data problem of ELM and the heterogenous datainfluence +e model integrated small-scaled labelled data oftransportation personnel and large-scaled unlabelled trafficdata to evaluate urban traffic congestion ELM speeded upthe processing time where kernel function optimized theaccuracy and robustness of the whole model However real-time labelled data collection was quite costly in terms ofhuman resources and working time and the number ofexperts for congestion state evaluation should have beenmore Another modification of EML was applied by Yiminget al [20] +ey applied asymmetric extreme learning ma-chine cluster (S-ELM-cluster) model for short-term trafficcongestion prediction by determining the CI +e authorsdivided the study area and implemented submodels pro-cessing simultaneously for fast speed

+e ELMmodel has the advantage in processing large scaledata learning at high speed ELM works better with labelleddata Where both labelled and unlabelled data are availablesemisupervised ELM has shown good prediction accuracy as itwas seen from the studies Other sectors where ELM wasapplied included air traffic flow prediction [90] traffic flowprediction [91] and traffic volume interval prediction [92]

Other than the models already discussed Zhang et al [93]proposed a deep autoencoder-based neural network modelwith symmetry of four layers for the encoder and the decoder

to learn temporal correlations of a transportation network+e first component encoded the vector representation ofhistorical congestion levels and their correlation +ey thendecoded to build a representation of congestion levels for thefuture +e second component of DCPS used two denselayers those converted the output from the decoder to cal-culate a vector representation of congestion level Howeverthe process lost information as the congestion level of all thepixels was averaged +is approach needed high iteration andwas computationally expensive as all the pixels regardless ofroads were considered Another study applied a generalisedversion of recurrent neural network named recursive neuralnetwork +e difference between these two is in recurrentNN weights are shared along the data sequence Whereasrecursive NN is a single neuron model therefore weights areshared at every node Huang et al [94] applied a recursive NNalgorithm named echo state network (ESN) +is modelconsists of an input layer reservoir network and output layer+e reservoir layer constructs the rules that connected pre-diction origin and forecasting horizon As the study took alarge study area with vast link number they simplified thetraining rule complexity applying recursive NN Table 3summarises some studies

5 Discussion and Research Gaps

Research in traffic congestion prediction is increasing ex-ponentially Among the two sources most of the studiesused stationary sensorcamera data Although sensor datacannot capture the dynamic traffic change frequent change

Table 3 Traffic congestion prediction studies in deep machine learning


state levels Reference

Convolutional neuralnetworks

Road network Probe Average traffic speedSpeed 3 Ma et al [80]

Average trafficspeed 5 Sun et al [45]

Camera Congestion level Congestion level 3 Chen et al [68]Highwaycorridor Sensor Traffic flow Traffic flow mdash Zhang et al [93]

Recurrent neuralnetwork

Road section Probe Weather data Congestion time 5 Zhao et al [12]Congestion time

Arterial road Online Congestion level Congestion level

3

Yuan-Yuan et al[79]

Road networkCamera Spatial similarity

feature Speed Lee et al [69]Sensor SpeedSurvey Peak hour


SpeedCongestion level 4 Zhang et al [83]Travel time

Volume

Road network Probe

Congestion state Congestion state 2 Ma et al [85]

Extreme learningmachine

Current time

CongestionIndex mdash Ban et al [19]

Road traffic statecluster

Last congestionindex

Road typeNumber of adjacent

roads


in source makes it complicated to evaluate the flow patternsfor probe data [95] Data collection horizon is an importantfactor in traffic congestion studies +e small horizon of afew days [3] cannot capture the actual situation of thecongestion as traffic is dynamic Other studies that used datafor a few months showed the limitation of seasonality[22 67]

+e condition of the surrounding plays an importantfactor in traffic congestion A few studies focused on thesefactors Two studies considered social media contribution ininput parameter [7 13] and five considered weather con-dition [12 13 27 34 73] Events eg national event schoolholiday and popular sports events play a big role in trafficcongestion For example Melbourne Australia has twopublic holidays before and during two most popular sportsevents of the country +e authorities close a few trafficroutes to tackle the traffic and the parade resulting in trafficcongestion +erefore more focus must be put in includingthese factors while forecasting

Dealing with missing data is a challenge in the dataprocessing Some excluded the respective data altogether[29] others applied different methods to retrieve the data[59 85] and some replaced with other data [45] Missingdata imputation can be a useful research scope in trans-portation engineering

Machine learning algorithm especially DML models isdeveloped with time+is shows a clear impact on the rise oftheir implementation in traffic congestion forecasting(Figure 10)

Probabilistic reasoning algorithms were mostly appliedfor a part of the prediction model eg map matching andoptimal feature number selection Fuzzy logic is the mostwidely used algorithm in this class of algorithms From otherbranches ANN and RNN are the mostly applied models Mostof the studies that applied hybrid or ensembled models belongto probabilistic and shallow learning class Only two studiesapplied hybrid deep learning models while predicting net-workwide congestion Tables 4 5ndash6 summarize the advantageand weaknesses of the algorithms of different branches

Among all DML models RNN is more suitable for timeseries prediction In a few studies RNN performed betterthan CNN as the gap between the traffic speeds in differentclasses was very small [12 69] However due to little re-search in traffic congestion field a lot of new ML algorithmsare yet to be applied

SML models showed better results than DML whileforecasting traffic congestion in the short-term as SML canprocess linearity efficiently and linear features have morecontribution to traffic flow in short-term All the short-termforecasting studies discussed in this article applying SMLshowed promising results At the same time DML modelsshowed good accuracy as these models can handle both

linear and nonlinear features efficiently Besides real-timecongestion prediction cannot afford high computation time+erefore models taking a short computational time aremore effective in this case

6 Future Direction

Traffic congestion is a promising area of research +ere-fore there are multiple directions to conduct in futureresearch

