| Summary |
Yen-Ching Chiou, 2006.07
Feng Chia University - Graduate Institute of Traffic and Transportation Engineering and Management
The real-time traffic information detected by traffic devices might be outdated to implement dynamic traffic management and real-time traffic responsive control, because of the time lag caused by the transitions and calculations of these information or because of the significant distances between these traffic detection devices and traffic control points. Traffic control operators have to recognize how these information will evolve in the near future. Thus, an accuracy and efficiency short-term traffic forecast model is extremely essential to traffic control. Although numerous short-term traffics forecasting models have been proposed and validated, some of them are too complex in field application or others do not excellently perform. In contrast, the grey prediction model has overwhelming advantages of high accuracy, easy calculation and few calibrated samples required, therefore, this study aims to develop rolling grey prediction models (RGM) to forecast the short-term traffics.
Two RGM models are developed, RGM(1,1) and RGM (1,N). RGM(1,1) only considers the historical time-series data of a specific traffic information detected by a specific vehicle detector (VD). RGM(1,N) further considers the other related information, such as volume, speed and occupancy detected by neighbored VDs. For investigating and validating the accuracy and applicability of proposed models, three time horizons of short-term traffic data of 20-sceonds, 1-minute and 5-minutes are collected. For comparison, two commonly used short-term traffic forecast models, statistical time-series model (ARIMA) and artificial neural network (ANN), are also developed and compared.
All traffic datasets, including three major traffics: volume, occupancy and speed, are collected from national No.1 freeway in Taiwan. Each of them is further divided into two subsets for training (70%) and validation (30%), respectively. The accuracies in term of mean absolute percent error (MAPE) of different numbers of rolling intervals (4-8interval) and forecast periods (1-5period) of the proposed model are also compared. The results show that, taking four rolling intervals and one forecast period RGM(1,1) model on 1-minute traffics for instance, it can predict volume with MAPE =17.51%, speed with MAPE=3.46%, and occupancy with MAPE=9.81%, which are significantly better than those of ARIMA (42.77% for volume, 11.51% for speed and 33.88% for occupancy). Furthermore, the RGM(1,6) can accurately predict volume with MAPE =5.44%, speed with MAPE=0.05% and occupancy with MAPE=4.44%, which are still remarkably better than those of ANN (15.34% for volume,7.54% for speed and 16.73% for occupancy) and those of RGM(1,1). Obviously, the performances and applicabilities of proposed RGM models are validated. |
