01049381

Component

The present paper adopts a pattern-based consideration to short-term prediction of traffic flow. A comprehensive methodological framework for detecting and identifying the characteristics of the temporal evolution of traffic flow is presented; this information is integrated into a set of genetically-optimized neural network structures that are trained to generalize in specific temporal behavior of traffic flow with respect to its deterministic and nonlinear characteristics. Predictions are significantly improved in terms of accuracy comparing to the ones of a dynamic neural network and an ARIMA model that operate on time-series thinking. Additionally, the pattern-based approach outperforms the other two methodologies in terms of the acquired knowledge of the anticipated traffic flow regime. Finally, the functional architecture of a prediction system is presented; the system’s performance is tested on arterials. Results of iterative predictions show that the system can provide multiple steps ahead predictions with better accuracy comparing to the one-step ahead predictions of dynamic neural networks and ARIMA.

01042056

07-0814

English

25p

2007

Transportation Research Board 86th Annual Meeting

CD-ROM

Figures (9) ; Tables (6)

Embedded systems; Highway operations; Information dissemination; Information systems; Methodology; Neural networks; Operations; Time series analysis; Traffic flow; Traffic surveillance

Pattern analysis; Short term

Administration and Management; Data and Information Technology; Highways; Operations and Traffic Management; Planning and Forecasting; I72: Traffic and Transport Planning

Transportation Research Board Annual Meeting 2007 Paper #07-0814

TRIS, TRB

Feb 8 2007 5:24PM