Travel Time Prediction for Urban Networks: the Comparisons of Simulationbased and Time-Series Models Ta-Yin Hu Department of Transportation and Communication Management Science, National Cheng Kung University, Taiwan, R. O. C. Wei-Ming Ho Department of Transportation and Communication Management Science, National Cheng Kung University, Taiwan, R. O. C. 2010. 28 17 th ITS WORLD CONGRESS
OUTLINE u u INTRODUCTION RESEARCH FRAMEWORK – – u OVERALL FRAMEWORK SIMUALTION-BASED MODEL TIME-SERIES MODEL MEASUREMENT CRITERIA NUMERICAL EXPERIMENTS AND RESULTS – NETWORK CONFIGURATION – DATA COLLECTION – RESULTS ANALYSIS u CONCLUDING COMMENTS 17 th ITS WORLD CONGRESS 2
u u INTRODUCTION RESEARCH FRAMEWORK – – u OVERALL FRAMEWORK SIMUALTION-BASED MODEL TIME-SERIES MODEL MEASUREMENT CRITERIA NUMERICAL EXPERIMENTS AND RESULTS – NETWORK CONFIGURATION – DATA COLLECTION – RESULTS ANALYSIS u CONCLUDING COMMENTS 17 th ITS WORLD CONGRESS 3
INTRODUCTION u With the development of modern technologies in Intelligent Transportation Systems (ITS), the route guidance with predicted travel time is an important issue, especially in Advanced Traveler Information System (ATIS) (Chen and Chien, 2001). 17 th ITS WORLD CONGRESS 4
u Travel time is a key factor that can influence drivers’ behavior, and can let travelers understand what kind of traffic situations they are going to encounter while traveling (Tam and Lam, 2009). u Travel time is defined as the time required to travel along a route between any two points within a traffic network (Ma et al. , 2009), (Ben-Akiva et al, 1991). 17 th ITS WORLD CONGRESS 5
u With the development of new technologies, many instruments can be utilized to collect the traffic data, such as global positioning system (GPS), automated vehicle identification (AVI), and vehicle detector (VD). 17 th ITS WORLD CONGRESS 6
u The travel time data can be directly/indirectly predicted through these technologies: – Travel time can be estimated and predicted directly by using probe vehicles, license plate matching, electronic toll stations, and automatic vehicle identification (AVI) etc. ; – travel time can also be estimated and predicted indirectly by vehicle detectors (VD) (Vanajakshi et al. , 2008). 17 th ITS WORLD CONGRESS 7
u Several approaches have been proposed for travel time estimation and prediction. – statistical-base algorithm: • Most of the statistical-based algorithms are based on the applications of regression, bayesian and time series models (Lam and Toan, 2008), (Hinsbergen and Lint, 2008), (Guin, 2006). • In general, the algorithms are easy to implement, but purely statistical-based algorithms may result in poor performance during abnormal traffic conditions (Ishaak and Al-Deek, 2003). 17 th ITS WORLD CONGRESS 8
– Simulation-based method: • The simulation-based method used traffic simulation software and can integrate other algorithms such as Kalman filter model and traffic flow theory model to simulate the traffic pattern (Liu et al. , 2006) • Several empirical studies show that the extensive data needs to be collected for the model validation (Tam and Lam, 2009). 17 th ITS WORLD CONGRESS 9
u In summary, how to apply these algorithms under different traffic situations is still a critical issue and so far there is no conclusive evidence to demonstrate the best algorithm for travel time estimation and prediction. 17 th ITS WORLD CONGRESS 10
u This paper presents a simulation-based model and a time-series model for travel time prediction for urban networks. – The simulation-based model, calibrated based on VD flow data, uses simulated vehicle trajectories to generate travel time information. – The ARIMA model, calibrated based on timeseries data, is integrated with signal delay for travel time prediction. 17 th ITS WORLD CONGRESS 11
u The main advantage of the simulationbased model is: – easy to be implemented with possible applications for normal and abnormal traffic conditions. u The main advantage of ARIMA is: – efficiency for short-term predictions. 17 th ITS WORLD CONGRESS 12
