Data-Driven Prediction Methodology of Origin-Destination Demand in Large Network for Real-Time Services

Abstract

Prediction of origin–destination (O-D) demand is an important topic in transportation engineering because it is a crucial input for a dynamic traffic management and control system. Previous literature has focused primarily on estimation and prediction of O-D demand with Kalman filtering; however, these forecasts lack efficiency when unusually fluctuating O-D demand of a large O-D network is predicted in real time. With true, real-time O-D demand data from South Korean expressways, a data-driven prediction framework of O-D demand in a large network for real-time service is proposed by modifying the k–nearest neighbor (k-NN) algorithm. Three strategies that implement different feature vectors for k-NN prediction of single-level O-D demand, multilevel O-D demand, and single-level point demand are proposed. The strategies were tested on hourly O-D demand in South Korea. The average mean absolute percentage of error values of the three strategies in terms of total demand are 5.52%, 5.34%, and 3.36%, respectively; single-level point demand performs slightly better than do the other two strategies. Similarly, for the average mean absolute percentage and weighted average mean absolute percentage in terms of individual O-D demand, single-level point demand performs better than do the two other strategies, especially for O-D pairs with larger demand and for further prediction horizons. In addition, the single-level point demand shows the highest computation efficiency. Therefore, the single-level point demand strategy for k-NN prediction shows the best combination of accuracy and computation efficiency among the three strategies. Furthermore, a historical database size of at least 300 dates for this data-driven prediction algorithm seems required for accuracy.