Numerous forecasting models have already been appliedin road traffic congestion forecasting However with thenewly developed forecasting models there is more scope tomake the congestion prediction more precise Also in thisera of information the use of increased available traffic databy applying the newly developed forecasting models canimprove the prediction accuracy

+e semisupervised model was applied only for the EMLmodel Other machine learning algorithms should be ex-plored for using both labelled and unlabelled data for higherprediction accuracy Also a limited number of studies havefocused on real-time congestion forecasting In future re-searches should pay attention to real-time traffic congestionestimation problem

Another future direction can be focusing on the levelof traffic congestion A few studies have divided the trafficcongestion into a few states However for better trafficmanagement knowing the grade of congestion is essen-tial +erefore future researches should focus on thisBesides most studies focused on only one traffic pa-rameter to forecast congestion for congestion prediction+is can be an excellent future direction to give attentionto more than one parameter and combining the resultsduring congestion forecasting to make the forecastingmore reliable

Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014

Figure 10 Application of AI models with time


Table 5 +e strength and weakness of the models of shallow machine Learning

Methodology Advantages Disadvantages


(i) It is an adaptive system that can change structurebased on inputs during the learning stage [96]

(i) BPNN requires vast data for training the model due to theparameter complexity resulting from its parameternonsharing technique [97]

(ii) It features defined early FNN shows excellentefficiency in capturing the nonlinear relationship ofdata

(ii) +e training convergence rate of the model is slow

Regression model

(i) Models are suitable for time series problems (i) Linear models cannot address nonlinearity making itharder to solve complex prediction problems

(ii) Traffic congestion forecasting problems can beeasily solved (ii) Linear models are sensitive to outliers

(iii) ARIMA can increase accuracy by maintainingminimum parameters (iii) Computationally expensive

(iv) Minimum complexity in the model (iv) ARIMA cannot deal multifeature dataset efficiently(v) ARIMA cannot capture the rapidly changing traffic flow[8]


(i) It is efficient in pattern recognition andclassification

(i) +e improperly chosen kernel function may result in aninaccurate outcome

(ii) A universal learning algorithm that can diminishthe classification error probability by reducing thestructural risk [1]

(ii) Unstable traffic flow requires improved predictionaccuracy of SVM

(iii) It does not need a vast sample size (iii) It takes high computational time and memory

Table 4 +e strength and weakness of the models of probabilistic reasoning


Fuzzy logic

(i) It converts the binary value into the linguisticdescription hence portraying the traffic congestionstate

(i) No appropriate membership function shape selectionmethod exists

(ii) iIt can portray more than two states (ii) Traffic pattern recognition capability is not as durable asML algorithms

(iii) As it does not need an exact crisp input it can dealwith uncertainty

(iii) Traffic state may not match the actual traffic state as theoutcome is not exact

Hidden Markovmodel

(i) +e model can overcome noisy measurements (i) Accuracy decreases with scarce temporal probe trajectorydata

(ii) Can efficiently learn from non-preprocessed data (ii) Not suitable in case of missing dataset(iii) Can evaluate multiple hypotheses of the actualmapping simultaneously

Gaussian mixturemodel

(i) Can do traffic parameter distribution over a periodas a mixture regardless of the traffic state

(i) Optimization algorithm used with GMM must be chosencautiously

(ii) Can overcome the limitation of not being able toaccount for multimodal output by a single Gaussianprocess

(ii) Results may show wrong traffic patterns due to localoptima limitation and lack of traffic congestion thresholdknowledge of the optimisation algorithm

Bayesian network

(i) It can understand the underlying relationshipbetween random variables (i) Computationally expensive

(ii) It can model and analyse traffic parametersbetween adjacent road links (ii) +e model performs poorly with the increment in data

(iii) +e model can work with incomplete data (iii) +e model represents one-directional relation betweenvariables only


7 Conclusions

Traffic congestion prediction is getting more attention fromthe last few decades With the development of infrastructureevery country is facing traffic congestion problem +ere-fore forecasting the congestion can allow authorities tomake plans and take necessary actions to avoid it +edevelopment of artificial intelligence and the availability ofbig data have led researchers to apply different models in thisfield +is article divided the methodologies in three classedAlthough probabilistic models are simple in general theybecome complex while different factors that affect trafficcongestion eg weather social media and event areconsidered Machine learning especially deep learning hasthe benefit in this case +erefore deep learning algorithmsbecame more popular with time as they can assess a largedataset However a wide range of machine learning algo-rithms are yet to be applied +erefore a vast opportunity ofresearch in the field of traffic congestion prediction stillprevails

Conflicts of Interest

+e authors declare that they have no conflicts of interest

Acknowledgments

+e authors would like to thank RMIT University and theAustralian Government Research Training Program (RTP)for the financial support

References

[1] Z Shi Advanced Artificial Intelligence World ScientificSingapore 2011

[2] M S Ali M Adnan S M Noman and S F A BaquerildquoEstimation of traffic congestion cost-A case study of a majorarterial in karachirdquo Procedia Engineering vol 77 pp 37ndash442014

[3] W Cao and J Wang ldquoResearch on traffic flow congestionbased on Mamdani fuzzy systemrdquo AIP Conference Proceed-ings vol 2073 2019

[4] X Kong Z Xu G Shen J Wang Q Yang and B ZhangldquoUrban traffic congestion estimation and prediction based onfloating car trajectory datardquo Future Generation ComputerSystems vol 61 pp 97ndash107 2016

[5] Q Yang J Wang X Song X Kong Z Xu and B ZhangldquoUrban traffic congestion prediction using floating car tra-jectory datardquo in Proceedings of the International Conference onAlgorithms and Architectures for Parallel Processing pp 18ndash30 Springer Zhangjiajie China November 2015

