Research And Application Of Traffic Forecasting Models In City Road Network

With the development of the urban economy and increment of private vehicle, theconventional traffic pattern without intelligent techniques gets a lot troubles, such as trafficcongestion, awful air pollution and traffic accidents. Frequently constructing new roads to easetraffic pressure is not realistic and untenable in social and economic aspects. In the effort to dealwith such troubles, the intelligent transportation systems (ITS) has been developed, which combinesthe information technologies and transportation theories to help traffic control, traffic guide, etc.After several decades, ITS has shown its great potential for alleviating traffic jams, enhancingtraffic efficiency, etc. As one of the most important research subfields of ITS, traffic forecastingalways play a very important role, especially in the dynamic control of traffic control. The dynamiccontrol of traffic network depends on proper short-term/mid-term forecasting of traffic states. Theability to make and continuously update predictions of traffic flows and link times for severalminutes into the future using real-time data is a major requirement for providing dynamic trafficcontrol.Reasonably accurate forecasting modelling of traffic states is essentially important for betteranalyzing traffic conditions, designing efficient control strategies, and etc. Over the past decades,many researchers with different background had made intensive studies on this subject and a widerange of traffic theories and models had been developed. In this dissertation, based on the existingtraffic forecasting models, several improved mathematical models are presented, and thecomparisons of the advantages and disadvantages of the models with corresponding error analysisare conducted. Applying real traffic data from two different sources, the dissertation analyzes theapplication of the proposed models in city road network. On the one hand, considering thedifference of traffic on freeways and in urban road networks, the research is based on the twooccasions; on the other hand, selecting different prediction time-intervals, the studies compareforecasting models detailedly in mid-term and short-term prediction. The major contributions of thethesis are as follows: 1. On the basis of spatial-temporal information, the dissertation applies traffic data from roadnetworks to properly evaluate the traffic states and provide accurate forecasts. The researchobtains the travel time index (TTI) data from the website of Freeway PerformanceMeasurement System (PeMS), California. The analysis of the1-hour aggregated data canprovide real-time information on the whole freeway networks. Meanwhile, the studies alsoapply traffic flow data to evaluate the urban road network. The15-minute traffic flow datacomes from the Sydney Coordinated Adaptive Traffic System (SCATS) settled in HengshanRoad and Wuxing Road, Xuhui District, Shanghai.2. The choice of traffic data from two data sources with different time interval can help comparedifferent forecasting models comprehensively. Considering the commonly existingincompleteness of traffic data, the research analyzes the traffic network in two different ways.One method spatially aggregates the traffic data to evaluate the whole network from amacroscopic point of view. Correspondingly, the other one analyzes the correlation of trafficdata from different lanes using topology analysis and produces forecasts according to bothspatial and temporal information. The choice of the data in the dissertation is based on the point.The experiments apply different error rates including mean absolute error (MAE), root meansquare error (RMSE), mean absolute percentage error (MAPE), variance of absolute percentageerror (VAPE), etc. to compare the forecasting models.3. Applying the obtained traffic data from two different sources, the research compares theconventional individual forecasting models. The proposed spatial-temporal state space methodincorporates with the predictors. Two nonparametric models least squares support vectormachines (LS-SVM) and Fuzzy T-S models are demonstrated to be efficient. Meanwhile, theperformance of Kalman filtering (KF), autoregressive moving average (ARMA), historicalmean (HM), linear least squares regression (LLSR), radial basis function neural network(RBF-NN), and support vector regression (SVR) models are compared in detail using erroranalyses with different error rates.4. Based on the application of the obtained traffic data from two different sources, the analyses ofindividual forecasting models can provide important background for further research.Considering that the combined models can effectively reduce the diverse effects from statisticfactors of circumstances in forecasting process, the dissertation propose six linear combinedforecasting models according to the real-time requirements of the distribution and process oftraffic information: equal weights (EW), optimal weights (OW), minimum error I&II (ME I&ME II) and minimum variance I&II (MV I&MV II) models. Similarly, the researchalso applies different error rates to produce error analyses for over100combined models.The performance comparison proves that the2-model combined forecasts are superior toindividual ones in both accuracy and robustness.5. Motivated by the good performance of combined models with varying weights and the similarcharacteristics of the interacting multiple model (IMM) algorithm in estimation of hybridsystems, the dissertation firstly proposes the IMM-based combined forecasting models. TheIMM algorithm is recursive, modular and has fixed computational requirements per cycle,which are the three desirable properties to adapt to real-time traffic forecasting. the choice ofIMM-based models corresponds to the requirements of traffic information processing. Applyingthe traffic data with different time intervals from different data sources, the research analyzesall the combined models on the basis of the permutations and combinations of the individualpredictors in the two-model combining process. Comparisons of the IMM-based models withthe corresponding individual models and combined ones prove that the effectiveness androbustness of IMM-based combined predictors.