Traffic Breakdown Prediction Beyond Stochastic Capacity Models: Machine Learning Approach

Abstract

This study focuses on the application of machine learning models for a short-term prediction of traffic breakdowns on freeways. Traffic breakdowns, which occur when demand exceeds the momentary capacity, are typically predicted using probabilistic methods, but these approaches do not fully capture the short-term variability inherent in traffic flow. In this work, the methodology is advanced by employing machine learning techniques to predict traffic flow conditions, relying exclusively on lane-by-lane analysis of current detector data without utilizing any upstream or downstream information. Traffic conditions are classified into distinct categories, including breakdowns, and a neural network is employed to predict them, providing a robust method for identifying intervals in which the momentary capacity of a freeway is reached. Capacity estimates from the neural network are then compared with those from widely accepted statistical methods, revealing minimal differences, and thereby validating the effectiveness of the neural network approach in capacity analysis. Moreover, comparing the short-term flow conditions predicted based on the two approaches revealed superiority of neural network in providing significantly more accurate classifications. These findings highlight the significant potential of machine learning methods as powerful tools for momentary capacity estimation, with applications across various transportation systems management and operations strategies.