[6] W Zhang Y Yu Y Qi F Shu and Y Wang ldquoShort-termtraffic flow prediction based on spatio-temporal analysis andCNN deep learningrdquo Transportmetrica A Transport Sciencevol 15 no 2 pp 1688ndash1711 2019

[7] T Adetiloye and A Awasthi ldquoMultimodal big data fusion fortraffic congestion predictionrdquoMultimodal Analytics for Next-Generation Big Data Technologies and Applications SpringerBerlin Germany pp 319ndash335 2019

[8] F Wen G Zhang L Sun X Wang and X Xu ldquoA hybridtemporal association rules mining method for traffic con-gestion predictionrdquo Computers amp Industrial Engineeringvol 130 pp 779ndash787 2019

[9] J Wang Y Mao J Li Z Xiong and W-X Wang ldquoPre-dictability of road traffic and Congestion in urban areasrdquoPLoS One vol 10 no 4 Article ID e0121825 2015

[10] Z He G Qi L Lu and Y Chen ldquoNetwork-wide identificationof turn-level intersection congestion using only low-fre-quency probe vehicle datardquo Transportation Research Part CEmerging Technologies vol 108 pp 320ndash339 2019

[11] K M Nadeem and T P Fowdur ldquoPerformance analysis of areal-time adaptive prediction algorithm for traffic conges-tionrdquo Journal of Information and Communication Technologyvol 17 no 3 pp 493ndash511 2018

[12] H Zhao X Jizhe L Fan L Zhen and L Qingquan ldquoA peaktraffic Congestion prediction method based on bus drivingtimerdquo Entropy vol 21 no 7 p 709 2019

[13] F-H Tseng J-H Hsueh C-W Tseng Y-T YangH-C Chao and L-D Chou ldquoCongestion prediction with bigdata for real-time highway trafficrdquo IEEE Access vol 6pp 57311ndash57323 2018

Table 6 +e strength and weakness of the models of deep machine learning



(i) Capable of learning features from local connections andcomposing them into high-level representation

(i) Computationally expensive as a huge kernel isneeded for feature extraction

(ii) Classification is less time-consuming (ii) A vast dataset is required(iii) Can automatically extract features (iii) Traffic data needs to be converted to an image

(iv) No available strategies are available on CNNmodel depth and parameter selection


(i) Shows excellent performance in processing sequentialdata flow

(i) Long-term dependency results in badperformance

(ii) Efficient in sequence classification (ii) No available firm guideline in dependencyelimination

(iii) Efficient in processing time-series with long intervalsand postponements


(i) Fast learning speed (i) Training time increases with the hidden noderise

(ii) Can avoid local minima (ii) Unlabelled data problem(iii) Modified models are available to deal with an unlabelleddata problem (iii) May produce less accurate results


[14] D Li J Deogun W Spaulding and B Shuart ldquoTowardsmissing data imputation a study of fuzzy K-means Clus-tering methodrdquo in Rough Sets and Current Trends inComputing S Tsumoto R Słowinski J Komorowski andJ W Grzymała-Busse Eds pp 573ndash579 Springer BerlinGermany 2004

[15] Y Yang Z Cui JWu G Zhang and X Xian ldquoFuzzy c-meansclustering and opposition-based reinforcement learning fortraffic congestion identificationrdquo Journal of Information ampComputational Science vol 9 no 9 pp 2441ndash2450 2012

[16] Z Chen Y Jiang D Sun and X Liu ldquoDiscrimination andprediction of traffic congestion states of urban road networkbased on spatio-temporal correlationrdquo IEEE Access vol 8pp 3330ndash3342 2020

[17] Y Guo L Yang S Hao and J Gao ldquoDynamic identificationof urban traffic congestion warning communities in hetero-geneous networksrdquo Physica A Statistical Mechanics and ItsApplications vol 522 pp 98ndash111 2019

[18] A Elfar A Talebpour and H S Mahmassani ldquoMachinelearning approach to short-term traffic congestion predictionin a connected environmentrdquo Transportation Research Re-cord Journal of the Transportation Research Board vol 2672no 45 pp 185ndash195 2018

[19] X Ban C Guo and G Li ldquoApplication of extreme learningmachine on large scale traffic congestion predictionrdquo Pro-ceedings of ELM-2015 vol 1 pp 293ndash305 2016

[20] X Yiming B Xiaojuan L Xu and S Qing ldquoLarge-scale trafficCongestion prediction based on the symmetric extremelearning machine Cluster fast learning methodrdquo Symmetryvol 11 no 6 p 730 2019

[21] Y Zheng Y Li C-M Own Z Meng and M Gao ldquoReal-timepredication and navigation on traffic congestion model withequilibrium Markov chainrdquo International Journal of Dis-tributed Sensor Networks vol 14 no 4 Article ID155014771876978 2018

[22] P Zhang and Z Qian ldquoUser-centric interdependent urbansystems using time-of-day electricity usage data to predictmorning roadway congestionrdquo Transportation Research PartC Emerging Technologies vol 92 pp 392ndash411 2018

[23] P Mishra R Hadfi and T Ito ldquoAdaptive model for trafficcongestion predictionrdquo in Proceedings of the InternationalConference on Industrial Engineering and Other Applicationsof Applied Intelligent Systems vol 9799 pp 782ndash793 Mor-ioka Japan 2016

[24] F Li J Gong Y Liang and J Zhou ldquoReal-time congestionprediction for urban arterials using adaptive data-drivenmethodsrdquo Multimedia Tools and Applications vol 75 no 24pp 17573ndash17592 2016

[25] J F Zaki A Ali-Eldin S E Hussein S F Saraya andF F Areed ldquoTraffic congestion prediction based on HiddenMarkov Models and contrast measurerdquo Ain Shams Engi-neering Journal vol 11 no 3 p 535 2020

