Short-term Traffic Flow Prediction Using Group Method Data Handling (GMDH)-based Abductive Networks

Abstract

This study investigates a method of predicting traffic flow that employs the use of abductive networks based on the group method data handling (GMDH) for short-term traffic flow prediction. The GMDH algorithm relies on high-order polynomial input variables and it starts by building regression equations of order 2 or 3 for each pair of input variables. The new input variables are used for predicting the output in lieu of the original input variables. Based on pre-specified selection criteria, the best new input variables (polynomials) in terms of predicting the output are retained and used as input variables for the next layer to generate newer input variables of higher order (order of 4 if started with order 2). This process continues until the added value of the predicting power becomes insignificant and/or the model becomes practically complex for predicting purposes. Models for a linear road network were developed first using both spatial and temporal information without differentiating between weekdays and weekends. In the subsequent efforts, different models were built for weekdays and weekends. It was found that day-specific mode performance is not better than the generic model in predicting traffic flow. Models developed for predicting traffic after 15 min had correlation coefficients between 0.97 and 0.98. Models which were developed to predict traffic after 30 min were also robust but with slightly lower values of correlation coefficient. Due to the self-organizing nature of the models and the minimum required interventions, the models can be easily used by practitioners.