u u INTRODUCTION RESEARCH FRAMEWORK – – u OVERALL FRAMEWORK SIMUALTION-BASED MODEL TIME-SERIES MODEL MEASUREMENT CRITERIA NUMERICAL EXPERIMENTS AND RESULTS – NETWORK CONFIGURATION – DATA COLLECTION – RESULTS ANALYSIS u CONCLUDING COMMENTS 17 th ITS WORLD CONGRESS 13
OVERALL FRAMEWORK u u u The collected input data sets for travel time model include network, O-D flows, and VD data. The true travel time values for validation are collected through probe vehicles. The travel time information can be sent to the information management center. 17 th ITS WORLD CONGRESS 14
SIMUALTION-BASED MODEL u u The input data sets for travel time model include O-D flows, network, alternative routes and other traffic conditions. The time-dependent flow can be simulated through Dyna. TAIWAN, and route travel time can be calculated based on the vehicle trajectory data. 17 th ITS WORLD CONGRESS 15
u The travel time is directly calculated by averaging the travel time of vehicles start from origin node to destination node for each time interval, as described in the following equation: u is the vehicle-based average travel time from origin node r to destination node s during time interval k is the average travel time of vehicle from origin node r to destination node s during time interval k. u 17 th ITS WORLD CONGRESS 16
u The features of simulation-based model are summarized as follows: – the algorithm is reliable since the vehicle trajectory is obtained through simulation. – The vehicle trajectory data is managed through relational data base management systems, such as Access, thus travel time information can be retrieved easily. 17 th ITS WORLD CONGRESS 17
TIME-SERIES MODEL u In time series analysis, an ARIMA model is a generalization of an autoregressive moving average (ARMA) model. – The common approaches for modeling univariate time series includes: • the autoregressive (AR) • the moving average (MA) models and • the ARIMA, which is a model after the combination of AR and MA models. 17 th ITS WORLD CONGRESS 18
u The AR, MA, and ARMA model can be expressed as follows: – AR Model: – MA Model: – ARMA Model: 17 th ITS WORLD CONGRESS 19
Where, – : time series – : constant – : white noise – : the mean – , : parameters of the model – p : the order of the AR model – q : the order of the MA model 17 th ITS WORLD CONGRESS 20
PROCEDURE OF TIMESERIES MODEL u u The historical speed data are collected as the basic time series in the ARIMA model. If the historical time series are stationary, the order of p, d, q of the model can be identified initially, or the data needs to differenced before identification. The parameters of the ARIMA model can be estimated and diagnosed to check the parameters of the model and to determine the final time-series formulations. The final model can be applied to predict the objective time series and compared with the true value to verify the accuracy of the model. 17 th ITS WORLD CONGRESS 21
MEASUREMENT CRITERIA u The MAPE indexes is applied to measure the performance of the models. Where, – – – M ：the number of samples ：the real values ：the estimated values 17 th ITS WORLD CONGRESS 22
u The interpretation of MAPE index: (Lewis, 1982). MAPE (% ) Interpretation <10 Highly accurate forecasting 10 -20 Good forecasting 20 -50 Reasonable forecasting >50 Inaccurate forecasting 17 th ITS WORLD CONGRESS 23
u u u INTRODUCTION LITERATURE REVIEW RESEARCH FRAMEWORK – – u OVERALL FRAMEWORK SIMUALTION-BASED MODEL TIME-SERIES MODEL MEASUREMENT CRITERIA NUMERICAL EXPERIMENTS AND RESULTS – NETWORK CONFIGURATION – DATA COLLECTION – RESULTS ANALYSIS u CONCLUDING COMMENTS 17 th ITS WORLD CONGRESS 24
NETWORK CONFIGURATION u Numerical experiments are conducted for an arterial street in Kaohsiung city in Taiwan to illustrate these two models. – The arterial street selected (the East bound and West bound) is located in the CBD with heavy traffic. 17 th ITS WORLD CONGRESS 25
u u The East bound is from tag A to B and the node number is from 2616 to 5142. The West bound is from tag B to A and the node number is from 5142 to 2616. There are 8 intersections and 2 ramps in this arterial street. 17 th ITS WORLD CONGRESS 26