[26] P Jiang L Liu L Cui H Li and Y Shi ldquoCongestion pre-diction of urban traffic employing SRBDPrdquo in Proceedings ofthe 2017 IEEE International Symposium on Parallel andDistributed Processing with Applications and 2017 IEEE In-ternational Conference on Ubiquitous Computing and Com-munications (ISPAIUCC) pp 1099ndash1106 IEEE GuangzhouChina 2017

[27] L Jiwan H Bonghee L Kyungmin and J Yang-Ja ldquoAprediction model of traffic congestion using weather datardquo inProceedings of the 2015 IEEE International Conference on DataScience and Data Intensive Systems pp 81ndash88 Sydney NSWAustralia December 2015

[28] E Onieva P Lopez-Garcia A D Masegosa E Osaba andA Perallos ldquoA comparative study on the performance ofevolutionary fuzzy and Crisp rule based Classificationmethods in Congestion predictionrdquo Transportation ResearchProcedia vol 14 pp 4458ndash4467 2016

[29] S Yang ldquoOn feature selection for traffic congestion predic-tionrdquo Transportation Research Part C Emerging Technologiesvol 26 pp 160ndash169 2013

[30] X Zhang E Onieva A Perallos E Osaba and V C S LeeldquoHierarchical fuzzy rule-based system optimized with geneticalgorithms for short term traffic congestion predictionrdquoTransportation Research Part C Emerging Technologiesvol 43 pp 127ndash142 2014

[31] Y Xu L Shixin G Keyan Q Tingting and C XiaoyaldquoApplication of data science technologies in intelligent pre-diction of traffic Congestionrdquo Journal of Advanced Trans-portation 2019

[32] J Zaki A Ali-Eldin S Hussein S Saraya and F Areed ldquoTimeaware hybrid hidden Markov models for traffic Congestionpredictionrdquo International Journal on Electrical Engineeringand Informatics vol 11 no 1 pp 1ndash17 2019

[33] S Jain S S Jain and G Jain ldquoTraffic congestion modellingbased on origin and destinationrdquo Procedia Engineeringvol 187 pp 442ndash450 2017

[34] J Kim and G Wang ldquoDiagnosis and prediction of trafficCongestion on urban road networks using bayesian net-worksrdquo Transportation Research Record Journal of theTransportation Research Board vol 2595 no 1 pp 108ndash1182016

[35] W-X Wang R-J Guo and J Yu ldquoResearch on road trafficcongestion index based on comprehensive parameters takingDalian city as an examplerdquo Advances in Mechanical Engi-neering vol 10 no 6 Article ID 168781401878148 2018

[36] E Onieva V Milanes J Villagra J Perez and J GodoyldquoGenetic optimization of a vehicle fuzzy decision system forintersectionsrdquo Expert Systems with Applications vol 39no 18 pp 13148ndash13157 2012

[37] P Lopez-Garcia E Onieva E Osaba A D Masegosa andA Perallos ldquoA hybrid method for short-term traffic Con-gestion forecasting using genetic algorithms and Cross en-tropyrdquo IEEE Transactions on Intelligent TransportationSystems vol 17 no 2 pp 557ndash569 2016

[38] A Daissaoui A Boulmakoul and Z Habbas ldquoFirst specifi-cations of urban traffic-congestion forecasting modelsrdquo inProceedings of the 27th International Conference on Micro-electronics (ICM 2015) pp 249ndash252 Casablanca MoroccoDecember 2015

[39] M Collotta L Lo Bello and G Pau ldquoA novel approach fordynamic traffic lights management based on Wireless SensorNetworks and multiple fuzzy logic controllersrdquo Expert Sys-tems with Applications vol 42 no 13 pp 5403ndash5415 2015

[40] M B Trabia M S Kaseko and M Ande ldquoA two-stage fuzzylogic controller for traffic signalsrdquo Transportation ResearchPart C Emerging Technologies vol 7 no 6 pp 353ndash367 1999

[41] Y Zhang and Z Ye ldquoShort-term traffic flow forecasting usingfuzzy logic system methodsrdquo Journal of Intelligent Trans-portation Systems vol 12 no 3 pp 102ndash112 2008

[42] H Wang L Zheng and X Meng Traffic Accidents PredictionModel Based on Fuzzy Logic Springer Berlin Germany 2011

[43] Y Zhang and H Ge ldquoFreeway travel time prediction usingtakagi-sugeno-kang fuzzy neural networkrdquo Computer-aidedCivil and Infrastructure Engineering vol 28 no 8 pp 594ndash603 2013


[44] J Zhao ldquoResearch on prediction of traffic congestion staterdquo inProceedings of the MATEC Web of Conferences Les UlisFrance 2015

[45] S Sun J Chen and J Sun ldquoTraffic congestion predictionbased on GPS trajectory datardquo International Journal of Dis-tributed Sensor Networks vol 15 no 5 Article ID155014771984744 2019

[46] Y Qi and S Ishak ldquoA Hidden Markov Model for short termprediction of traffic conditions on freewaysrdquo TransportationResearch Part C Emerging Technologies vol 43 pp 95ndash1112014

[47] B Jiang and Y Fei ldquoTraffic and vehicle speed prediction withneural network and hidden markov model in vehicularnetworksrdquo in Proceedings of the 2015 IEEE Intelligent VehiclesSymposium (IV) pp 1082ndash1087 IEEE Seoul 2015

[48] G Zhu K Song P Zhang and L Wang ldquoA traffic flow statetransition model for urban road network based on HiddenMarkov Modelrdquo Neurocomputing vol 214 pp 567ndash5742016a

[49] L Zhu R Krishnan F Guo J Polak and A SivakumarldquoEarly identification of recurrent congestion in heterogeneousurban trafficrdquo in Proceedings of the 2019 IEEE IntelligentTransportation Systems Conference (ITSC) pp 4392ndash4397IEEE Auckland New Zealand October 2019

