A Novel Method to Predict Traffic Features Based on Rolling Self-Structured Traffic Patterns

Abstract

In this study, a novel method is proposed to predict the traffic features in a long freeway corridor with a number of time steps ahead. The proposed method, on the basis of rolling self-structured traffic patterns, utilizes the growing hierarchical self-organizing map model to partition the unlabeled traffic patterns into an appropriate number of clusters and then develops the genetic programming model for each cluster to predict its corresponding traffic features. For demonstration, the proposed method is tested against a 110-km freeway stretch, on which 48 time steps of 5-min traffic flows are predicted (i.e., a 4-h prediction). The prediction accuracy of the proposed method is compared with other models (ARIMA, SARIMA, and naive models) and the results support the superiority of the proposed method. Further analyses indicate that applications of the proposed method to larger scale freeway networks require sufficient lengths of observation to acquire enough traffic patterns for training and validation in order to achieve higher prediction accuracy.