DATA COLLECTION u Empirical travel time data: – The empirical travel time is used in the comparison process for the simulation-based model and the ARIMA model. The data is gathered from 7: 30~8: 00 AM in August 19, 2009. 1 2 3 Average Travel Time East bound: 2616→ 5142 6. 73 7. 23 6. 97 0. 31 West bound: 5142→ 2616 6. 33 6. 43 6. 9 6. 56 0. 15 Stander Deviation （Unit: min） 17 th ITS WORLD CONGRESS 27
u Historical speed data: – The historical speed data is from 7: 30~8: 00 AM for 17 days. – The speed data is collected through VDs. 17 th ITS WORLD CONGRESS 28
u Signal cycle data: – The signal cycle data are collected and used to calculate the expected delay for 8 intersections and 2 ramps, and the total expected delay is about 264. 5 seconds. 17 th ITS WORLD CONGRESS 29
RESULTS ANALYSIS u Simulation-based Model – In order to observe the system performances under different demand levels, several loading factors are tested. • The loading factor is defined as the ratio of the total number of vehicles generated in the network during the simulation period. • The loading factors are tested from 0. 5 to 0. 9. 17 th ITS WORLD CONGRESS 30
u The results show that the average travel time increases with respect to the loading factor. Loading Factor East Bound: 2616 -5142 West Bound: 5142 -2616 0. 5 0. 6 0. 7 0. 8 0. 9 4. 38 5. 19 6. 18 7. 39 8. 35 6. 35 8. 58 13. 73 10. 16 13. 15 （Unit: min） 17 th ITS WORLD CONGRESS 31
u Time-Series Model u The ACF and PACF values are shown below. – The models are ARIMA(1, 1, 1) for East bound street and ARIMA(2, 0, 1) for West bound street. Time-Series Model, ARIMA(p, d, q) Parameters p d q (p, d, q) East Bound: 2616 -5142 1 1 1 (1, 1, 1) West Bound: 5142 -2616 2 0 1 (2, 0, 1) 17 th ITS WORLD CONGRESS 32
u The predicted speed from 7: 30~8: 00 and average values on East bound street and West bound street shows that the timeseries model can predict the travel time reasonably if the signal delay is included. – The speed data from VDs does not able to reflect signal delay and possible reasons might be the location of installed VD. 17 th ITS WORLD CONGRESS 33
Predicted Average Speed (km/hr) 07: 30~ 07: 35~ 07: 40~ 07: 45~ 35 40 45 50 07: 50~ 55 0755~ 08: 00 Ave East Bound: 50. 15 43. 87 47. 85 50. 24 52. 50 49. 02 48. 94 West Bound: 40. 09 40. 77 39. 50 40. 03 38. 03 35. 20 38. 71 Time Predicted Average Travel Time from 7: 30~8: 00 (min) East Bound: 3. 56 West Bound: 4. 50 Predicted Average Travel Time + Delay from 7: 30~8: 00 (min) East Bound: 7. 96 West Bound: 8. 91 17 th ITS WORLD CONGRESS 34
u The best results are compared: – The simulation-based model shows possible potential for travel time prediction; however, the timedependent O-D demand is still a crucial issue in applying the simulation-based model. – For normal traffic conditions, the ARIMA model could also have reasonable results. Model East Bound West Bound ARIMA 7. 96 8. 91 Dyna. TAIWAN 7. 39 6. 35 True Value 6. 97 6. 56 （Unit: min） 17 th ITS WORLD CONGRESS 35
Outline u u INTRODUCTION RESEARCH FRAMEWORK – – u OVERALL FRAMEWORK SIMUALTION-BASED MODEL TIME-SERIES MODEL MEASUREMENT CRITERIA NUMERICAL EXPERIMENTS AND RESULTS – NETWORK CONFIGURATION – DATA COLLECTION – RESULTS ANALYSIS u CONCLUDING COMMENTS 17 th ITS WORLD CONGRESS 36
CONCLUDING COMMENTS u In this research, – the simulation-based model based on Dyna. TAIWAN and – the time-series with ARIMA model for travel time prediction are presented. u The numerical results indicate the proposed simulation-based travel time prediction model and the ARIMA model provides reasonable travel time information. 17 th ITS WORLD CONGRESS 37
u Some important issues for simulation-based model include: – (1) the network configuration and – (2) time-dependent O-D demand data are very important in the simulation-based model, and these data sets can significantly influence the performances of travel time predictions. u Some important issues for time-series model include: – (1) the accuracy of the model is influenced by the outliers of the historical time-series; – (2) the determinations of orders of p and q may highly influence the performances. 17 th ITS WORLD CONGRESS 38
u Thanks for your listening. 17 th ITS WORLD CONGRESS 39
Скачать презентацию Travel Time Prediction for Urban Networks the Comparisons
2d6492b7b99d1db791488557c1d730d5.ppt