[50] Y Xie K Zhao Y Sun and D Chen ldquoGaussian processes forshort-term traffic volume forecastingrdquo Transportation Re-search Record Journal of the Transportation Research Boardvol 2165 no 1 pp 69ndash78 2010

[51] S Jin X Qu and D Wang ldquoAssessment of expressway trafficsafety using Gaussian mixture model based on time to col-lisionrdquo International Journal of Computational IntelligenceSystems vol 4 no 6 pp 1122ndash1130 2011

[52] J Jun ldquoUnderstanding the variability of speed distributionsunder mixed traffic conditions caused by holiday trafficrdquoTransportation Research Part C Emerging Technologiesvol 18 no 4 pp 599ndash610 2010

[53] S Shiliang Z Changshui and Y Guoqiang ldquoA bayesiannetwork approach to traffic flow forecastingrdquo IEEE Trans-actions on Intelligent Transportation Systems vol 7 no 1pp 124ndash132 2006

[54] G Asencio-Cortes E Florido A Troncoso and F Martınez-Alvarez ldquoA novel methodology to predict urban trafficcongestion with ensemble learningrdquo Soft Computing vol 20no 11 pp 4205ndash4216 2016

[55] Z Zhu B Peng C Xiong and L Zhang ldquoShort-term trafficflow prediction with linear conditional Gaussian Bayesiannetworkrdquo Journal of Advanced Transportation vol 50 no 6pp 1111ndash1123 2016

[56] S Wang W Huang and H K Lo ldquoTraffic parameters es-timation for signalized intersections based on combinedshockwave analysis and Bayesian Networkrdquo TransportationResearch Part C Emerging Technologies vol 104 pp 22ndash372019

[57] S Wang W Huang and H K Lo ldquoCombining shockwaveanalysis and Bayesian Network for traffic parameter estima-tion at signalized intersections considering queue spillbackrdquoTransportation Research Part C Emerging Technologiesvol 120 Article ID 102807 2020

[58] X Wang K An L Tang and X Chen ldquoShort term predictionof freeway exiting volume based on SVM and KNNrdquo Inter-national Journal of Transportation Science and Technologyvol 4 no 3 pp 337ndash352 2015

[59] Y Liu X Feng QWang H Zhang and XWang ldquoPredictionof urban road Congestion using a bayesian network

approachrdquo ProcediamdashSocial and Behavioral Sciences vol 138no C pp 671ndash678 2014

[60] T Ito and R Kaneyasu ldquoPredicting traffic congestion usingdriver behaviorrdquo in Proceedigs of the International Conferenceon Knowledge Based and Intelligent Information and Engi-neering Systems Marseille France 2017

[61] K Kumar M Parida and V K Katiyar ldquoShort term trafficflow prediction for a non urban highway using artificial neuralnetworkrdquo Procedia - Social and Behavioral Sciences vol 104pp 755ndash764 2013

[62] K Kumar M Parida and V K Katiyar ldquoShort term trafficflow prediction in heterogeneous condition using artificialneural networkrdquo Transport vol 30 no 4 pp 397ndash405 2013

[63] R More A Mugal S Rajgure R B Adhao andV K Pachghare ldquoRoad traffic prediction and congestioncontrol using artificial neural networksrdquo in Proceedings of the2016 International Conference on Computing Analytics andSecurity Trends (CAST) pp 52ndash57 IEEE Pune India De-cember 2016

[64] C Jacobe de Naurois C Bourdin A Stratulat E Diaz andJ-L Vercher ldquoDetection and prediction of driver drowsinessusing artificial neural network modelsrdquo Accident Analysis ampPrevention vol 126 pp 95ndash104 2019

[65] P Kumar S P Nigam and N Kumar ldquoVehicular traffic noisemodeling using artificial neural network approachrdquo Trans-portation Research Part C Emerging Technologies vol 40pp 111ndash122 2014

[66] V Nourani H Gokccedilekus I K Umar and H Najafi ldquoAnemotional artificial neural network for prediction of vehiculartraffic noiserdquo Science of De Total Environment vol 707p 136134 2020

[67] T Alghamdi K Elgazzar M Bayoumi T Sharaf and S ShahldquoForecasting traffic congestion using arima modelingrdquo inProceedings of the 2019 15th International Wireless Commu-nications amp Mobile Computing Conference (IWCMC)pp 1227ndash1232 IEEE Tangier Morocco October 2019

[68] M Chen X Yu and Y Liu ldquoPCNN deep convolutionalnetworks for short-term traffic congestion predictionrdquo IEEETransactions on Intelligent Transportation Systems vol 19no 11 pp 3550ndash3559 2018

[69] C Lee Y Kim S Jin et al ldquoA visual analytics system forexploring monitoring and forecasting road traffic Conges-tionrdquo IEEE Transactions on Visualization and ComputerGraphics vol 26 no 11 pp 3133ndash3146 2020

[70] H Wang L Liu S Dong Z Qian and HWei ldquoA novel workzone short-term vehicle-type specific traffic speed predictionmodel through the hybrid EMD-ARIMA frameworkrdquoTransportmetrica B Transport Dynamics vol 4 no 3pp 159ndash186 2016

[71] X F Wang X Y Zhang Q Y Ding and Z Q SunldquoForecasting traffic volume with space-time ARIMA modelrdquoAdvancedMaterials Research vol 156-157 pp 979ndash983 2010

[72] L Kui-Lin Z Chun-Jie and X Jian-Min ldquoShort-term trafficflow prediction using a methodology based on ARIMA andRBF-ANNrdquo in Proceedings of the 2017 Chinese AutomationCongress (CAC) pp 2804ndash2807 Jinan China October 2017

