A Lane-based Predictive Model of Downstream Arrival Rates in a Queue Estimation Model Using a Long Short-Term Memory Network

Abstract

Abstract In this study, we develop a mathematical framework to predict cycle-based queued vehicles at each individual lane using a deep learning method – the long short-term memory (LSTM) network. The key challenges are to decide the existence of residual queued vehicles at the end of each cycle, and to predict the lane-based downstream arrivals to calculate vertical queue lengths at individual lanes using an integrated deep learning method. The primary contribution of the proposed method is to enhance the predictive accuracy of lane-based queue lengths in the future cycles using the historical queuing patterns. A major advantage of implementing an integrated deep learning process compared to the previously Kalman-filter-based queue estimation approach (Lee et al., 2015) is that there is no need to calibrate the co-variance matrix and tune the gain values (parameters) of the estimator. In the simulation results, the proposed method perform better in only straight movements and a shared lane with left turning movements.