[73] Y Liu and HWu ldquoPrediction of road traffic congestion basedon random forestrdquovol 2 pp 361ndash364 in Proceedings of the2017 10th International Symposium on Computational Intel-ligence and Design (ISCID) vol 2 pp 361ndash364 IEEEHangzhou China December 2017

[74] W Alajali W Zhou S Wen and Y Wang ldquoIntersectiontraffic prediction using decision tree modelsrdquo Symmetryvol 10 no 9 p 386 2018


[75] M Balta and I Ozccedilelik ldquoA 3-stage fuzzy-decision tree modelfor traffic signal optimization in urban city via a SDN basedVANET architecturerdquo Future Generation Computer Systemsvol 104 pp 142ndash158 2020

[76] X Feng X Ling H Zheng Z Chen and Y Xu ldquoAdaptivemulti-kernel SVMwith spatial-temporal correlation for short-term traffic flow predictionrdquo IEEE Transactions on IntelligentTransportation Systems vol 20 no 6 pp 2001ndash2013 2019

[77] S Lu and Y Liu ldquoEvaluation system for the sustainabledevelopment of urban transportation and ecological envi-ronment based on SVMrdquo Journal of Intelligent amp FuzzySystems vol 34 no 2 pp 831ndash838 2018

[78] Q Shen X Ban and C Guo ldquoUrban traffic Congestionevaluation based on kernel the semi-supervised extremelearning machinerdquo Symmetry vol 9 no 5 p 70 2017

[79] C Yuan-Yuan Y Lv Z Li and F-YWang ldquoLong short-termmemory model for traffic congestion prediction with onlineopen datardquo in Proceedings of the 2016 IEEE 19th InternationalConference on Intelligent Transportation Systems (ITSC)pp 132ndash137 Rio de Janeiro Brazil December 2016

[80] X Ma Z Dai Z He J Ma Y Wang and Y Wang ldquoLearningtraffic as images a deep convolutional neural network forlarge-scale transportation network speed predictionrdquo Sensorsvol 17 no 4 p 818 2017

[81] R Ke W Li Z Cui and Y Wang ldquoTwo-stream multi-channel Convolutional neural network for multi-lane trafficspeed prediction Considering traffic volume impactrdquoTransportation Research Record Journal of the TransportationResearch Board vol 2674 no 4 pp 459ndash470 2020

[82] T Bogaerts A D Masegosa J S Angarita-Zapata E Onievaand P Hellinckx ldquoA graph CNN-LSTM neural network forshort and long-term traffic forecasting based on trajectorydatardquo Transportation Research Part C Emerging Technologiesvol 112 pp 62ndash77 2020

[83] T Zhang Y Liu Z Cui J LengW Xie and L Zhang ldquoShort-term traffic Congestion forecasting using attention-based longshort-termmemory recurrent neural networkrdquo in Proceedingsof the International Conference on Computational ScienceFaro Portugal June 2019

[84] X Di Y Xiao C Zhu Y Deng Q Zhao andW Rao ldquoTrafficcongestion prediction by spatiotemporal propagation pat-ternsrdquo in Proceedings of the 2019 20th IEEE InternationalConference on Mobile Data Management (MDM) pp 298ndash303 Hong Kong China June 2019

[85] X Ma H Yu Y Wang and Y Wang ldquoLarge-scale trans-portation network congestion evolution prediction using deeplearning theoryrdquo PloS One vol 10 no 3 Article ID e01190442015

[86] Z Zhene P Hao L Lin et al ldquoDeep convolutional mesh RNNfor urban traffic passenger flows predictionrdquo in Proceedings ofthe 2018 IEEE SmartWorld Ubiquitous Intelligence amp Com-puting Advanced amp Trusted Computing Scalable Computingamp Communications Cloud amp Big Data Computing Internet ofPeople and Smart City Innovation (SmartWorldSCALCOMUICATCCBDComIOPSCI) pp 1305ndash1310 IEEEGuangzhou China October 2018

[87] P Li M Abdel-Aty and J Yuan ldquoReal-time crash riskprediction on arterials based on LSTM-CNNrdquo AccidentAnalysis amp Prevention vol 135 Article ID 105371 2020

[88] W Xiangxue X Lunhui C Kaixun X Lunhui C Kaixunand C Kaixun ldquoData-Driven short-term forecasting for ur-ban road network traffic based on data processing and LSTM-RNNrdquo Arabian Journal for Science and Engineering vol 44no 4 pp 3043ndash3060 2019

[89] Q Shen X Ban C Guo and C Wang ldquoKernel based semi-supervised extreme learning machine and the application intraffic congestion evaluationrdquo Proceedings of ELM-2015vol 1 pp 227ndash236 2016

[90] Z Zhang A Zhang C Sun et al ldquoResearch on air traffic flowforecast based on ELM non-iterative algorithmrdquo MobileNetworks and Applications pp 1ndash15 2020

[91] Y-m Xing X-j Ban and R Liu ldquoA short-term traffic flowprediction method based on kernel extreme learning ma-chinerdquo in Proceedings of the 2018 IEEE International Con-ference on Big Data and Smart Computing (BigComp)pp 533ndash536 IEEE Shanghai China January 2018

[92] L Lin J C Handley and A W Sadek ldquoInterval prediction ofshort-term traffic volume based on extreme learning machineand particle swarm optimizationrdquo in Proceedings of the 96thTransportation Research Board Annual Washington DCUSA 2017

[93] S Zhang Y Yao J Hu Y Zhao S Li and J Hu ldquoDeepautoencoder neural networks for short-term traffic Conges-tion prediction of transportation networksrdquo Sensors vol 19no 10 p 2229 2019

[94] D Huang Z Deng S Wan B Mi and Y Liu ldquoIdentificationand prediction of urban traffic congestion via cyber-physicallink optimizationrdquo IEEE Access vol 6 pp 63268ndash63278 2018

[95] V Ahsani M Amin-Naseri S Knickerbocker andA Sharma ldquoQuantitative analysis of probe data character-istics coverage speed bias and congestion detection preci-sionrdquo Journal of Intelligent Transportation Systems vol 23no 2 pp 103ndash119 2019

[96] S J Kwon Artificial Neural Networks Nova Science Pub-lishers New York NY USA 2011

[97] Y Liu Z Liu and R Jia ldquoDeepPF a deep learning basedarchitecture for metro passenger flow predictionrdquo Trans-portation Research Part C Emerging Technologies vol 101pp 18ndash34 2019


the search process included ldquotraffic congestion predictionrdquoOR ldquotraffic congestion estimationrdquo OR ldquocongestion pre-diction modellingrdquo OR ldquoprediction of traffic congestionrdquo ORldquoroad congestion forecastrdquo OR ldquotraffic congestion forecastrdquoFor efficient screening research paper search was doneaccording to year using search engines like Scopus GoogleScholar and Science Direct After collecting all the peer-reviewed journal and conference papers written in theEnglish language 48 articles were found for review Anystudies focusing on the cause of traffic congestion trafficcongestion control traffic congestion impact traffic con-gestion propagation traffic congestion prevention etc wereexcluded from this manuscript

A general layout of the prediction approaches is pro-vided in Section 2 +e data collection sources and con-gestion forecasting models are explained in Sections 3ndash6 andthey provide the overall discussion and concluding remarks

2 General Layout

Traffic congestion forecasting has two basic steps of datacollection and prediction model development Every step ofthe methodology is important and may affect the results ifnot done correctly After data collection data processingplays a vital role to prepare the training and testing datasetsCase area differs for different research After developing themodel it is validated with other base models and groundtrue results Figure 1 shows the general components of trafficcongestion prediction studies +ese branches were furtherdivided into more specific sub-branches and are discussed inthe following sections

3 Data Source

Traffic datasets used in different studies can be mainly di-vided into two classes including stationary and probe dataStationary data can be further divided into sensor data andfixed cameras On the other hand probe data that were usedin the studies were GPS data mounted on vehicles

Stationary sensors continuously capture spatiotemporaldata of traffic However sensor operation may interruptanytime Authorities should always consider this temporaryfailure of the sensor while planning by using this data +eadvantage of the sensor data is that there is no confusion onthe location of the vehicles +e most used dataset wasPerformance Measurement System (PeMS) that collectshighway data across all major metropolitan areas of the Stateof California of traffic flow sensor occupancy and travelspeed in real-time Most of the studies used dataset from theI-5 highway in San Diego California every 5 minutes [3ndash6]Other systems included the Genetec blufaxcloud travel-timesystem engine (GBTTSE) [7] and the Topologically Inte-grated Geographic Encoding and Referencing (TIGER) linegraph [8]

On the other hand probe data has the advantage ofcovering the entire road network A network consists ofdifferent structured roads +erefore studies especiallythose that considered the network wide area used probedata +e most used dataset was GPS data collecting every

second from approximately 20000 taxies of Beijing ChinaData included the taxi number the latitude-longitude of thevehicle timestamp when sampling and whether there was apassenger or not Data updating frequency of this datasetvaries from 10 s to 5min according to the quality of GPSdevice [4 5 9] Other probe data included low-frequencyProbe Vehicle Data (PVD) [10] and bus GPS data [11 12]However sometimes probe data show significant fluctua-tion Besides map matching is usually a must for probe dataBut data can minimize this limitation Probe data collectedfrom one city cannot be used directly for modelling othercity networks+is is because the data collected from BeijingChina includes latitude-longitude of the vehicle which isunique However a generalised model using probe data canbe generated for different cities

Other data sources eg data from tolling system anddata provided by transportation authority will add morereliable data as the sources are dependable However a lot ofthe times study area needs to be adjusted as in most casestolled road information is not available Tracking cellularphone movements without privacy breach can also be asource of data However the heterogeneity of the vehicledistribution will be hard to determine from this dataset ifnot impossible Besides due to pedestrian or cyclists trav-elling through the sidewalk there might be many outliers inthe dataset if modelling is done for a road network Datacollected from a questionnaire to the general publicdriversmay provide a misleading result [13]

31 Clustering Algorithms Some studies use clustering theacquired data before applying the main congestion modelsof prediction +is hybrid modelling technique is applied tofine-tune the input values and to use them in the trainingphase Figure 2 shows the commonly used AI clusteringmodels in this field of research +e models are describedbriefly in this section

Fuzzy C-Means (FCM) is a popular nondeterministicclustering technique in data mining In traffic engineeringresearches traffic pattern recognition plays an importantrole Besides these studies often face the limitation ofmissing or incomplete data To deal with these constraintsFCM has become a commonly applied clustering technique+e advantage of this approach is unlike original C-meansclustering methods it can overcome the issue of gettingtrapped in the local optimum [14] However FCM requiressetting a predefined cluster number which is not alwayspossible while dealing with massive data without any priorknowledge of the data dimension Besides this model be-comes computationally expensive with data size incrementDifferent studies have applied FCM successfully by im-proving its limitations Some studies changed the fuzzyindex value for each FCM algorithm execution [15] somecalculated the Davies-Bouldin (DB) index [10] while othersapplied the K-means clustering algorithm [16 17]

K-means clustering is an effective and relatively flexiblealgorithm while dealing with large datasets It is a popularunsupervised machine learning algorithm Depending onthe features cluster number varied from two [18] to 50








K-means


Methodology


Data preprocessing

Scope of study area


Traffic parameters

Traffic volume

Traffic density

Sensor occupancy

Traffic speed

Traffic congestion

index

Model validation

AI models




Dataset


Probe vehicle data










MAE 1n

1113944

n

i1Yi minus Yi

11138681113868111386811138681113868111386811138681113868 (1)

sMAPE 1n

1113944

n

i1

Yi minus Yi

11138681113868111386811138681113868111386811138681113868

Yi

11138681113868111386811138681113868111386811138681113868 + Yi

111386811138681113868111386811138681113868111386811138681113872 11138732

100 (2)

RMSE

1113936ni1 Yi minus Yi( 1113857

2

n

1113971

(3)


FPR FP

TN + FP (4)

DR TP

FN + TP (5)




















Bayesian network





Variable 1

Variable 2



Variable 3






































Speed Speed

4



[28]



Congestion Index

Cao and Wang[3]

Density




Hidden Markov model




Main road Probe



Bayesian network



Yi Liu et al [59]

Bridge Sensor





Networkdirection


Kim and Wang[34]

















Output layer

Hidden layer

Input layer










Reference



Highwaycorridor


[11]Speed




Regression model


Temperature


HumidityRainfall

Traffic speedTime







Speed



Rainfall volume







Output layer

Output layer

Output layer
























3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References




























































































































K-means


Methodology


Data preprocessing

Scope of study area


Traffic parameters

Traffic volume

Traffic density

Sensor occupancy

Traffic speed

Traffic congestion

index

Model validation

AI models




Dataset


Probe vehicle data










MAE 1n

1113944

n

i1Yi minus Yi

11138681113868111386811138681113868111386811138681113868 (1)

sMAPE 1n

1113944

n

i1

Yi minus Yi

11138681113868111386811138681113868111386811138681113868

Yi

11138681113868111386811138681113868111386811138681113868 + Yi

111386811138681113868111386811138681113868111386811138681113872 11138732

100 (2)

RMSE

1113936ni1 Yi minus Yi( 1113857

2

n

1113971

(3)


FPR FP

TN + FP (4)

DR TP

FN + TP (5)




















Bayesian network





Variable 1

Variable 2



Variable 3






































Speed Speed

4



[28]



Congestion Index

Cao and Wang[3]

Density




Hidden Markov model




Main road Probe



Bayesian network



Yi Liu et al [59]

Bridge Sensor





Networkdirection


Kim and Wang[34]

















Output layer

Hidden layer

Input layer










Reference



Highwaycorridor


[11]Speed




Regression model


Temperature


HumidityRainfall

Traffic speedTime







Speed



Rainfall volume







Output layer

Output layer

Output layer
























3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References



























































































































MAE 1n

1113944

n

i1Yi minus Yi

11138681113868111386811138681113868111386811138681113868 (1)

sMAPE 1n

1113944

n

i1

Yi minus Yi

11138681113868111386811138681113868111386811138681113868

Yi

11138681113868111386811138681113868111386811138681113868 + Yi

111386811138681113868111386811138681113868111386811138681113872 11138732

100 (2)

RMSE

1113936ni1 Yi minus Yi( 1113857

2

n

1113971

(3)


FPR FP

TN + FP (4)

DR TP

FN + TP (5)




















Bayesian network





Variable 1

Variable 2



Variable 3






































Speed Speed

4



[28]



Congestion Index

Cao and Wang[3]

Density




Hidden Markov model




Main road Probe



Bayesian network



Yi Liu et al [59]

Bridge Sensor





Networkdirection


Kim and Wang[34]

















Output layer

Hidden layer

Input layer










Reference



Highwaycorridor


[11]Speed




Regression model


Temperature


HumidityRainfall

Traffic speedTime







Speed



Rainfall volume







Output layer

Output layer

Output layer
























3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References






























































































































Bayesian network





Variable 1

Variable 2



Variable 3






































Speed Speed

4



[28]



Congestion Index

Cao and Wang[3]

Density




Hidden Markov model




Main road Probe



Bayesian network



Yi Liu et al [59]

Bridge Sensor





Networkdirection


Kim and Wang[34]

















Output layer

Hidden layer

Input layer










Reference



Highwaycorridor


[11]Speed




Regression model


Temperature


HumidityRainfall

Traffic speedTime







Speed



Rainfall volume







Output layer

Output layer

Output layer
























3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References

























































































































































Speed Speed

4



[28]



Congestion Index

Cao and Wang[3]

Density




Hidden Markov model




Main road Probe



Bayesian network



Yi Liu et al [59]

Bridge Sensor





Networkdirection


Kim and Wang[34]

















Output layer

Hidden layer

Input layer










Reference



Highwaycorridor


[11]Speed




Regression model


Temperature


HumidityRainfall

Traffic speedTime







Speed



Rainfall volume







Output layer

Output layer

Output layer
























3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References

































































































































Output layer

Hidden layer

Input layer










Reference



Highwaycorridor


[11]Speed




Regression model


Temperature


HumidityRainfall

Traffic speedTime







Speed



Rainfall volume







Output layer

Output layer

Output layer
























3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References





























































































































Reference



Highwaycorridor


[11]Speed




Regression model


Temperature


HumidityRainfall

Traffic speedTime







Speed



Rainfall volume







Output layer

Output layer

Output layer
























3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References



























































































































Output layer

Output layer

Output layer
























3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References






































































































































3

Yuan-Yuan et al[79]





Volume

Road network Probe



Current time





roads










6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References






























































































































6 Future Direction





Deep ML

Shallow ML

Reasoning

0 2 4 6 8 10 12

201920182017

201620152014










Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References





























































































































Regression model













Fuzzy logic






Hidden Markovmodel








Bayesian network





7 Conclusions




Acknowledgments


References






















































































































7 Conclusions




Acknowledgments


References






















































































































AReviewofTrafficCongestionPredictionUsing …2020/08/08 · English language, 48 articles were...

Documents

Transcript of AReviewofTrafficCongestionPredictionUsing …2020/08/08 · English language, 48 articles were...